Studies that remove or modify certain components of the model to understand their contribution to the overall performance.
Produces coherent summaries by conducting cross-sentence information ordering, compression, and revision
A semantic formalism that encodes core concepts and models relations in the text at a high level of abstraction.
A reward used in the mixed objective to encourage the generation of words not in the source document in a summarization model
Methods that aim at constructing new sentences as summaries, thus they require a deeper understanding of the text and the capability of generating new sentences, which provide an obvious advantage in improving the focus of a summary, reducing the redundancy, and keeping a good compression rate.
Summarization algorithms that may generate new text that is not present in the initial document.
Summaries that use natural language to convey the meaning of the original document.
A subjective task in natural language processing that involves summarizing text in a way that captures the main ideas and meaning of the original text.
Preferred by users over extractive summaries
A type of QFS that incorporates the query relevance into existing neural summarization models.
Summarization systems that generate summaries based on specific queries or questions
The process of summarizing questions in a concise and informative manner
Automatically generating concise and coherent summaries of user reviews.
A subtask in CASAS that generates aspect/sentiment-aware abstractive summaries of reviews.
Summarization techniques that generate summaries with novel words.
Models that generate a summary of a text by paraphrasing and rephrasing the original content, rather than simply selecting and extracting sentences
Summarisers that generate new sentences that capture the essence of the input document to form the summary
A technique for generating summaries that go beyond verbatim copying of the original text and instead generate new and abstract concepts that reflect high-level semantics.
A summarization strategy that involves paraphrasing the source document.
A system that produces summaries that resemble reference summaries on a word-to-word basis.
A summary that uses natural language to convey the meaning of the original text
Summarization systems that generate summaries by paraphrasing and rephrasing the input text
Generate novel words and sentences
Attentional feed-forward network and recurrent neural network-based encoder-decoder models.
Defining bins corresponding to three abstractiveness levels and designing constraints that allow users to control the summary’s abstractiveness
A component of the hybrid architecture that rewrites and compresses each of the extracted sentences.
A metric introduced to extract summaries by exploiting the abstract of a paper
Functions used in Bayesian optimization to quantify the value of querying the user about a particular pair of candidates.
Extracting action triplets from the text and constructing an action graph to encode the structural information in the unstructured text.
A common solution to reduce the number of labels a user must provide by iteratively acquiring labels and training a model on the labels collected so far.
A reinforcement learning technique used to connect the extractor and abstractor networks and learn sentence saliency.
Existing methods designed for specific scenarios, limiting their applicability in a unified framework.
Summaries that possess desired properties and accurately convey the main concepts of the original text.
A technique used to generate hard factual inconsistent examples for training a model
A framework that jointly trains a generative model G and a discriminative model D, where G takes the original text as input and generates the summary, and D learns to classify the generated summaries as machine or human generated
A methodology for evaluating the performance of factuality metrics by applying adversarial transformations to test datasets with different distributions.
A summary given to patients after their clinical visit, intended to summarize patients’ clinical visits and help their disease self-management.
A framework in which the similarity of two summaries is calculated based on an alignment between the summaries’ tokens
Multiple perspectives on the same event.
A model that rewrites pronominal mentions to increase expressivity.
A specific implementation of the protocol that utilizes a set of lexical items in the source document to compute ROUGE metric.
The design of the manual annotation process, including the number of annotators and the distribution of annotators to annotation items
A method for collecting human judgements on the factuality of model-generated summaries
The degree to which different annotators agree on the same evaluation
A set of guidelines used to annotate or mark up text for specific purposes.
Refers to a study conducted to identify errors made by state-of-the-art summarizers on two benchmarks
The process of evaluating the overlap between the answers generated by the QA model and the selected answers.
A score that predicts the answer relevance of the given source documents to the query.
A task in CQA that involves selecting the correct answer from a set of candidates.
A method for answer summarization that involves selecting relevant sentences from a set of candidate sentences to answer a question.
The paper introduces the concept of answer summarization, which is a form of query-based, multi-document summarization that involves several subtasks, including query sentence relevance, clustering, cluster summarization, and fusion.
The task of verifying whether the QA model's prediction is correct
A noise-estimating model that can be trained from a single noisy corpus to distinguish appropriate and inappropriate pairs of source and target texts.
An aspect that may not be explicitly mentioned but is related to portions of the document
Models used to evaluate the quality of arguments
The process of mining product-related aspects and identifying sentences related to those aspects.
Words indicating product information
The process of analyzing the sentiment expressed towards different aspects of a product or service in a review.
A type of opinion summarization that allows for the control of the number and type of aspects included in the summary
Keywords that represent specific aspects or topics within a document
Summarization that focuses on specific aspects or topics within a document
Representations used in QT for identifying specific aspects of an entity.
Detailed summaries on individual aspects of an entity
A type of opinion summarization that focuses on creating summaries for individual aspects of a product or service
Transcriptions of speech generated by Automatic Speech Recognition technology
Individuals hired to collect relevance judgments for tasks such as test collection construction
Articles from alternative news resources covering the same news event that can complement the background knowledge in a generated summary.
A variant of sequence-to-sequence models that attend to relevant parts of the input when producing an output
A method of ensuring factual consistency by focusing on triples of subject, predicate, and object
An update unit based on current decoding state, designed to retain the attention on salient parts but weaken the attention on irrelevant parts of input.
A technique used in recurrent neural networks to inform content selection and boost summarizer performance.
An encoder-decoder approach that uses attention mechanisms to focus on important parts of the input
A relevant model to the task of abstractive summarization, originally developed for machine translation, which has produced state-of-the-art performance in machine translation.
Information about users who post reviews, such as gender, age, and occupation.
A model that incorporates attribute information into review summarization using a sequence to sequence (S2S) model with an attribute encoder, attribute-aware review encoder, and attribute-aware summary decoder.
Words and phrases that are specific to certain user attributes and can be incorporated into review summarization to improve performance.
A method of compressing data into a latent representation without supervision
A decoding method where the model generates one token at a time based on the previously generated tokens.
A type of neural network architecture used for text generation that generates one word at a time based on the previous words generated.
A type of language model that generates text one word at a time, conditioned on the previous words.
Using previous predictions to inform future predictions in sentence extraction
Algorithms that automatically generate new, smaller, customized models for a custom dataset.
Highly desirable since it requires a lot of effort and language skills for generating summaries from varying information sources such as social media, databases, web articles etc.
The evaluation of a summarization model without human intervention
The generation of a summary without human intervention
A strategy to automate the pipeline for model creation, including automated generation of the model itself.
Different types of measures, including question-answering, text reconstruction, semantic similarity, and lexical overlap, with various families of measures within each type
The process of automatically finding and retrieving information
The process of generating a summary sentence from a set of input sentences.
An efficient evaluation method for the quality of learner summaries that can enhance reading comprehension tasks.
The paper discusses the growing demand for automated summarization of online conversations due to the increasing amount of information exchanged.
Systems that produce a summary of a text automatically
The process of generating a summary of a given text using machine learning techniques.
A method for producing concise and coherent summaries to facilitate quick information consumption
Designed for comparative summaries on a new dataset of three ongoing controversial news topics.
Automatic detection of factual inconsistency in summarization models
The process of creating a shorter version of a longer text document while retaining its most important information.
Summarizing to-do items from given emails to help people overview overwhelming numbers of emails they receive every day and schedule their daily work.
A metric used to evaluate system performance for the task of document summarization.
Challenging and often inconsistent with human evaluation
The use of fact descriptions brings significant improvement in this aspect.
Method for fully automatic key point analysis that selects short, high quality comments as key point candidates and leverages previous work on argument-to-key-point matching to select a subset of candidates that achieve high coverage of the data
Transcriptions of speech from meetings that are generated automatically using speech recognition technology
Methods used to automatically evaluate the quality of summarization
A model that models the unidirectional dependency between sentences, where the state of the current sentence is based on previously sentence labels.
A solution to the issue of redundant phrases by introducing an autoregressive decoder, which extracts sentences one by one and allows different sentences to influence each other.
A type of pre-training objective for language models where the model is trained to predict the next word in a sequence
A technique that encourages the model's scores for sentences to mimic an estimated score distribution over the sentences
Added to strengthen the robustness of the node representations. Aims to classify the shortest path length (e.g., N-hops) from query nodes to others.
Additional tasks designed to improve the performance of the main summarization task, such as recognizing salient entities and ensuring factual consistency.
A technique used in training neural networks to adjust the weights of the model based on the error between the predicted output and the actual output.
A method of representing text data as a bag of its words, disregarding grammar and word order.
The parameterized models between two learning stages are relatively independent, preventing the meta model from fully utilizing the knowledge encoded in the base systems.
Algorithms used to compare and evaluate the performance of example-based summarization systems
A parameter used in the beam search algorithm for text generation. Raffel et al. (2019) used a beam search length penalty of 0.6 in their study.
An optimisation strategy that iteratively evaluates the semantic coverage scores of a set of candidate summaries to obtain the optimal extraction.
A mechanism that helps to ensure that the generated summary is similar to the reference summary.
A ranking-based method in which the annotator selects the best and worst example out of a set of examples
The labeling of summaries as either factual or non-factual, which can be difficult to determine and may not provide a fine-grained understanding of factual errors
A method of solving the non-convex sparse optimization problem in compressive summarization.
Resampling techniques used to generate new instances of matrices of values, which are used to calculate the correlation of an evaluation metric to human judgments.
An algorithmic approach that divides a problem into smaller sub-problems and solves them recursively.
The quality of being concise and to the point
Given a collection S of sets with associated costs and a budget L, find a subset S′ ⊆ S such that the total cost of sets in S′ does not exceed L, and the total weight of elements covered by S′ is maximized.
A technique used to split rare words into sub-units to eliminate the out-of-vocabulary (OOV) problem.
The quality of candidate summaries generated by the abstractive model, which can be estimated more accurately using the proposed training paradigm.
A type of architecture for neural text summarization that involves a pipeline approach where one or two sentences are selected from the source document, their summary-worthy segments are highlighted, and a summary sentence is composed using a neural generator.
A method used to align each section of the structured summary with a set of input sentences classified as the same type.
A method for determining causal relationships between variables
Identifying the most important content in a document cluster
A measure used to determine the importance of each sentence in a document
Factors considered by selection models in extraction-based summarization systems
An attention function that is extended to consider the centrality score of source words.
important for generating high-quality summaries by allowing models to learn from reference papers
The set of citing sentences to a paper, which contains extra information that does not appear in paper abstracts and is more focused in describing the paper's main contributions.
A short textual description regarding related work in scientific literature that highlights certain contributions of the referenced paper and can provide useful information about that paper.
A type of scientific summary which is formed by utilizing a set of citations to a referenced article. This set of citations has been previously indicated as a good representation of important findings and contributions of the article. Contributions stated in the citations are usually more focused than the abstract and contain additional information that is not in the abstract.
The textual spans in the reference articles that reflect the citation.
A process involving clinical evidence acquisition with integration and abstraction over medical knowledge to synthesize a conclusion in the form of a diagnosis.
Terms related to the clinical domain, which can be used to improve the content selection and summary generation of radiology reports.
An algorithm that takes two review sets as input to compare and contrast the token probability distributions of the models to generate more distinctive summaries.
A matrix that shows the frequency of co-occurrence of words or phrases in a corpus
A technique used to compensate for the missing regularization requirement of abstractive summarization in the standard framework, learning a category-specific text encoder and improving the quality of locating salient aspect information of the review.
A two-layer hierarchical attention method that divides the input into chunks and sparsely attends to one or a few chunks at a time using hard attention, then applies full attention over those chunks
Linguistically motivated patterns created by entities in connected sentences that ensure coherence in automatic summarization.
A method that ranks summaries based on their coherence using human annotated coherence scores.
Summaries that are logically connected and easy to read
Words or phrases that tie sentences together into a coherent text.
Dividing the task of encoding a long text across multiple agents, each in charge of a different subsection of the text
The process of generating knowledge or solutions through the collaboration and input of multiple individuals
The most related sentence to the to-do item.
Prior knowledge that facilitates and enhances human reading, which is expected to have a large influence on the reading process as it helps the reader to construct a coherent mental representation of the document and gain an overview of the content in the document.
A type of question answering system that involves real-world applications such as Yahoo! Answer and StackExchange.
A comparison of well-known summarization systems regarding their implicit choices of θ by measuring the correlation of their θ functions with human judgments on two datasets from the Text Analysis Conference (TAC).
A novel task that generates two contrastive summaries and one common summary by comparing multiple entities.
Selecting documents that represent each group and highlight differences between groups.
A method for modeling the interaction between QA pairs by aggregating comparison signals from low-level elements into high-level representations.
post-editing is less likely to lead to problems of factual correctness and consistency
The existence of complementarity introduced by decoding algorithms or system combination among the current state-of-the-art summarization systems, which can be effectively utilized by Refactor for boosting the system performance.
The different types of guidance signals investigated in the paper are complementary to each other, and their outputs can be aggregated together to obtain further improvements in summarization quality.
The property of word embeddings that allows the combination of vectors to represent the meaning of a phrase or sentence.
Lacking it might result in generating puzzling words where some subtopics are unnecessarily accessed for multiple times and generating faultiness summary in which some salient information is mistakenly unexplored.
Summaries that are compressed automatically for each sentence
A technique historically used in heuristic-driven systems or in systems with only certain components being learned. It involves identifying permissible deletions in text.
Summarization techniques used to reduce the length of the summary
A technique used in hybrid methods to discard uninformative phrases in selected sentences.
A paradigm used to train unsupervised encoder-decoder-based summarizers in the absence of paired data
A system that compresses text in a data-driven way, offering a tradeoff between the robustness of extractive models and the flexibility of abstractive models.
Visual summaries structured as directed graphs
A novel model that encourages the generation of conceptual and abstract words by leveraging context-aware conceptualization and a concept pointer, both of which are jointly integrated into the generator to deliver informative and abstract-oriented summaries.
The process of reducing the number of concepts in the model to find optimal solutions efficiently.
A knowledge graph used to expand the aspect scope and enrich the supervisions
The quality of being concise, expressing much in few words.
A method of generating summaries that uses control codes to condition the output.
Models used for single-document news summarization that include sequence-to-sequence architectures with attention and copy mechanisms, Transformers, and pre-trained language modeling.
A method to directly test abstractive summarizers in terms of how they score potential candidate summaries, allowing for testing of specific desired qualities such as semantic consistency and entailed by the source text.
A type of recurrent neural network that generates output based on a conditioning input
The selection of the most salient points and how those points are expressed are explicitly conditioned on an ad-hoc context, such as a question or topic of interest.
Results are reported on three canonical conditional text generation tasks of increasing complexity
A range of values that is likely to contain the true value with a certain level of confidence
The impact of output length on the evaluation of summarization systems.
Summaries that represent dominant opinions in reviews and can be useful for quick decision making and getting an overall feel for a product or business.
A process of generating summaries that represent dominant opinions in reviews, which can be useful for quick decision making and to get an overall feel for a product or business.
Improvement of controllable summarization models based on two different architectures
Providing summarized content with an additional constraint, i.e., the commonality criteria.
Ensuring that critical pronoun references are clear in the final summary.
Ensuring that sentence realizations are well-formed.
Questions related to health asked by patients and their families, which tend to include numerous peripheral details that are not always needed to find correct answers.
The degree to which the summary accurately represents the content of the original reviews.
The degree of importance of information in a given document
The degree to which the content of different summaries of the same editorial overlap.
The process of planning the content of a generated text.
The quality of the content in a summary
Measures used to evaluate the quality of a summary, such as relevance, coherence, and completeness.
The degree to which a summary accurately captures the information in the original text
A module designed for coarse filtering that retains the most promising candidates from a sheer number of sentences in the original document.
A phenomenon in abstractive summarization where the contents of the original document may be reordered in its summary.
The process of selecting important content from the source text to generate a summary.
A system that decides on relevant aspects of the source document by identifying tokens from a document that are part of its summary.
Similar text segments that form a content unit, where the contributing text segments of a content unit should have similar semantic meanings
Errors in the content of the summary that can be verified by checking the source document.
a new evaluation metric based on abstracting away from the particular surface form of the target summary, but representing it as facts using Semantic Role Labelling (SRL)
Interprets textual sequential meaning on the Transformer.
A method of formulating extractive summarization that greatly reduces the size of the space that must be explored, removes the need to perform supervised pre-training, and prevents systematically privileging earlier sentences over later ones.
A problem where the model takes an action which is a to-be-selected sentence set and then receives a reward based on the correlation between extractive summary and gold-standard reference summary.
used for alignment of facts present in the source document according to the facts selected by a human-written summary
Information from the source article that is lost in post-processing based approaches
A model used in the data-driven approach that learns a latent soft alignment over the input text to help inform the summary
A network responsible for extracting and compacting the source document in a summarization model
A method used to build a content selection model that can identify correct tokens with a recall of over 60%, and a precision of over 50%.
Adding the appropriate context from the reference article to the citation texts to better understand the context for the ideas, methods or findings stated in the citation text.
A model that can be finetuned based on the deep bi-directional Transformer for content selection in summarization.
A type of neural network model that is pre-trained on large amounts of data and can be fine-tuned for specific tasks.
A representation that takes the whole document into consideration to learn the document-level context.
Word embeddings that capture the meaning of a word in context
The approach used in this paper to rely on continuous latent representations, in contrast to MeanSum, which treats the summary itself as a discrete latent representation of a product.
A method for correcting hallucinations in which named entities in a potentially hallucinated summary are replaced with ones with compatible semantic types that are present in the source, and variants of candidate summaries are created and ranked with a discriminative model trained to distinguish between faithful summaries and synthetic negative candidates generated given the source.
A mechanism that encourages the contribution from the conventional attention that attends to relevant parts of the source sentence, while penalizing the contribution from an opponent attention that attends to irrelevant or less relevant parts
Clearly incorrect summaries generated using a rule-based procedure to test how well neural abstractive summarizers distinguish human-written abstracts.
Summaries that highlight the differences between two entities.
A module that encourages the model to better differentiate factual summaries from nonfactual ones by paying attention to the document using contrastive learning.
Codes used to condition summary generation based on sub-aspect functions, such as importance, diversity, and position.
A proposed model for abstractive summarization that incorporates entity information and generates summaries with selected entities, resulting in improved content accuracy and topic coherence.
A model that enables personalized generation of summaries and allows the reader to control important aspects of the generated summary, such as length, focus on entities, style, and portion of the article to be summarized.
Algorithms that allow for controlling various dimensions of the output summary, such as length, entities, and topics.
The ability to generate summaries that conform to specific topic distributions or sentiment polarity.
Text that scatters main points across multiple utterances and between numerous writers.
A method of summarization that formulates the problem as a decomposable row-sparsity regularized optimization problem.
Associates each word with a topic vector capturing whether it is representative of the document's content.
A unit used in global encoding to perform global encoding on the source context
A type of neural network architecture that uses convolutional layers instead of recurrent layers, which can be faster and more stable when processing long sequences.
Models that have scores that are coordinated with the actual quality metrics by which the summaries will be evaluated – higher model scores should indicate better quality summaries.
Aggregations that can be directly copied from the source text
A bias observed in pseudo summaries generated from a Seq2Seq teacher model, where more continuous text spans from original documents are copied than in reference summaries.
A neural component that allows a summarization system to copy words or phrases from the input text to the summary
The percentage of summary n-grams (sequences of words) appearing in the source text.
Tokens in the generated summary that are copied directly from the input document
An indicator of a word's saliency in terms of forming an impression, which can be learned via a sequence-tagger.
Relations between mention phrases of the same entity
The bias towards certain sub-aspect functions in different types of documents.
An estimation of the accuracy of generated summaries in abstractive summarization.
A module that removes hallucinations existing in reference summaries, allowing training on the full training set without learning unfaithful behaviors.
The process of comparing system-generated texts with human-generated reference texts to evaluate the quality of the system
The degree to which the generated summary matches the judgment of a human evaluator.
A technique used to evaluate the model's performance with different subsets of the input data.
A method for estimating the causal effect of language prior on the generated summary and removing it from the total causal effect
A proposed evaluation metric for text summarization that evaluates factual consistency via counterfactual estimation and does not rely on auxiliary tasks
Extrinsic hallucinations that introduce information that is not true in real life.
Added to prevent repetitions on long summaries
A model that takes as input the original document with keywords masked out and uses the current best automatically generated summary to try to uncover the missing keywords
A report generated by a script reader consisting of a logline, a synopsis, comments explaining its appeal or problematic aspects, and a final verdict as to whether the script merits further consideration
Used to track and control coverage of the source document, remarkably effective for eliminating repetition
A regularizer that helps to ensure that all parts of the source text are covered in the summary.
The relationship between the prototype document-summary pair used to obtain a summary pattern and prototype facts.
Applications where the summarization model is tested on data from a different domain than the one it was trained on
Evaluating a summarization system on a range of out-of-dataset corpora to test its generalization ability
A scenario where the model is tested on a dataset that is different from the one it was trained on.
The ability of the proposed model to perform well on another publicly available clinical dataset (OpenI).
A core step in extractive document summarization, where the relations between sentences are modeled to effectively extract summary-worthy sentences.
Particular biases in the data that the summarizer can learn to exploit.
The ability of a model to perform well with a small amount of training data
Includes synthesis and augmentation
Incomplete or irrelevant information in a dataset that can negatively impact model performance.
The degree to which data is accurate, reliable, and consistent, and must be evaluated either during dataset construction or post hoc.
A method of summarization that selects sentences that can best reconstruct the original document.
The limited availability of high-quality data for training and evaluating automatic summarization systems.
A property of the content selection model that can be trained with less than 1% of the original training data, providing opportunities for domain-transfer and low-resource summarization.
The tendency of a summarization system to perform better on certain types of datasets due to biases in the training data
The issue that released datasets do not contain both annotation and social information, making it challenging to evaluate summarization systems that consider both sentences and social messages.
A proposed method to effectively demote the lead bias learned by the neural news summarizer and improve its generalizability.
A desideratum that suggests the selected sentences should lead to the same decision as using the full text based on the model.
A summarization task that emphasizes supporting decision making by identifying the most relevant information for decisions.
A novel architecture that adds another ‘closed book’ decoder without attention layer to a popular pointer-generator baseline, such that the ‘closed book’ decoder and pointer decoder share an encoder.
A method suggested in a previous study to regulate the overconfidence of the decoder in a summarization model
Part-of-speech tagging (POS), Dependency Labeling (DEP), Semantic Role Labeling (SRL), and Named Entity Labeling (NEL) used to explore the information encoded in the role, filler, and TPR space.
The entropy of decisions made by the model during generation
A type of abstractive summarization framework that uses only a decoder, as opposed to an encoder-decoder architecture.
Reinforcement learning methods that use feedback to improve the quality of generated summaries
a neural extractive summarizer that estimates salience for guiding the extraction procedure instead of learning an end-to-end mapping
a lightweight method for post-editing extractive summaries, which involves predicting whether articles in the summary should be kept as is or modified
The process of discussing and considering different options or viewpoints in order to make a decision or reach a conclusion
A graph built from sentences to decode a tree using integer linear programming, which is finally linearized to generate a summary sentence.
A graph that represents the syntactic structure of a sentence or set of sentences.
A rule-based sentence compression module that operates on the dependency parse of the answer sentence can yield better results than query-based extractive summarizers trained for the specific dataset.
A structure that naturally combines with the copy mechanism of an abstractive summarization system to encourage salient source words/relations to be preserved in summaries.
The relationships between words in a sentence that ensure grammaticality in compressed sentences.
The factors that need to be taken into account when designing human-AI interaction in text summarization and broader text generation tasks.
The potential areas for improvement in the design of human-AI interaction in text summarization and broader text generation tasks.
Properties of a summary such as capturing the most important information, being faithful to the original text, grammatical and fluent.
Optimization method that selects a diverse subset from a ground set of items, characterized by quality and diversity scores
Computerized systems that assist healthcare providers in accurately understanding a patient’s condition and reducing the effort in document review during time-sensitive hospital events.
Computer systems that can engage in conversation with humans.
The trend of diminishing benefits from intermediate pretraining as the amount of pretraining data increases
An approach for measuring centrality in single-document summarization that uses directed edges and considers the relative position of nodes
Augmenting the document graph with directionality and hierarchy to reflect the rich discourse structure of long scientific documents.
The unit of information used to relate sentences in automatic summarization. They are referred to as head nouns of noun phrases.
Words or phrases that connect sentences or parts of sentences to show the relationship between them.
Techniques used to analyze the relationships between sentences in a text
Relations between Elementary Discourse Units (EDUs) within a document
Patterns in the structure of long scientific documents that are highly useful for determining sentence importance in summarization systems.
A method of representing text as a tree structure, where each node represents a sentence and its children provide additional information about the parent sentence
A model component that generates the summary and attends to different discourse sections
A method of ranking the importance of each sentence in terms of the number of descendants to generate a summary that focuses on the main review point
Probability distributions that take on a finite or countably infinite number of values
The ability to distinguish between relevant and non-relevant documents
The variables used in the deterministic transformations in existing seq2seq models, which lead to limitations on the representation ability of the latent structure information.
A transformer-based model that scores the factuality or discourse quality of candidate summaries using one of four different objectives
Improving human judgments of linguistic clarity and referential structure.
Interruptions, hesitations, and filler words such as "um" and "uh-huh" that occur in speech
A model that computes two scalar values, one from a content feature and the other from a context feature, to predict the popularity of a post and exclude the effect of context.
Constraints used to ensure the specific embeddings in each feature space.
Labels used to train models for summarization tasks that do not require manual labeling, such as categories of news articles and ratings of online reviews.
A pre-training method that uses filtered sentences of the documents as noisy targets to pre-train all the parameters of the NHG model.
A training strategy that uses external knowledge sources to provide supervision for the model, allowing it to adapt and generalize better.
A method of optimizing the information selection process
A reward signal that measures the similarity between the generated summary and the input based on their distributional semantics
The subject and structure of the document that the system should have a global view of in order to decide whether to choose a particular sentence.
A hierarchical architecture that suits the compositionality of documents
A layer in the model that encodes the document
A module that learns representations of documents.
The model consists of a document reader (encoder) and a sentence extractor (decoder).
A method for segmenting the document into facet-aware semantic blocks
The way a document is organized, including sections and paragraphs, can facilitate information searching, reading comprehension, and knowledge acquisition.
Topics, categories, sentiments, and other meta-information about a document.
Summaries of documents created based on a model of relevance for the topic
Generating a summary for a long document or multiple documents on the same topic
Features that depend on the specific document being summarized, such as term frequency or position.
Features that do not depend on the specific document being summarized, such as stopword ratio or word polarity.
A feature that captures global information and plays a key role in sentence selection for summarization.
Terminology specific to a particular domain.
differences in the language and terminology used in different domains or fields
A pre-training method based on an unlabeled substantial domain-related corpus.
The effectiveness of summarization models varies depending on the domain of the text being summarized.
Resources that are customized for a specific domain, such as chemistry, to aid in NLP pipelines.
Enables dropping out hallucinated entities from the predicted content plan and prompting the decoder with this modified plan to generate faithful summaries
Alleviates the problem of generating repetitive words and incomplete summary, allowing the model to track the comprehensive information typically for each salient facet within the source document.
Focus on when to output eos (end of sequence), indicating the end of the summary. An ad-hoc method generates the eos by assigning a score of −∞ to all candidate words at the position of the desired length during test. Others learn the relationship between length and the decoder state at training time.
The degree to which a reference summary is easy to summarize.
A method that maximizes a heuristically defined scoring function to evaluate the quality of the generated summary
A technique that generates a summary by operating an edit action (e.g., keep, remove, or replace) for each word in the input sentence
The three parts of an editorial - lead, body, and conclusion - each with a specific contribution to the overall argument.
Improvement in faithfulness over a baseline system operating at the same level of extractiveness
Summaries that convey the most important information in a text
Comparisons with multiple abstractive and extractive baselines, including traditional syntax-based systems, integer linear program-constrained systems, information-retrieval style approaches, and statistical phrase-based machine translation
The effectiveness of selecting sentence singletons and pairs for abstractive summarization.
The ability of the AI-assisted text generation system to generate summaries quickly and accurately.
The idea that the effectiveness of content consumption is not only determined by the information contained within it, but also by the tone and style of presentation.
The process of creating a shorter version of a longer document while retaining its key information.
A benefit of curriculum learning without the need for external data
Vector representations of text that encode meaning and allow the application of statistical and geometrical methods to words, sentences, and documents.
A framework for generating abstractive summaries using two complementary models, transformer and seq2seq.
The part of the model that encodes the source passage to a fixed-size memory-state vector.
A component of the model that processes the input sentence and produces a representation that is used by the decoder
Proposed models for extractive summarization using structured transformers that enable stepwise summarization by injecting the previously planned summary content into the structured transformer as an auxiliary sub-structure.
A type of neural network architecture used for sequence-to-sequence learning
a common approach for abstractive summarization based on encoding the original text sequence and decoding the summary sequence
Methods for abstractive summarization that use a neural network to encode the input and decode the summary
A type of attention mechanism used in sequence-to-sequence models where the decoder attends to the encoder's hidden states.
a framework that encodes a document and decodes its summary
A method widely used for single document extractive summarization, where the encoder encodes one sentence into vector representation and the decoder with top-k strategy predicts the probability scores of those sentence vectors, sorts sentences in descending order, and picks sentences until exceeding the length limit.
Modern neural summarization systems that aim at producing abstractive summaries and rely on the attention mechanism to focus on different parts of input during the decoding stage.
Benefit several tasks including document classification, natural language inference, and machine translation
The procedure of representing a document into high-dimensional representations
Recent work used to instantiate the theoretical graph framework for conversation summarization.
A type of training where the entire system is trained together, rather than training individual components separately.
A framework that directly learns to detect summary-worthy content as well as generate fluent sentences, circumventing efforts in feature engineering and template construction.
Gives higher weight to summaries whose sentences logically follow from the ground-truth summary.
Summaries whose sentences logically follow from the ground-truth summary.
Knowledge that ensures a correct summary is semantically entailed by the source sentence
A model that uses entailment RAML training to encourage the decoder of the summarization system to produce summary entailed by the source
A model that incorporates entailment knowledge into summarization models by jointly modeling summarization generation and entailment recognition
Metrics that determine whether the content in the summary is entailed by the input document.
Ordered sequences of entities in the summary
A desirable summary aspect that is encouraged in reinforcement learning approaches.
Designing constraints that guide the generated summary to cover the salient information of user-specified entities
A technique that guides the model learning process by utilizing entity coverage precision between the training document and its reference summary as faithfulness guidance
A graph used for summarization where one set of nodes corresponds to entities
A local coherence model used for summary coherence evaluation
A coherence modeling method that constructs a grid to represent grammatical and semantic transitions of entities between sentences.
A problem in text summarization where a model-generated summary contains named entities that never appeared in the source document.
The type of hallucination that occurs when the generated summary contains entities that do not exist in the source document
A system used to extract linked entities from the original text
Nodes representing entities in the document
A method of sentence fusion that involves replacing entities in the original sentences with pronouns or other entities
A component of the proposed controllable neural model that identifies the most important entities and sends their representations to the summary generation phase.
An NLP task on TV show transcripts that focuses on identifying and tracking entities
Selects important sentences from the input that includes references to salient entities
Enables enhanced input text interpretation, salient content selection, and coherent summary generation
A majority of work on opinion summarization is entity-centric, aiming to create summaries from text collections that are relevant to a particular entity of interest, e.g., product, person, company, and so on.
Metrics proposed to evaluate the quality of generated plot summaries based on bags of characters and character relations
A measure of the amount of uncertainty or randomness in a system, used in Peyrard's method to model importance
A model used to treat extractive summarization as a multi-step process that is aware of the extraction history.
A mechanism that automatically detects errors in the generated after-visit summary, including missing medical events and hallucinations.
Issues faced by autoregressive models.
an open challenge due to decoders being amenable to pathogeniessuch as hallucination and/or omission of important information, which are hard to capture using existing evaluation metrics
The paper discusses the challenges of evaluating the quality of summaries and the importance of human evaluation as the gold standard.
A method of answer verification that compares the prediction to the expected answer by exact match
A type of automatic evaluation metric that counts the number of matching n-grams between the generated summary and a reference summary
A paradigm that subsumes generic, query-based, question-based, and even abstractive summarization systems
A method of summarization that introduces an additional sentence dissimilarity term to encourage diversity in summary sentences.
The process of identifying all possible content points in a text passage.
A technique that leverages a small pool of strong sampled candidates to smartly inform the reward function
The process of identifying the reasons behind a label
Specialized modules added to the factual consistency model that explain which portions of both the source document and generated summary are pertinent to the model’s decision.
Supervision signals that prevent the model from violating the specified attribute requirement
Incorporates external linguistic structure (e.g., coreference links)
A problem where the model is trained with teacher forcing, which leads to a discrepancy between training time and inference time.
Systems that are difficult to control to produce correct and fluent output.
A model that selects important content with lexical features and allows aggressive compression of individual sentences by combining two different formalisms.
A summary of a long document, typically containing 400-600 terms, that conveys more detailed information than a short summary
A large scale set of experiments comparing different approaches to summarization.
Data sources such as Web search results, clickthrough data, query logs and Wikipedia, comments from news readers, and tweets corpus used for improving summary quality.
Representations of external knowledge that can be used in summarization
an extractive model first selects a subset of opinions and an abstractive model then generates the summary while conditioning on the extracted subset
A method that first extracts the summary-worthy sentences and then abstracts each of them, but suffers from an information loss in abstract stage and lacks an effective reinforcement learning framework to bridge together two modules.
A subaspect of summarization that is determined by how well an extractive model performs on the data, compression ratio between the source document and summary, and lead bias.
The process of automatically extracting a set of sentences that represent the information of a whole document by ranking the importance of sentence features.
A model that combines representations of the topic and partial summary with representations of the document sentences through an attention mechanism to extract one reference sentence.
A component of the hybrid architecture that selects salient sentences or highlights from the original document.
A new single-document summarization task that requires an abstractive modeling approach and aims to create a short, one-sentence news summary answering the question "What is the article about?"
Producing one to two summary sentences in extreme compression and high abstraction.
Refers to errors in summaries that are related to the source text and errors that are unrelated to the source text, respectively.
The type of hallucination that occurs when the generated summary contains information that is present in the source document, but is not relevant to the main content
The problem of centrality-based models tending to select sentences from one facet of a document, rather than important sentences from different facets.
A proposed evaluation metric for summarization that measures whether the system summary covers the facets in the reference summary
A model proposed in this paper to address the facet bias problem in document summarization.
Mappings from each facet (sentence) in the reference summary to its support sentences in the document
A method that forces a centrality-based model to select summary sentences from different facets by incorporating the relevance between the candidate summary and the document
A measure of information coverage in summarization based on facet overlap
Summarizing an article from distinct aspects including purpose, method, findings, and value.
A dataset consisting of 60,024 scientific articles collected from Emerald journals, each associated with a structured abstract that summarizes the article from distinct aspects including purpose, method, findings, and value.
A mechanism designed to prevent the generator from copying irrelevant facts from the prototype by providing mutual information between the generated summary and the input document.
Fact checking focuses on verifying facts against the whole of available knowledge, whereas factual consistency checking focuses on adherence of facts to information provided by a source document without guarantee that the information is true.
A framework proposed to assess the factual consistency of summarization models
Short sentences formed by merging words in a triple or tuples to represent a fact.
A task that aims to verify the truthfulness of a given statement.
A type of question answering system that can be answered by a certain word or a short phrase.
Inaccuracies in the information presented in a summary compared to the source document
The process of verifying the credibility and usability of models by checking for factual consistency
The paper introduces entailment-based and semantic area RL rewards to analyze their effect on factual consistency and semantic coverage, ensuring that all factually relevant perspectives are captured.
A machine learning model trained to identify factual inconsistencies in summaries.
A model used to evaluate the factual consistency of a summary
Improvement in the factual consistency of summarization models
Existing abstractive summarization models are optimized to generate summaries that highly overlap with human references, but this does not guarantee factual correctness. Maintaining factual correctness of generated summaries remains a critical yet unsolved problem.
Can be objectively annotated with detailed classification of factual errors
A graph that contains all the entities with their relationships mentioned in the documents, extracted from the original documents to support the summarization task.
A model capable of leveraging fine-grained human annotations to detect errors in generated texts.
A metric that shows the degree of factuality degradation observed for different models.
A model used to evaluate the factual consistency of a generated summary
Measurements used to evaluate the faithfulness of abstractive summarization.
Models trained on synthetic data to detect and correct errors in generated summaries.
The accuracy of the generated summary in reflecting the content of the source document.
Summaries that accurately reflect the content of the source document.
Generating information in the summary that is not present in the original text
A measure of how well a summary conveys the same meaning as the original document
The quality of accurately representing the content of the source text
The degree to which a summary accurately conveys the main concepts of the original text.
An algorithm developed to address the problem of multiple optimal solutions and achieve near-optimal performance without the need for concept pruning.
Techniques used to understand the contribution of input features to the output of a model.
The process of selecting and designing features to be used in a machine learning model.
The ability of a model to perform well on a task with only a small amount of labeled data.
A type of content selection that guides the neural text generator to stitch selected segments into a coherent sentence.
Annotations at the word- or span-level used to detect errors and localize them within generated texts.
A metric developed to evaluate the output quality of the proposed method in semantic vector level, assessing not only the relevance between model outputs and manual references but also the extra and omissive information in generated summaries.
The process of adapting a pre-trained language model to a specific task using supervised learning.
A method of content selection that can benefit from splitting documents into sub-sentential segments following its discourse structure, as it provides a more refined granularity of semantic information.
Finetuning the model using a labeled dataset for the target task
Demands extra focus on controlling the length of output summaries, specifically, it requires generating summaries with a preset number of characters or words
The issue of a fixed number or proportion of summary sentences in the top-k strategy, which has been a popular method employed by many models. A flexible extractor should generate a non-fixed number of summary sentences based on source document length, topics, or other aspects.
A type of summarization model that does not take into account the word-sentence hierarchy.
The ability to customize content selectors to be combined with general-purpose neural text generators.
Structured transformers are flexible in terms of content type (e.g., text or tables) that can be modeled.
The proposed approach is easily adaptable from single-document to multi-document summarization tasks.
The ability of the architecture to extract any number of sentences with a threshold, instead of a constant number.
A model that ensures the generated language is fluent
A technique for improving document summarization by focusing on supported and topical content.
A technique for generating diverse yet topically consistent and faithful summaries.
A mechanism that employs a learnable Gaussian focal bias as a regularization term on attention scores to emphasize the corresponding part of the document
Produced by the model according to human evaluations
Summaries generated by the same model, tailored towards different levels of formality.
A stylistic parameter that can affect the tone and style of text, with formal language being more appropriate for certain audiences, such as corporate executives.
The degree to which the generated text conforms to a specific format.
The different ways in which a summary can be presented
A discussion initiated by a user on an online forum
The task of generating concise summaries of lengthy and comprehensive forum threads
The authors propose a framework where an external information extraction system is used to extract information in the generated summary and produce a factual accuracy score by comparing it against the human reference summary. They further develop a training strategy where they combine a factual correctness objective, a textual overlap objective, and a language model objective, and jointly optimize them via reinforcement learning (RL).
A type of generation where the model is not constrained to copying from the input document and can generate new text
Failing to understand the sender's intent and failing to identify the roles of the sender and receiver
Common data mining techniques for calculating frequent sequences of words in transactional data.
The original pre-training objective of PEGASUS that transforms any text into a pseudo-summarization dataset by selecting important sentences using ROUGE as output summaries.
A part of the information selection layer that filters unnecessary information in the original document
A type of activation function used in neural networks that can help prevent the vanishing gradient problem.
a bias term used in CoCoNet to model positional correlations between source words, considering the relative distances between them and the scope of the local context when copying.
A type of summarization that is different from scientific summarization in three main aspects. First, the length of scientific papers are usually much longer than general articles. Second, in scientific summarization, the goal is typically to provide a technical summary of the paper which includes important findings, contributions or impacts of a paper to the community. Finally, scientific papers follow a natural discourse.
The ability of a model to generate summaries that are applicable to different domains.
The ability of the model to perform well on different datasets and scenarios
Improved by saliency and entailment skills.
The ability of a model to perform well on new, unseen data.
A model that replaces the hard copy component of the pointer generator architecture with a soft "editing" function by learning a relation embedding to transform the pointed word into a target embedding.
The probability of generating the next word in a summary
The style in which the summary is generated, which may differ between domains.
The process by which neural summarizers generate summary text.
A component of the proposed pipeline framework that is trained to reconstruct reviews from their extracted opinion phrases and can then generate opinion summaries based on the selected opinions.
A summarization method that creates a single summary for a document, not taking into account user preferences or control aspects.
Describes how all reference tokens should assign the attention to each source token and can be predicted from the source by training an attention-prediction model.
The overall coherence and consistency of the summary
A method for summarisation that considers the entire document at once
The overall information contained in a document
A model proposed to tackle the problems in attention mechanism in abstractive summarization
Methods that formulate summarization as a combinatorial optimization problem, selecting a subset of sentences that maximizes an objective function under a length constraint, and use Integer Linear Programming (ILP) to solve it exactly.
A suggested training method for abstractive summarization models
Proposed method for neural summarization models to control the output length better than existing methods
A hypothesis of the highest probability among all possible sentences, consisting of words in vocabulary V.
A novel mechanism composed of attention scores and length rewards to guide beam search based on the predicted global attention distribution.
The overall meaning and context of a document.
A loss function introduced to directly optimize the attention from the global perspective by preventing assigning high weights to the same locations multiple times.
High-quality summaries created by expert writers, requiring a lot of effort
The summary of a document that is considered to be the most accurate and informative.
Annotations that are considered to be the most accurate and reliable.
A network that simultaneously updates the representations of sentence and topic nodes in a heterogeneous document graph.
A method that pushes close topic representations of documents and sentences that have high semantic similarity with the gold summary and pulls away otherwise.
A model that incorporates the graph contrastive topic model (GCTM) empowered by the semantic information of the gold summary and the global document context, with PLMs for long document extractive summarization.
A technique used in the decoding procedure of the graph-based encoder-decoder model to guide the summary generation process by incorporating the graph structure.
A structured representation that produces a summary and highlights the proximity of relevant concepts.
Intuitive way to model long-range dependencies among text spans throughout a document
Methods for summarization that use graph-based representations of the input
A novel approach inspired by graph-based extractive summarization methods, introduced in the encoder-decoder framework to discover the salient information of a document.
A model that improves both the document representation and summary generation process of the Seq2Seq architecture by leveraging the graph structure.
A method to impose structure information on BERT directly to jointly learn contextual representations of different text granularities within a single BERT
A method that uses sentences as nodes and weighted edges to represent the degree of similarity between sentences.
A technique for summarization that uses graphs to represent the relationships between sentences
A method of adding one sentence at a time incrementally to a summary, with the goal of maximizing the ROUGE score and stopping when no remaining candidate sentences improve the score.
A method used to solve the maximal coverage problem by ranking sentences by their coverage of best compressing frequent word sequences and selecting the top-ranked sentences to a summary.
Summaries that are closely tethered to the original audio, providing a preview of notable podcast clips and reducing the risk of misleading or inaccurate information.
Factors that arise in summarization evaluation when one annotator rates multiple summaries
Specialized techniques employed during training to guide the model away from pathological behavior, including reducing repetition, encouraging the model to complete sentences, and avoiding frame filling patterns
Input to enhance factual consistency of the summary
Methods that provide various types of guidance signals to constrain the summary so that the output content will deviate less from the source document and allow for controllability through provision of user-specified inputs.
Signals extracted from the input source document such as keywords, highlighted sentences, and others to aid the model architecture in summarizing the input document.
A model proposed in this paper that combines extractive and abstractive models and uses keywords as guidance for the latter
Quantities that are not supported by the source text and are introduced by abstractive summarization models
A known shortcoming of current text generation and summarization models, which has been established for both abstractive and extractive summarization models.
A measurement of the risk of generating false or inaccurate information in the summary
Features that are manually designed by experts.
attention methods that perform abstraction only on text regions that were initially selected by some extraction process
0/1 labels used in traditional training methods.
A proposed approach for extractive summarization that introduces more semantic units as additional nodes in the graph to enrich the relationships between sentences.
Aspects of text semantics that can be derived from word embeddings computed from a general corpus.
A method for question-driven extractive answer summarization that integrates hierarchical interaction information between question-answer pairs in both word-level and sentence-level into a sequential extractive summarization model.
A technique used in the student model of DistilSum to encode sentences into hidden vectors.
A novel supervised thread summarization approach that learns effective sentence and thread representation by attending to important words and sentences
Learnable biases that adjust attention weights between tokens based on their relative positions with regard to the document structure, enabling summarizers to capture long-range relatedness for better document understanding.
A type of decoder architecture that generates a summary from latent sentence representations.
A proposed method to tackle the constraints of saliency, nonredundancy, information correctness, and fluency under a unified framework.
A model component that captures the discourse structure of the document
A model proposed in this work that consists of a summarization layer and a sentiment classification layer.
A graph constructed from a document where both words and sentences are nodes and the relations between them are constructed as different types of edges.
Into word embedding and paragraph attention, to expose the critical words and paragraphs for summarization.
A model proposed by Cheng and Lapata (2016) to encode a document and predict binary labels for each sentence
Deep learning based sequence learning methods for summarization
End-to-end deep learning framework leveraging encoder-decoder transformer architecture
Required to comprehend the hierarchical multi-facets information within the article, encoding the temporal dependencies with different timescales.
Captures multiple hierarchical levels of abstraction of the source document, encoding the temporal dependencies with different timescales.
Networks that model a document as a sequence of sequences to tackle the challenge of modeling long-range inter-sentence relationships for summarization.
A neural network used for rough reading which consists of a neural net to encode the whole document and another one to capture features in paragraphs.
A new summarization task that produces hierarchically organized question-summary pairs to facilitate information consumption, inspired by the top-down knowledge learning process.
Incorporates topic information into words embedding and paragraph attention.
A model that fully embeds the global context of long documents, to inform topic representations of documents and sentences.
The crucial criterion for selecting a good evaluation measure
Data that is accurate, reliable, and consistent, and forms the foundation for building meaningful statistical models in NLP.
A metric for which a small difference in score reliably indicates a difference in quality
A summary that is shorter than the original document, conveys only the most important and no extraneous information, and is semantically and syntactically correct
Text operations such as paraphrasing, generalization, text reduction, and reordering that pose a considerable challenge to natural language understanding.
The range of scores where summarization systems aim to perform well.
The model can holistically learn any document-level coherence properties, such as saliency, redundancy, and ordering, embodied in the gold summaries.
Used to derive labels for extraction units, but not necessarily the best method due to generalization, paraphrasing, and words not present in the source text.
The process of manually marking or labeling data by humans.
Two evaluations conducted to assess which type of summary participants prefer and how much key information from the document is preserved in the summary.
A method of training language models by collecting human preferences between pairs of summaries and using reward learning to fine-tune the model.
The gold standard for evaluating the quality of generated summaries, but is time-consuming and labor-intensive.
Evaluation of the generated summaries by human judges.
People who evaluate the quality of summaries.
Summaries written by humans that are used as a benchmark for evaluating system-generated summaries.
Can improve summarization performance by taking advantage of complementary strengths.
The interaction between humans and artificial intelligence in the context of text summarization and broader text generation tasks.
A method of incorporating human input to improve the performance of a system.
A concept that allows humans to actively participate in supervising AI systems by approving, rejecting, or re-labeling current outputs, and providing expert-guided advices to the system. It also acts as the unique source of external knowledge from humans.
A three-phase process that includes general understanding of the document, task-specific reading comprehension, and polishing process. It involves leveraging prior knowledge, making inferences, and evaluating and polishing the generated summary.
A representation of the input source text in an auto-encoder architecture that complies with human grammar and can be comprehended by humans
Typically high quality, but require heavy cognitive load due to limited time and energy.
A combination of both extractive and abstractive summarization methods, using policy-based reinforcement learning to bridge the two networks.
A novel approach that mimics the human-like reading strategy for abstractive text summarization. It consists of three components
A data-driven, end-to-end enhanced encoder-decoder based deep network that summarizes a news article by extracting salient sentences.
Text summarization methods that combine both abstractive and extractive approaches.
Facilitates copying words from the source text via pointing, which improves accuracy and handling of OOV words, while retaining the ability to generate new words
A combination of token-level and sentence-level representations used to encode text.
A model used in the document summarization process to import both structural-compression and structural-coverage regularizations.
Combines coreference based dependency graph and latent structure attention module output
A combination of extractive and abstractive methods that employ specialized components.
Summarization methods that combine extraction and abstraction techniques
The process of selecting the best parameters for a model.
A task of selecting the best output from multiple models, related to text-generation tasks.
Statistical tests used to determine whether the difference between two metrics' correlations is reflective of one metric being better than the other or if it is an artifact of random chance.
Texts are divided into IUs to deal with complex sentences that convey multiple ideas. The IU is defined as a minimal fragment of a text that conveys an "idea" or "thought" coherently.
Stressed by the poor correlation between automatic metrics and human judgment
The paper proposes the use of "the probability of being quoted" as an alternative aspect for summarization. The ability to extract quotes can help extract important sentences irrespective of how frequently the same topic appears in the text.
A section of the radiology report that summarizes the most prominent observation.
A novel loss function to encourage the consistency between two levels of attentions in the unified model
Incorporating source syntactic structure in neural sentence summarization to help the system identify summary-worthy content and compose summaries that preserve the important meaning of the source texts.
Information in the summary that is not accurate or true to the source text
A measure of centrality that assumes a word receiving more relevance score from others is more likely to be important.
Summaries that require the reader to infer information not explicitly stated in the original text.
The time it takes for a model to generate an output.
A mechanism proposed in this work that works on encoder-decoder attentions to underscore salient content from the source and improve the quality of abstractive summaries.
Techniques used to understand the effect of individual training examples on the model's output.
The focus of the methodical empirical evaluation, which assesses the summary quality with reading comprehension tasks and compares favorably with automatic metrics against state of the art.
The process of selecting important information from a document for summarization
Two-stage methods that extract top l most important tokens from the source document as a prototype summary where l is the desired length, and in the second stage encodes the original source document and prototype summary by a dual-encoder.
Measures used for controllable summarization, such as relevance and redundancy.
A technique used to understand the contribution of input features by computing the gradients of the output with respect to the input.
A novel method for training models with stylistic feedback on sampled and ground-truth summaries together.
An RL-based summarization system that employs a handcrafted reward function to learn a policy specifically for a given input, without requiring parallel data or a reward oracle.
The degree to which different automatic metrics agree in ranking summaries.
A reasoning module in the encoder that infers useful information from historical summaries to learn a more comprehensive representation for the given input review.
The ability to connect sentences in a summary to create a coherent text
A framework that continuously collects user-feedback to improve model prediction robustness. The user's intention is implicitly acknowledged as a factor influencing the extraction of important sentences from the source documents. The AI model is trained with human-produced summaries and adapted as more human-feedback is fed in.
Users can query relationships with a concept map interface
A method that efficiently gathers user feedback and combines it with predictions from pretrained, generic models to solve text ranking tasks.
Supervised pretraining using labeled datasets from different domains for a task that is related to or is the same as the target task
A visualisation of the semantic coverage of a generated summary by visualising the transport plan between summary tokens and document tokens.
A personalized attention mechanism that selects informative words in the input review.
Attention that records previous attention weights for each of the input tokens
The type of hallucination that occurs when the generated summary contains information that is not present in the source document
Metrics that are based on the properties of the dataset itself, rather than external factors, and are used to evaluate the quality of summarization datasets.
The problem of performing the two tasks of relevance ranking and saliency ranking in isolation.
including problems with verbosity and coherence, as well as coreference and pragmatic context issues
A process of refining sentence representation by fusing redundant information between selected sentences through iterative refinement, which is supervised by knowledge distillation.
Iterative Text Summarization is an iteration based summary generator which uses a sequence classifier to extract salient sentences from documents. It consists of a novel “iteration mechanism” and “selective reading module”. ITS is an iterative process, reading through the document many times. There is one encoder, one decoder, and one iterative unit in each iteration. They work together to polish document representation. The final labeling part uses outputs from all iterations to generate summaries.
A news structure commonly used in journalism where a narrative hook is presented first to catch the reader's attention, followed by the main story presented in a Synopsis and Body Section.
Short list of high-level arguments that summarize the pros and cons of a proposal. Salience of each key point is represented by the number of matching arguments. Can be composed by domain experts without looking at the arguments themselves. Used for measuring the distribution of key points in a large collection of arguments, interactive exploration of arguments, and novelty detection. Proposed method demonstrates feasibility and effectiveness of summarizing a large set of arguments by mapping them to a small set of key points. ArgKP dataset developed for argument-to-key point mapping task, comprising about 24,000 (argument, key point) pairs labeled as matching/non matching. First dataset for this task. Empirical evaluation and analysis of various classification methods performed.
a type of summarization that incorporates knowledge structures to improve the quality of the summary.
The process of constructing a graph that captures interactions among entities in a document.
The delay between the news exposure and the tweets linking to it, which might affect the summarization performance.
Alignment relations that cannot be identified by the standard pointer generator.
A discourse tree that is induced without a parser
A model for topic modeling where topic probabilities are assigned words in documents
Topic modeling method for summarization
A representation of chemical texts produced by deep learning models.
A higher-level representation of the input document that captures the main ideas and concepts.
The inherent structures that people may naturally follow when they write abstractive summaries, such as "What", "What-Happened", "Who Action What", etc.
A model that views labels of sentences in a document as binary latent variables and directly maximizes the likelihood of human summaries given selected sentences
Sentences in news articles that contain key information and are placed early due to journalistic conventions
A baseline method for summarization that selects the first three sentences of the original text as the summary.
A bias observed in pseudo summaries generated from a Seq2Seq teacher model, where the summaries tend to summarize the leading part of a document.
Algorithms that approximate the ground-truth reward oracle from weak supervisions, such as numeric scores indicating the quality of the summary or preferences over summary pairs.
A decoder that does not have complete context when predicting each word
An effective length controlling unit that allows summarizers to generate high-quality summaries with a preset number of characters, breaking the trade-off between length controllability and summary quality
The summary should be as long as the width of target devices such as smart-phones and digital signage
Dividing summary length into disjoint length bins and restricting the summary length according to the desired length bin
The ability to generate summaries with a preset number of characters or words
Added to the Entail reward to avoid misleadingly high entailment scores to very short sentences.
A metric used to evaluate the ability of a summarization model to generate summaries of different lengths.
Extends a transformer seq2seq model with the ability to select information in the context according to the length constraint. LAAM re-normalizes the attention between encoder and decoder to boost the tokens with higher attention scores based on the desired length, helping with selecting length-aware information from source document.
A dataset created by predefining the length ranges and constructing extractive summaries within different length ranges, which helps model to select different information from source document via desired lengths.
A method based on dynamic programming to satisfy the constraint of output lengths
A multi-objective optimization problem that includes generating complete summaries within desired lengths and selecting proper information to summarize based on desired lengths. Existing length-controllable summarization based on encoder-decoder models can be divided into two categories
The degree to which a summary is concise and does not only copy long passages from the source document
Relating to the vocabulary or words used in a language.
A measure of the importance of a passage in a text where the sentences of the document are connected by the similarity of their vocabularies.
The process of selecting the specific words to use in the summary
Simple features used to pre-compute review relevance in training, as opposed to deep encoder representations, allowing for selection of reviews from large collections without a significant computational burden.
The frequency of occurrence of words in a text. Concept
A commonly used evaluation metric for summarization that measures the n-gram overlap between system and reference summaries
The use of the same words or phrases in different texts
The use of different expressions to communicate the same or similar meanings
The analysis of both the lexical and semantic aspects of a text
A rating scale used to measure attitudes or opinions
A framework for analyzing factuality in summarization systems based on frame semantics and linguistic discourse theory
Included in the study of coherent abstractive summarization
A modeling approach for abstractive summarization
Entities found in the original text that can be used to infer the summary topic
Information about the immediate surroundings of a token in a document that contributes to producing high-quality summaries with adequate salient information
A method of exploiting standard transformer models by constraining the attention mechanism to be local, allowing longer input spans during training.
A part of the information selection layer that selects important sentences while generating each summary sentence sequentially
A loss function introduced to encourage the model to put most of the attention on just a few parts of input states at each decoding step.
A method proposed by ETC and Longformer to reduce computational overhead of fully-connected attention in Transformers by limiting each token to attend to a subset of other tokens.
Biases in summarization models that result from the tendency of key sentences to be located at the beginning of the text.
The ability to logically follow and entailed by the input document
Summarization of documents with input sequences that are longer than the limits of transformer models.
A complex summarization scenario that poses challenges to Seq2Seq models due to the even distribution of numerous details and salient content.
Tasks that involve processing long documents, such as scientific paper summarization and long-text reading comprehension, are challenging in natural language processing due to the varying details and subjects covered in such documents.
Document types that are significantly longer and structured, such as scientific papers
modelling concepts that span more than a few sentences, which is still a challenging task
Techniques that rely on long-range dependencies within a large local context for sentence ranking tasks.
Techniques that estimate word probabilities by considering long-range dependencies within a large local context.
Dependencies among distant tokens in a document that contribute to producing high-quality summaries with adequate salient information
Summaries with more sentences
A technique for modifying the loss computation during training to alter the learned behavior of summarization models
A key challenge in summarization is to optimally compress the original document in a lossy manner such that the key concepts in the original document are preserved, whereas in machine translation, the translation is expected to be loss-less.
The process of summarization, which involves compressing a text while retaining its meaning.
Another common drawback of neural abstractive summarization models where the generated content is inconsistent with the source document, also known as hallucination.
A common drawback of neural abstractive summarization models where the generated summaries fail to capture critical facts in the source document.
Languages that have limited resources and tools available for natural language processing tasks
Units that are less informative than units connected with verbs
An approach that approximates the co-occurrence matrix to tackle the challenge of lexical variety in summarizing student feedback
A setup where there are limited training examples or no training examples available for the downstream task.
A method of selecting the most frequent output from multiple models in post-processing.
A high time and cost demanding task that is often replaced with multiple choice or short answer questions in modern English exams.
An earlier approach for extractive summarization that involves implementing graphs and integer linear programming.
Parameters that are adjusted by hand to optimize the performance of a model
The paper addresses the lack of relevant datasets and techniques for effective answer summarization by developing a human-annotated dataset for multiperspective abstractive answer summarization.
Randomly masks some sentences and predicts the missing sentence from a candidate pool.
A STEP objective that learns to recover a masked document to its original form.
A novel method proposed in the paper for generating negative summaries.
A technique used to constrain copying words to the selected parts of the text, which produces grammatical outputs.
An unsupervised approach for extractive text summarization where the extract that maximally covers the information contained in the source text is selected.
A training method used to optimize the parameters of a model to maximize the likelihood of generating the correct output.
A method used in the mixed objective to optimize the n-gram overlap with the ground-truth summary in a summarization model
A loss function used in prior work to train the model
A representative example of early work in extractive summarization for selecting sentences based on both their relevance (to the central theme of the document) and the diversity of the selected sentences.
An approach for extracting relevant sentences into a summary that represents documents as a sequential transactional dataset and then compresses it by replacing frequent sequences of words by codes.
A form of content representation for factuality evaluation.
A metric used to indicate the popularity of a post in social media, such as the number of votes, shares, or bookmarks.
A system that takes a meeting recording and its transcript as input and produces a concise text summary as output, which preserves the most important content of the meeting discussion.
Difficulty in scaling complex summarization models to long-form documents due to memory concerns.
A technique used to preserve salient information learned from previous windows and enrich local texts
We propose the use of memory networks and convolutional bidirectional long short term memory networks for capturing better document representation.
The process of evaluating the quality of evaluation metrics.
How similarly a metric replicates human judgments of systems
The consistency and accuracy of automatic evaluation metrics in measuring the quality of summaries.
An approach to training Seq2Seq models that directly optimizes the model with the corresponding evaluation metrics.
Two kinds of methods proposed to adapt DPPs to large scale computing
A game where the generator G and the discriminator D are optimized, with D trying to distinguish the ground truth summaries from the generated summaries by G, and G trying to maximize the probability of D making a mistake
A principle defining the best summary as the one that leads to the best compression of the text by providing its shortest and most concise description.
An alternative to RL-based training, used in language generation tasks, but the accuracy of the estimated loss is restricted by the number of sampled outputs.
Used to optimize a model globally for an arbitrary evaluation metric
Language models that generate summaries that contain errors or inaccuracies.
The discrepancy between the objective of fine-tuning language models (maximizing the likelihood of human-written text) and the objective of generating high-quality outputs as determined by humans.
Conclusions that are based on incomplete or biased information.
a discrepancy or lack of alignment between the current research directions in automatic summarization and the needs of users, specifically university students
Relevant concepts that are not displayed in the search result summary
Important details such as medication dosage and route that are left out in a summary, undermining patients’ medical self-management.
A method of training an abstractive summarizer to generate summaries with varying amounts of copied content by gradually transitioning from copying only to both copying and generating new words not present in the source text.
Statistical models that account for both fixed and random effects
An objective function used to capture importance, non-redundancy, and coherence in automatic summarization.
An objective that jointly optimizes the n-gram overlap with the ground-truth summary while encouraging abstraction in a summarization model
A multi-task learning framework that optimizes jointly over several measures.
Maximum likelihood estimation, a standard training approach for seq2seq learning
Maximum Marginal Relevance, a strategy to control redundancy
The speed and ease with which a model can be created.
A method of combining multiple models to improve performance, often used in text-generation tasks.
Human-generated summaries used as a benchmark to evaluate the quality of system-generated summaries.
An evaluation approach that does not rely on human-generated model summaries.
A proposed architecture that can dynamically select salient input sentences to constrain the encoder-decoder attention without having to compute complete attention at inference time.
Variants of the transformer architecture that reduce the quadratic complexity of the self-attention mechanism.
An objective that leverages information about error spans in gold summaries, derived from factuality models, to train the summarizer.
Summarization of text in a single language.
A method of statistical analysis that uses random sampling to simulate possible outcomes
Leads to better performance when the decoder interacts with multiple agents
An automatic evaluation metric that performs well on the TAC dataset but is significantly worse than ROUGE-2 on the CNNDM dataset.
A policy used to select sentences during careful reading, with an adapted termination mechanism to select various but proper numbers of sentences.
A network designed to explicitly integrate summary-related features, like sentence semantics, importance, and position.
Articles from a variety of genres, including newswire, academic papers, movie scripts, and product or restaurant reviews.
The problem of correcting multiple erroneous facts in generated summaries.
A model that exploits the recent success of the encoder-decoder framework to generate aspect/sentiment-aware review summaries. It uses a mutual attention mechanism to capture the correlation of context words, sentiment words, and aspect words, and explores three kinds of attentions (i.e., semantic attention, sentiment attention, and aspect attention) to selectively attend to the context information when decoding summaries.
A feature in Transformers where multiple attention heads are used at all layers to highlight salient content.
Considering the semantic relevance to the question as well as the information consistency among different sentences to enable human-like multi-hop reasoning in question-driven summarization.
A task where data points are assigned multiple labels
A dataset where each data point can have multiple labels
A selector architecture that can be used for content selection in summarization.
Novel multi-task learning architectures
A novel memory network model for abstractive summarization that stores information from different levels of abstraction and can build representations of multiple ranges.
An approach to topic-focused summarization that uses multi-modality manifold ranking.
State-of-the-art solutions in summarization utilize multi-objective training strategies, including reinforcement learning techniques.
Conversations involving more than two participants.
The paper emphasizes the importance of multi-perspective summaries that cover the varying perspectives found in the answers to a question.
Datasets with multiple summaries, which are unrealistic to expect at scale for neural network training
Optimizes multiple rewards simultaneously in alternate mini-batches.
Including paragraph-level and document-level ones, induced to capture local and global semantic and syntactic information of a document.
Summarization tasks where the goal is to generate a summary consisting of multiple sentences
Proposed methods for generating meeting summaries
Utilized in current approaches to text summarization
Using multiple tasks to fine-tune a model for a downstream task.
Proposed approach to solve the problem of extractive summarization of the MD&A section of 10-K reports
An approach that involves training a model to perform multiple tasks simultaneously
A technique proposed in this work to obtain different representation of the texts for summarization and sentiment classification.
Addressing the repetition issue along with the multiview pointer network and generating informative answers.
A framework that can effectively incorporate multiple guided signals for the scientific document summarization task.
A learning paradigm in semi-supervised learning that encourages models to learn from multiple views of the same data.
A type of machine learning where the input data consists of bags of instances, rather than individual instances
The problem that arises when reducing the number of concepts in the model, allowing different sentences to have the same score, and ultimately leading to multiple optimal summaries.
A type of graph that involves multiple types of edges to capture different types of relationships among sentences and words.
A method that ranks sentences through an extractive labeling-based module and an attention-based module.
A machine learning approach where a model is trained to perform multiple tasks simultaneously
Terms that describe molecules and techniques in chemistry that can be accurately substituted for chemical formulae or accurate hyponyms like ketone without ontological knowledge.
Surveys conducted by a municipality to gather feedback from citizens
Metrics such as ROUGE that poorly reflect human preference for summarization
The amount of copying performed by different models when generating summaries.
A model used to convert traditional CNN architecture into an unsupervised learning regime
A commonly used automatic evaluation metric based on the overlap of n-gram
Queries for health-related content in the form of natural language
Outputs that are close to optimal in terms of the objective function being optimized.
Models that look up similar transcripts or recaps
A novel baseline introduced in this work that uses a sentence containing information that the summarization model should not focus on.
A technique used in machine learning to train models to distinguish between positive and negative examples.
Summaries that are factually inconsistent with the source text.
A model that estimates the coherence degree between two sentences by their distributed representation in an end-to-end fashion.
recent advances that led to a number of single-document summarisation systems that exhibit some level of abstraction in their outputs
A type of neural network used for natural language processing tasks
A framework used to generate summaries tuned to the target topic of interest.
Produces a short summary for a document by selecting a set of representative sentences.
Over-emphasize sentence importance and pay little attention to reducing redundancy in the selection phase.
State-of-the-art systems that use machine learning to automatically generate summaries
A novel approach that exploits pre-trained LMs for sentence classification in abstractive summarization
A novel neural framework for the identification and extraction of salient customer opinions that combines aspect and sentiment information and does not require unrealistic amounts of supervision.
The process of automatically generating a headline based on the text of the document using artificial neural networks.
Models that can provide an approximation of the global semantics captured from document contents, i.e., latent topics, as well as their posterior topic representations.
Models that use a seq2seq framework to generate a summary after encoding a full document.
Summarization models that use encoder-decoder structures with either recurrent neural networks or Transformer.
Five defined elements in a news story, including two ledes (Standard Lede and Image Lede) and three other categories (Synopsis, Narration, and Body Section), each with a specific function in building a news story and featuring a writing style (narrative or expository).
The process of creating a brief and concise summary of news articles.
A STEP objective that generates the next segment of the original document given the first segment of a document.
The process of selecting nodes (i.e., sentences) that are semantically similar to other nodes to be included in the final summary.
Techniques introduced to make both the teacher and student robust to noise
a probabilistic approach for sentence-level and document-level compression
Extracting sentences independently without using previous predictions
A method that utilizes non-autoregressive decoders to generate all output tokens in parallel
A decoder that can extract a non-fixed number of summary sentences simultaneously and individually, which is formulated instead of extracting sentences one by one to form a top-k summary.
A target distribution for abstractive models in which candidate summaries are also assigned probability mass according to their quality, as opposed to the one-point deterministic distribution assumed by MLE training.
Questions that require detailed analysis to explain or justify the final answers, such as questions in community QA or explainable QA.
Part of the current research setup for text summarization
The supplemental information that explains the introductory information in more detail
A type of QA that deals with questions that do not have a straightforward factual answer, such as opinion-based or explanatory questions.
Approaches that use frequency and information theoretic measures as proxies for content salience in summarization
ROUGE scores that are adjusted to account for differences in summary length.
A technique that combines deep learning models of encoder-decoder architecture and semantic-based data transformations to generate summaries in a generalized form.
Aggregations that are not found in the source text and must be generated by the summarization system
Words that are not present in the source document but are generated by the model in the summary.
a concept in summarization that refers to the degree to which a sentence contributes new information to the summary
The final summary is formed by maximizing both novelty and informativeness of the sentences in the summary.
The degree to which a model generates novel output
The process of aggregating customer satisfaction scores across different aspects of an entity
Information gathered by the care team that is based on observable and measurable data.
A technique used to understand the contribution of input features by masking them and observing the change in the model's output.
A reinforcement learning algorithm that uses a pre-collected dataset to train a model.
Existing components for tasks other than query-based summarization can be competitive with state-of-the-art methods in the field.
Frameworks such as entailment models or QA systems that have been explored in past work to detect and correct errors in generated summaries.
A summarizer that mimics how a human might approach a lengthy transcript, identifying important and relevant portions to produce a new summary piece.
The assumption that model output is evaluated based on similarity to general-purpose reference summaries reflecting the full content of the original document.
A set of concepts and categories in a subject area or domain that shows their properties and the relations between them.
Information about the relationships between medical concepts.
Out-of-vocabulary words or words of limited occurrences that the proposed framework is capable of coping with, achieving semantic content generalization.
A pre-trained component of the proposed pipeline framework that identifies opinion phrases in reviews.
A subsequence of tokens within a review that expresses the attitude of the reviewer towards a specific aspect of the entity. It may not be contiguous in the review, and a word can be part of multiple opinions.
A simple and controllable component of the proposed pipeline framework that merges, ranks, and optionally filters the extracted opinions.
The set of opinion phrases within a review, defined as Or = {(oi, poli, ai)}|Or|i=1, where poli is the sentiment polarity of the i-th phrase (positive, neutral, or negative) and ai is the aspect category it discusses.
A dataset consisting of Amazon reviews from six product domains, and includes development and test sets with gold standard aspect annotations, salience labels, and multi-document extractive summaries.
A summary that is considered to be the best possible version of the original text
A novel non-learning based extractive summarisation method that formulates extractive summarisation based on the optimal transport theory.
A method for measuring the distance between probability distributions
A problem of finding the best solution among all possible solutions.
A method used at training time to select informative guidance signals to encourage the model to pay close attention to the guidance.
Approaches that extract the most important sentences from the input document to generate a summary
One-to-one matching of each fact in the gold summary to one fact in the source text to obtain the oracle label
Extractive labels created using a greedy oracle labeling algorithm for the Pubmed and arXiv datasets
Sentences that tend to be crowded at the beginning of news articles while distributed more evenly in scientific papers.
A solution proposed to prevent the repeated phrases problem by ensuring that successive context vectors are orthogonal to each other.
Experiments conducted to examine the cross-domain generalizability of the compressive system.
A problem in natural language processing where a word is not present in the vocabulary of a model, making it difficult to generate accurate summaries.
Words that are not present in the vocabulary of the model
words that are not present in the vocabulary of a language model, making it difficult for the model to handle them.
A measure of centrality that assumes a word sending out more relevance score to others is more critical.
The length of the summary produced by a summarization system.
A category of ensemble algorithms that corresponds to the majority vote method in classification tasks.
The length of the summary produced by the system
When a model performs well on the training data but poorly on new, unseen data
when a model is too complex and fits the training data too closely, resulting in poor performance on new, unseen data
Words that appear in both the query and the source document
Summaries that exceed the desired length constraint
A method used to derive sentence prestige by building a connectivity graph
A manual evaluation approach that compares two summaries
An approach to summarization evaluation that collects preference labels over sentences in documents or over summaries from a human assessor, requiring less cognitive effort than writing a reference summary or manually scoring a machine-generated summary.
Labels provided by the user that compare two candidates and label the best one.
Simple and inexpensive annotations used to train the evaluation model
A set of texts in two or more languages that are aligned at the sentence or phrase level
A method of decoding the summary in parallel, rather than sequentially.
A technique used in global encoding to refine the representations at each time step with consideration of the global context
Methods used to induce a latent variable model for unsupervised sentence summarization tasks
A task where the model determines if two sentences have the same meaning
The ability of models to generate novel text by rephrasing the source document.
A technique used in this work where only some attention heads are masked to guide the summarization process.
Generating summaries that conform to a specific pattern, such as court judgments, diagnosis certificates, abstracts in academic papers, etc.
One of the link analysis algorithms applied to graph attention network (GAT) to obtain node representations that better reflect the query and document relationships.
An algorithm used to leverage repetitive random walks on a semantic network to identify the relevance of different senses of a word
A technique based on person names to reduce extrinsic hallucinations involving named entities.
A method that creates summaries from extracted phrases rather than from sentences
Models that condition the generation process on themes of interest and text style transfer controls selected attributes, such as politeness, emotions, or humor of the generated text.
The ability of the proposed algorithm with global awareness to enhance beam search for neural abstractive summarization without requiring any model or parameter modification.
A condensed version of the podcast episode that can serve as a basis for decision-making or as a synopsis
A mode in the pointer generator architecture where words are directly copied from an aligned source context.
Statements made by participants in online discussions, either agreeing or disagreeing with others.
The establishment of correspondence between sentences, which is crucial for fusing sentences. It includes entity and event coreference, shared words/concepts between sentences, and more.
Text chunks that convey the same or similar meanings and tie two sentences together into a coherent text.
A metric used to measure the relevance and redundancy of a summary. It is based on the probability of a sentence or summary occurring in a document.
A model trained via reinforcement learning to generate summaries that maximize the score given by the reward model.
A method used in the mixed objective to encourage abstraction in a summarization model
A set of 8 metrics on the Accuracy and Fluency aspects designed to quantify the primary sources of errors over representative models
A problem in extractive summarization where sentences appearing earlier in a document tend to be selected as the most important, resulting in sub-optimal models.
Indicators of important content in news articles based on the sentence's position in the document
A channel used to learn sentence position features.
The idea that important sentences appear in preferred positions in a document.
Information about the position of words in the input sentence
An adversary introduced to optimize the neural extractive summarizer in an alternating manner.
A method that significantly outperforms strong baselines in single-document summarization by using position information to transform undirected edges into directed ones
Editing generated summaries to improve their quality
The process of correcting errors in a generated summary after it has been generated
The probability of an entity being in a summary given the source document.
Language models that are trained on large amounts of data and can be used to improve the decoder's ability to learn summary representations, context interactions, and language modeling together
A pre-trained encoder is an essential component for sequence generation tasks and often these tasks benefit from sharing the weights between the encoder and the decoder.
A common approach in natural language processing where a model is first pre-trained on a large corpus of text and then fine-tuned on a specific task.
The objective used during pre-training to improve the quality of the generated output in downstream tasks.
The degree to which repeated measurements under unchanged conditions show the same results
A mechanism that predicts the extent of key information covered in the final summary to further guide the summary generation
A problem formulation for reward learning that involves predicting the relative preference between two summaries
A form of feedback in which a user provides a preference over a pair of predictions, which is less cognitively burdensome than providing ratings or categorical labels.
Teaching the model how to rewrite a summary which is a directed-logical subset of the input document
Embeddings that are trained on large amounts of data and used in the summarization task
A task designed to summarize a patient’s problems and generate relevant diagnoses to assist healthcare providers and overcome the cognitive burden and information overload.
A combination of the pretrained language model and a smoothed problem specific target language model to guide the fluency of the generation process
Techniques used to guide the generation process of text generation models.
Features of speech such as intonation, stress, and rhythm that convey meaning and emotion
Models that use prototype document-summary pairs to improve summarization performance.
An effective distillation method for Seq2Seq models, where the teacher model generates pseudo summaries for all documents in the training set and the resulting document–pseudo-summary pairs are used to train the student model.
The reason why a summary is being created
An analysis that brings insights as to what current unsupervised models are missing in automatic text summarization
An approach to topic-focused summarization that uses query attention.
A type of summarization that highlights those points that are relevant in the context of the query.
Interactive and non-interactive techniques proposed to translate input questions into structured queries covering specific elements of the questions.
Summaries that consist of a few sentences around query terms in the results
Features like query word overlap that are designed to learn the relevance ranking.
A technique that creates a brief, well-organized and fluent summary that answers the need of the query. It is useful in many scenarios like news services and search engines, etc.
The process of ordering a set of systems based on their performance.
A function that maps text documents to scores and is used to rank candidates.
Metrics used to evaluate the informativeness of posts based on their ranking in a summary.
Models used to rank sentences in order of importance for summarization.
An additional module in a summarization system that re-scores candidate summaries generated by the main summarizer.
A type of question answering system that focuses on answer span extraction in long documents.
The proportion of true positives that are correctly identified by a model
A system that provides recommendations to users based on their preferences and behavior.
A type of neural network that can process sequences of inputs
Successful in a variety of NLP tasks where an encoder obtains representations of input sequences and a decoder generates target sequences
Information that is repetitive and should not be included in summaries
The issue of redundant phrases between selected sentences in the naive approach for the first prediction step of the decoder, which makes independent binary decisions for each sentence, leading to the absence of overlap or redundancy modeling between the selected target sentences.
A general framework that serves as a base system to construct a summary and as a meta system to select the best system output from multiple candidates, allowing the base and meta learners to share a set of parameters.
The bias that occurs when evaluating against a single reference summary
Texts written by humans used as a basis for comparison with generated summaries.
Reference-free evaluation metrics focus on evaluating summaries without the need for human-annotated summaries as reference. A high-quality summary should be concise and contain the most important information of its document. Some reference-free evaluation metrics unsupervisedly construct a pseudo-reference summary by selecting salient sentences from the source document, while others evaluate the summary quality by measuring how much information from the document is represented in the summary. QA-based evaluation metrics achieve this possibility by first asking the same questions to document and summary and then comparing their answers.
Evaluating single document summaries without reference summaries, using embedding similarity between the full document and system summaries.
A statistical method used to analyze the relationship between variables
A problem formulation for reward learning that involves predicting a continuous value
A technique used by the summarizer to visit portions of the transcript in chronological order, while allowing zigzags to produce a coherent summary.
A term added to the loss function during training to prevent overfitting.
A problem in text summarization where the entities in the summary exist in the source document, but the relations between them are not accurately reflected in the summary.
An easy-to-compute model-free metric that evaluates factual consistency given a summary and the article, employing the extracted relations and not requiring human-labelled summaries.
A technique for maintaining reasonable performance even in the case of a sub-sequence with errors, which involves accurately estimating the relative quality of different generated outputs, since effective inference requires comparison among these candidates.
The process of assessing the relevance of documents for a given task or search topic
The degree to which a metric can be trusted to produce consistent and accurate results.
A metric that can automatically evaluate the content of a summary
A recurring problem in models based on the encode-attend-decode paradigm where the summaries produced by such models contain repeated phrases.
The quality of the summary in terms of how much it repeats information from the original text
Different ways of representing text, including sentence embedding, un-contextualized word embedding, and contextualized word embedding.
A process of learning unbiased sentence representations using deep neural networks
A technique used in neural networks to help prevent the degradation of performance that can occur when training very deep networks.
A simpler manual evaluation approach that does not rely on reference summaries and can be attained via crowdsourcing
A manual evaluation method where human annotators score summaries on a LIKERT scale ranging from 1 to 5.
The process of adjusting and updating a machine learning model to improve its performance on a specific task or in a specific domain
A model that utilizes word embeddings and domain-specific knowledge for finding the appropriate context of citations, aimed at capturing terminology variations and paraphrasing between the citation text and its relevant reference context.
A function that can guide RL-based summarisers to generate more human-appealing summaries
A model trained via supervised learning to predict the human-preferred summary.
A method used to alleviate the sparsity of training signals in abstractive summarization models
A theory that inspires the formulation of sub-sentence highlights, where subsentence highlights resemble the nuclei which are text spans essential to express the writer’s purpose.
A theory that provides a coherent and well-organized representation of documents and suggests discourse-level segmentation can help model semantic information with more refined granularity.
The agents that are misled by the poor performance of ROUGE at summary level, as existing RL-based summarisation systems rely on summary-level ROUGE scores to guide the optimisation direction
The ability of a model to perform well on different document genres and lengths.
Refers to tasks where mistakes made by AI systems can have serious consequences.
Two aspects rated by humans to evaluate the model-generated summaries
The balance between selecting sentences with high semantic similarity to the gold summary and resolving redundancy between selected sentences.
the process of estimating the importance of each sentence in a document
Labels provided by externally trained content selectors that indicate the importance of different parts of the source text.
the process of extracting the most salient sentences to form a summary
Metrics used to evaluate the importance of the information in the generated summary.
A network that manages the information flow from encoder to decoder explicitly and assigns a salient score for each token in source documents according to their encoded representations
The most significant clinical terms occurring in the findings, which can be used to improve the final impression generation.
The procedure of information representation and discrimination to ensure generated summaries contain adequate salient information of the original documents
Opinions that are important or relevant to the product or service being reviewed.
Important phrases that are not common across different domains
A third type of guidance signal investigated in the paper, which involves providing the model with salient relational triples in the form of (subject, relation, object).
The inability of a machine learning model to generalize well to new data due to insufficient training examples
a potential problem when recruiting participants to evaluate summarization systems, as different demographics may exhibit different preferences in rater studies
Mechanisms that selectively sample offline data in favor of human feedback learning. The sampling strategies focus on low-rewarded samples or documents that are similar to fine-tuning data.
The process of summarizing research papers, usually generating paper abstracts.
The process of generating summaries from professional texts like COVID-19-related papers, which are difficult due to their long texts with complicated structures.
Summary content unit, which is a unit of information in a summary
a strategy for summarization where an extractor selects salient sentences, then an abstractor generates a summary
A technique used to solve the issue of extracted linked entities being too ambiguous and coarse to be considered relevant to the summary
A modified version of a Gated Recurrent Unit (GRU) network, which can decide how much of the hidden state of each sentence should be retained or updated based on its relationship with the document.
The average cosine similarity between sentences in a paper's citation summary, which is consistently higher than in its abstract and measures the focus in describing the paper's main contributions.
The characteristic of a summary that conveys the majority opinion of the reviews and does so in a self-consistent manner.
Highlighted text that is understandable on its own, without the need for specific information from surrounding context
A baseline used in reinforced abstractive summarization methods that is obtained by greedily searching for a sequence that maximizes the likelihood probability of the current model.
A method used to train the network end-to-end
A technique employed in this paper to optimize the ConvS2S architecture enhanced by topic embedding and SCST, which yields high accuracy for abstractive summarization, advancing the state-of-the-art methods.
Clusters with titles that accurately reflect their contents
A score that measures the degree of semantic similarity between different parts of a summary.
Repetition of n-grams within the output of a model
A method of training a summarization model without the need for human-generated reference summaries.
Methods that do not require labeled data for training
A method to boost the zero-shot capability of the model.
Pretraining with prohibitively-large datasets to facilitate adaptation to new tasks with less abundant data.
A training method that uses unlabeled data to train error detectors.
The process of combining multiple entities into a more general expression in order to change the level of detail in a summary
An unsupervised extractive model that learns a representation of text over latent semantic units using dictionary learning.
Continuing sentences that describe the same facet
A desirable summary aspect that is encouraged in reinforcement learning approaches.
The meaning and context of the text beyond its literal words
the aim of the new evaluation metric to better capture this aspect of a summary, i.e. be more sensitive to hallucinations and omissions
A representation of predicate-argument relations between content words in a sentence that can guide summary generation.
A model that leverages the input sentence and semantic dependency graph to generate a summary in a complementary way.
Output summary by existing graph-based methods tends to deviate from input text due to statistical level graphs
A metric clustering paradigm used to estimate the student coverage of each phrase in the summary
The semantics of the document may drift from section to section.
A channel used to learn sentence linguistic features.
The degree to which a summary conveys the same meaning as the input text
A type of multiplex graph that connects two sentences sharing similar meanings.
The process of using context and background knowledge to understand the meaning of a text
The exploration of more compelling ways to evaluate summarization, translation, and dialog beyond token overlap, including using word embeddings and universal sentence representation.
The meaning of words and phrases in context
Evaluation of abstractive summarization needs a semantic overlap based method.
A new NLP task for summarizing multiple alternative narratives with different perspectives by cross-verifying their information contents against each other.
The relationships between sentences that are ignored when existing language models model sentences word-by-word.
A type of relational information among words that has been proven to be useful for downstream tasks.
a large-scale scientific papers summarization dataset with citation graph
Another task that can be aided by the transfer of the latent semantic representation into useful editing tasks.
The similarities in meaning between different words or phrases
Scores that measure the degree of similarity in meaning between different words or phrases
The ability of models to understand the meaning of the source document and generate meaningful paraphrases.
The degree of overlap between the meaning of the reference and model summaries.
A set of diagnostics proposed in this work for measuring the sensitivity of factuality metrics to factual inconsistency.
The method of shortening text by removing words or rephrasing parts of a sentence
State-of-the-art techniques used to prepare groundtruth rankings of sentences from the original document by computing the semantic similarity between each individual sentence of the original document and the entire human-written summary.
Design choices for encoding sentences in neural network architectures for summarization
A crucial step in extractive summarization where a representative subset of sentences is selected, which contains the information of the entire set.
Methods that select sentences in a document to create its summary, with advantages of truthfulness compared with abstractive methods and of fluency compared with word extraction methods
Built with recurrent neural networks that remember the partial output summary and provide a sentence extraction state to score sentences
Design choices for extracting sentences in neural network architectures for summarization
The degree to which the generated text is grammatically correct and coherent.
Extracting full sentences as the extraction unit
A model that includes a planning step at the sentence level before generating the summary word by word, in order to generate more abstractive summaries without sacrificing ROUGE and coherence.
The bias in news summarization where sentence position dominates the learning signal
A process in sentence regression that evaluates the importance of each sentence with a ranking model
Extractive summarization is modeled as a sentence ranking problem with length constraints
A branch of extractive summarization methods that models the relative importance of a sentence given a set of sentences
The first stage of the fact consistency assessment framework where top-K pieces of evidence are selected from the original document
Individual sentences selected for summarization.
A type of extractive summarization that involves selecting important sentences from the original text to create a summary. Concept
a component of ESCA that ensures the abstractive summary focuses on both correct and desired concepts
A measure of the relevance between summary sentences and the original document.
A method for extracting the final summary from candidate sentences
The proper way to evaluate the SOS task, which improves the inter-rater agreement compared to the traditional ROUGE metric and shows a higher correlation with human judgments.
Focusing on summarizing individual sentences
A sequence-to-sequence model that uses attention mechanism.
A neural sequence-to-sequence architecture used for text generation tasks like machine translation and image captioning.
A framework that has achieved state-of-the-art performance on abstractive sentence summarization task.
A type of machine learning problem where the input and output are both sequences, and the goal is to learn a mapping between them.
A statistical machine translation model that uses an encoder to convert the input text as a vector representation, and then feeds this representation into a decoder to generate summary.
A type of neural network architecture that has been demonstrated to be the state-of-the-art for SEQ2SEQ modeling in natural language generation tasks, such as abstractive summarization.
A model that uses a sequence-to-sequence architecture to incorporate the salient clinical terms into the summarizer.
A method of extractive summarization that formulates it as a problem of labeling each sentence in the original text as either included or excluded in the summary.
A deep learning-based approach that encodes input documents as vector representations with a long short-term memory (LSTM) and uses another LSTM as the decoder to generate corresponding summaries. It captures the semantic and syntactic relations between raw documents and their summaries in a scalable and end-to-end way.
A neural network architecture that learns to compare two inputs and output a similarity score.
A method of computing the similarity between the source document and the candidate summary in extractive summarization, using a pre-trained BERT model in a Siamese network structure to derive semantically meaningful text embeddings that can be compared using cosine-similarity.
A statistical test used to determine if the differences between two sets of data are significant
A method of aggregating latent vectors by taking their average.
A phenomenon where different conclusions are drawn depending on which subset of a population is considered.
Datasets with only one summary, which might not be optimal for summarization due to human variation
Story highlights of an article provided by news websites consisting of three or four succinct itemized texts for readers to quickly capture the gist of the document.
A technique that generates a general summary of popular opinions for a single entity.
Summaries created by humans that serve as the sole reference for evaluating the quality of machine-generated summaries.
A practical solution for summarizing long-form documents by processing them separately in multiple windows
A task that extracts important sentences and representative comments as the summarization, utilizing the social information of a web document to support sentences for generating a high-quality summarization.
Informal language and massive noise within social media content make training deep neural networks on such datasets challenging.
A way of aligning the input sentence with the summary that allows the decoder to focus on the most important parts of the input sentence
The de facto standard attention mechanism that assigns attention weights to all input encoder states.
A mechanism that calculates aspect/sentiment-aware review representations.
attention methods that first locate salient text regions within the input text and then bias the abstraction process to prefer such regions during decoding
Labels that are not binary or definitive, but rather represent the degree of relevance or importance
A method of optimizing shared parameters that achieves higher performance than hard sharing
Class probabilities produced by the teacher model in knowledge distillation.
The original document(s) from which the summary is generated.
The domain used for training the neural summarization system.
a knowledge structure that contains few connections between concepts or ideas.
A transformer model with the sparse attention mechanism for abstractive summarization, which supports the encoder to model longer input sequences with limited GPU memory.
The degree to which the attention weights are concentrated on a small subset of the input sentences, which can be exploited to reduce computation cost.
Achieving faster processing time compared to existing methods
Summaries that do not adapt to the search query
Models for abstractive sentence summarization trained on relatively small scale training corpora
A method of selecting a sample of data from a larger population for analysis
Improvements that are unlikely to have occurred by chance
Three sequence-to-sequence pre-training objectives, namely Sentence Reordering (SR), Next Sentence Generation (NSG), and Masked Document Generation (MDG), which can be used to pretrain a SEQ2SEQ model on unlabeled text.
A method where a summary is constructed incrementally by choosing new content conditioned on previously planned content.
A measure of how much a summarization system's performance changes when evaluated on different datasets
Added to articles to provide a teaser summary of the most important points of the article, but most straplines in the Newsroom corpus are not summaries of their associated articles. Distinguishing straplines aimed at piquing a reader’s interest from abstractive summaries is necessary to obtain high quality data.
Patterns in the organization of text that can be exploited by extractive summarization methods.
Each summary sentence is generated by compressing several specific source sentences.
Different summary sentences usually focus on different sets of source sentences to cover more salient information of the original document.
Mechanisms that facilitate copying source words and relations to the summary based on their semantic and structural importance in the source sentences.
Incorporating structured action representations to generate more faithful todos.
Used as both the objective and attention weights for extractive summarization.
Enable the model to better control both the content to be conveyed and the syntactic structure needed to express it, ultimately improving the factuality and grammaticality of the generated summaries.
A representation that facilitates the connection of relevant subjects and the preservation of global context.
Potentially beneficial for reducing data sparsity and localizing generation errors in abstractive scenarios.
Summaries that allow humans to browse specific aspects of interest more readily.
Transformer-based architectures that have the flexibility to model some form of structure of the input, e.g., hierarchical document structure.
The degree to which the generated text matches the style of the input or a specific genre.
The non-informational or non-factual aspect of text that drives the quality of response from its audience.
Summaries that are tailored towards specific stylistic preferences, such as formality.
Functions of summarization including position, importance, diversity, and information.
The idea that summarization is a combination of sub-aspect functions, such as information and layout.
Identifying a single most informative textual unit from each sentence to create highlights.
Identified from a document to strike a balance between the quality and amount of highlights
Extracting non-terminal nodes in a constituency parsing tree to separate important and unimportant contents
Position, importance, and diversity, which determine the output form of summarization.
Based on personal opinions, feelings, and attitudes
Information gathered by the care team that is based on the patient’s experiences and perceptions.
Intents that are difficult to express through queries and require more personalized examples
Metrics used to measure the quality of generated summaries are important for the development of summarization systems. Previous evaluation metrics require human-annotated summaries as reference and measure summary quality through the similarity between generated summaries and their reference summaries. However, such reference-based evaluation metrics cannot accurately evaluate the summary, because a document has many correct but different summaries. Thus, it is useful to develop reference-free evaluation metrics for this task.
Graphs built among sentences to capture inter-sentence relationships and rank them by estimating summary-worthy features of sentence importance.
A shortened version of a text document which maintains the most important ideas from the original article. Automatic text summarization is a process by which a machine gleans the most important concepts from an article, removing secondary or redundant concepts. Extractive summarization is a technique for generating summaries by directly choosing a subset of salient sentences from the original document to constitute the summary.
A new task for content selection in topic-focused summarization that involves producing the next sentence in the summary given a topic, a partial summary, and a reference document(s).
The degree to which the summary is logically connected and easy to understand
A layer in the model that decodes the selected information into a summary
Techniques used to assess the quality of system-generated summaries, including manual and automated pyramid and ROUGE scores.
The generated summaries must accord with the facts expressed in the source.
Overlaid sub-sentence segments on source documents to enable users to quickly navigate through content
The length of a summary, which can be controlled by the generator and predicted by the selectors.
Phrases or groups of words that appear in a summary
The fact that certain sentences are more appropriate for inclusion in a summary regardless of the specific document they appear in.
Different aspects of a summary that can be specified for evaluation, such as factual consistency, fluency, coherence, and informativeness.
A phenomenon where the decoder generates less informative and overly generic summaries due to simple average vector aggregation.
A complex task involving various linguistic operations that is useful for developing student linguistic proficiency including text comprehension and composition.
A scenario where the correlation between the candidate metric and human judgments is computed for each topic individually and then averaged over topics.
The process of evaluating the quality of a summary as a whole, rather than evaluating individual sentences or phrases.
A novel approach to extractive summarization that formulates it as a semantic text matching problem, where a good summary should be more semantically similar as a whole to the source document than the unqualified summaries.
A training strategy that optimizes the summary-level quality of a summary, rather than optimizing individual sentences or phrases.
Data that has been labeled or annotated by humans for machine learning algorithms to learn from
Training models with large training corpora comprising pairs of long texts and their summaries
The process of providing labeled data to train a machine learning model
Features like the TF-IDF cosine similarity between a sentence and the query that are inadequate to measure the query relevance.
A measure of how unexpected a word or phrase is in a given context
The components of a sentence that provide information about the verb, such as the subject and object.
A method of analyzing the grammatical structure of a sentence by breaking it down into its constituent parts.
Changes in the form of a word to indicate tense, number, or gender.
A type of relational information among words that has been proven to be useful for downstream tasks.
Beneficial for generating compressed yet informative summaries
A related task that exploits word-level syntax to generate high-quality summaries from the language modeling perspective, and thus alleviates the issues of incomplete sentences and duplicated words.
Artificial pairs of summaries and reviews created for training a summarization model.
Data generated specifically to train models on the factuality detection task.
Oracle methods, baselines, and state-of-the-art approaches are used to evaluate summarisation quality.
The bias towards certain sub-aspect functions in different summarization systems.
Bias learned on a news corpus that can be reduced by modulation with semantic sub-aspects.
Summaries that match the interests of the reader and are required in manifold settings, such as summarization of complex event streams with a focus on regions, entities or topics of interest for journalists or analysts, understanding reviews or opinions from different perspectives, the summarization of electronic health records with a focus on the medical sub-specialty of the physician reader, or any other form of personalized summarization targeting explicitly defined or implicitly mined preference parameters.
The domain for which the neural summarization system is being trained.
Strategies for combining multiple tasks during fine-tuning.
A pre-training method based on an unlabeled small-scale task-related corpus.
The model's task-agnostic approach allows it to implicitly learn and leverage content plans directly from the data.
These are pretraining methods that do not take into account the specific downstream task. Examples of such methods include corrupted span prediction (T5), masked language model (BERT), denoising objective (BART), and vanilla language model (GPT).
Modules added to the model to enable it to effectively share knowledge from multiple tasks.
A classification system used to categorize different types of interactions in AI-assisted text generation.
A traditional approach to abstractive summarization that uses manually defined rules to fill in incomplete sentences.
An approach to traditional abstractive summarization that involves manually creating hard templates by domain experts and populating key snippets to form the final summaries.
Explicitly-compositional vector embeddings of symbolic structures that encode a constituent in a symbolic structure as a composite of a role (encodes structural information) and a filler (encodes content).
The discourse and terminology variations between the citing and the referenced authors that traditional IR models relying on term matching for finding relevant information are ineffective.
Neural networks that can encode and decode text.
A related task that improves the quality of locating salient information of the text by learning a category-specific text encoder.
A theory that covers a broad range of points of correspondence, including entity and event coreference, shared words/concepts between sentences, and more.
A model that converts text into a numerical representation
Techniques used to convert text into numerical representations that can be processed by neural networks
Model that takes as input a pair of a document and a corresponding gold summary and perturbs the summary to render it factually inconsistent with the original document
Systems that aim to produce text that is fluent, coherent, relevant, and factually correct.
Tasks that involve generating new text based on existing text.
Automatic evaluation metrics that compare two summaries based on matching their tokens, either through some lexical or embedding-based similarity
Tasks that involve ranking text documents based on their relevance to a user's topic of interest.
The step where the proposed architecture conducts sentence search based on fluency to return the final extracted summaries.
a way to evaluate summarization systems by comparing machine-generated summaries to human summaries
The process of making a text easier to read and understand, often by using simpler vocabulary and sentence structures.
a self-supervised objective used in CoCoNet's pre-training, where each sequence in the corpus is divided into two spans with some overlapping words, and the first span is used to generate the second by copying.
Short overviews of long documents or document collections that allow readers to understand the content without the need to read full documents.
The coherence of a summary achieved by considering keywords in a sentence for coherence to other sentences and capturing interactions between sentences through discourse dependency trees
Natural language descriptions of the relationships between concepts shown on the edges
A desideratum that encourages summaries to cover diverse information in the input documents.
Methods for modeling importance that formalize the concept of importance and develop general-purpose systems by modeling the background knowledge of readers
The main argument or point of an editorial.
Document encoding, information selection, and summary decoding
The likelihood of a token appearing in the summary
A specific type of loss truncation that involves downweighting certain tokens during training
The optimization of the likelihood of individual tokens in a summary.
A strategy used by previous models to extract a constant number of sentences from different documents, which conflicts with the real world.
The ratio of sentences within a dataset that are relevant to the query
Produces document-level topic vector and merges it to traditional dense word embedding obtained by an extension of BERT.
important words or phrases that capture the main idea of a text.
An alternative approach to document clustering that discovers hidden thematic structure
A vector representing the topic of the summary to be generated
Introduce topic as guidance to help generate abundant topic-related words and maintain the original ideas of documents.
Summaries produced by extractive summarization using sentence-level features that have been leveraged for producing query-focused or topic-based summaries.
A novel approach used to artificially create a dataset containing articles with multiple topic-oriented summaries.
The task of generating a summary given a source text and a specific query or topic.
The measure of semantic relatedness between words and the topical coherence of a document
A graph used for summarization where one set of nodes corresponds to topics
A graph that consists of sentence and topic nodes and efficiently captures inter-sentence relationships for summarization.
A method of ordering nodes in a tree based on their dependencies.
The limitation of current models that requires a balance between generating abstract summaries and maintaining faithfulness to the source documents.
A curve indicating the trade-off between abstractiveness and faithfulness
A distribution gap between the training and test distributions in the meta-learning stage.
The paper explains that IQE uses pairs of a post and reply candidate to train the model. The model requires replies only during the training and not during the evaluation.
The data used to train and evaluate a model
Requires large amounts of data
Abundance of training data is necessary for the performance of neural models. Lack of sufficient training data worsens the model’s ability to generalize patterns in training data to unseen data.
Training data is generated from source documents by applying a series of rule-based transformations inspired by error-analysis of neural summarization model outputs.
The study of how models learn during the training process
The paper uses a Transformer conditional language model (CLM) that is trained with a ‘leave-one-out’ objective by attending to other reviews of the product.
An alternate method for summarization model generation that is superior in terms of performance and resource utilization.
A negative sampling strategy that considers not only negative examples directly sampled from the data, but also negative examples that are indirectly related to the positive examples.
The novel conceptualization of extractive summarization as a problem of inducing document-level dependency trees.
A decoder that predicts dependency arcs between words in the partial summary.
Guides neural summarization system to identify summary-worthy content and compose summaries that preserve vital meaning of source texts
a neural network architecture that integrates three separate encoders to consider the context of the original text, topic keywords, and knowledge structure simultaneously.
A method of removing duplication in extractive summarization by skipping sentences that have trigram overlapping with the previously selected sentences.
The process of cutting off parts of a document, which can lead to information loss
The level of confidence users have in the accuracy and reliability of the AI-assisted text generation system.
Concern about whether all facts of a generated summary are mentioned in the source text
The important feature of summarization for it to be widely accepted in real-world applications
A parameter that can be adjusted to change the output of the system
One for words and the other for sentences
A process that generates the summary using a left-context-only decoder in the first stage and predicts the refined word one-by-one using a refine decoder in the second stage
False positives in statistical analysis
A dataset created by permuting the order of sentences in training articles to reduce position bias
Part of the current research setup for text summarization
The summarization task is underconstrained in that the importance of a piece of information highly depends on the expectations and prior knowledge of a reader
Inaccurately reproducing factual details, inability to deal with out-of-vocabulary (OOV) words, and repeating themselves
Content that is not faithful to the original text.
The generation of summaries that contain information that is not factually consistent with the source documents.
Query-specific information utilized by recent unsupervised approaches for effective performance.
A method of summarizing text without paired summaries
One of the reasons why the model sometimes exhibits such an untruthful behavior lies in untruthful article-headline pairs, which are used for training the model
The highest possible performance that can be achieved in the task of generating summary-worthy aggregations
The maximum possible performance of a metric
The degree to which a summary is helpful or valuable
Reviews written by users on products sold online.
A user-based selective mechanism that considers different user preferences on review content when summarizing a review and applies user-specific vocabulary to consider user’s writing styles when generating a summary.
Content created by users, such as comments, that can be combined with a news article to provide a perspective viewpoint regarding an event.
The wants and needs of people who use automatic text summarization
the requirements and expectations of individuals who use automatically generated summaries
Fragments of speech transcripts that are not well-formed grammatical sentences
A function that represents the goodness of an action on a given state in RL
An RL technique that takes a value-based approach such as Q-learning or the combination of policy and value-based approaches such as Asynchronous Advantage Actor-Critic
Summaries that can have varying lengths as opposed to fixed lengths
Traditional models used to characterize sentence singletons and pairs.
Incorporating linguistic preferences of the readers into the summary
Supervision data that is not directly paired with documents but is used to train models
A pooling technique that represents the document by taking the sum of sentence embeddings weighted by the automatically learned query relevance of a sentence.
The accuracy of aligning target words with their corresponding source words.
Scores derived from matching words against hidden topics, used to determine the importance of a sentence in a text.
A constructed heterogeneous graph that connects each sentence to its contained words, allowing for different granularities of information to be fully used through multiple message passing processes.
A structure that takes into account the relationship between words and sentences in a document.
The process of directly applying a pre-trained model to a target task without fine-tuning
allows users to control important aspects of the generated summary at test time