Overview

This tool provides a unified way of exploring literature on neural single document English text summarization. It currently hosts 514 publications (spanning 2015-2022), from top conferences such as ACL, EMNLP, AAAI (and more).

Each paper has been carefully annotated for multiple attributes to provide a quick overview, following the annotation scheme described below. This allows researchers (especially those new to the field) to easily find relevant papers and to better organize their literature reviews.

visual search
Conceptual View of a Neural Text Summarization Pipeline. A paper may target one or more components.
Annotation Scheme

Each paper is annotated with the following attributes:

Domain
The domain(s) of the source documents that are being summarized.
Dataset
The dataset(s) used for training and evaluation. For example CNN/DailyMail, Gigaword, etc.
Learning Paradigm
If the model was trained via supervised, unsupervised, or reinforcement learning.
Paper Type
If the paper proposes a new method, analysis (evaluation), metric, dataset, or combinations of these.
Metric
The metrics used for automatic evaluation. For example, ROUGE, BLEU, etc.
Human Evaluation
The summary quality criteria that were evaluated manually. For example, Informativeness, Fluency, etc.
Pipeline Component
The component of the conceptual pipeline shown above that the paper focuses on. For example, a new objective function, efficient encoding of the input document, post-processing of the generated summaries, etc.
Challenge Type
The type of challenge outlined in Challenges that the paper addresses. For example, controlled (or tailored) summarization, hallucinations in the generated summaries, etc.
Purpose of the Summaries*
This answers three questions:
  1. Who is the target audience for the generated summaries?
  2. Why are the summaries being generated?
  3. How will the generated summaries be used?
This allows for a better alignment between the goals and the results of the paper.
Claims & Contributions*
The claims (or research questions) outlined and the corresponding contributions (solutions) presented by the paper.
Summary*
A brief summary of the paper.
Year
The year the paper was published.
Venue
The conference the paper was published in. For example, ACL, EMNLP, etc.
Code URL
The link to the code repository for the paper.
Paper URL
The link to the paper.
*These attributes are automatically generated and may require closer inspection. If you find any errors, please report them for correction.
Pipeline Components
Input Encoding
Consists of methods to better encode long documents which cannot be processed (without truncating) by standard Transformer models, for instance, by using hierarchical (or graph) attention, and/or feeding additional context to the model such as the discourse structure of the document (if available), or user-specific aspects.
Unit Relationship
Consists of methods that explicitly consider the relationship between the source document's units (words, sentences, or even passages). A key motivation for this is the information redundancy in a long document, or key connections between various parts of the document that may be ignored via standard encoding of the document.
Data Augmentation
Consists of methods that employ various data augmentation techniques to either introduce specific aspects (facets) into the model, or to create contrastive (or adversarial) examples for robustness, or to simply overcome the lack of suitable data in low-resource domains.
External Knowledge
Consists of methods that use external knowledge to improve the model's performance, for instance, via knowledge graphs, domain-specific vocabularies, or information from pre-trained language models.
Objective Function
Consists of methods that introduce new objective functions alongside standard the cross-entropy loss to better suit the task at hand, for instance, to emphasis on diversity, faithfulness, or custom rewards (in case of reinforcement learning).
Auxiliary Tasks
Consists of methods that employ additional tasks, for instance, via multi-task learning, or via pre-training on related tasks (such as textual entailment, paraphrasing, gap-sentence prediction, etc.) that help the summarization task.
Unit Selection
Consists of methods that explicitly select the units (words, sentences, or even passages) that are most relevant to the summary, for instance via copying or pointing to specific text spans in the source document. A key motivation for this is the loss of information when the model is forced to generate a summary of a fixed length.
Controlled Generation
Consists of methods that guide the model to generate summaries with specific attributes such as style, length, tone etc., for instance, by providing additional text as guidance to condition the summary generation process, or to restrict the model's vocabulary to a specific domain.
Post Processing
Consists of methods that post-process the generated summaries to improve their quality, for instance, via re-ranking, re-writing, or swapping specific spans of the summary to achieve the desired goal.