This category includes challenges such as difficulty in generating style-tailored summaries, difficulty in controlling the length of the summaries, inability to guarantee the inclusion of specific named entities in the summaries, and inability to generate summaries that are consistent with the user's interests.

This category includes challenges such as difficulty in modeling highly representative latent vectors for long documents, inability of the sequence-to-sequence models to capture long range dependencies among the document's contents, loss of information due to truncating long inputs, and summary vector degeneration due to simple averaging of the latent vectors.

This category includes challenges such as difficulty in modeling the hierarchical structure of long documents such as scholarly documents, abd the inability to leverage the discourse structure, topical structure or the narrative flow emphasized by the document layout.

This category includes challenges such as unfaithful and factually incorrect summaries resulting from hallucinating information that is not present in or inconsistent with the source document. Hallucinations may be intrinsic (wrongly modifying the original contents) or extrinsic (adding inconsistent novel content that is not present in the original document).

This category includes challenges such as difficulty in identifying the most important or salient or relevant contents in the document to be included in the summary.

This category includes challenges of extractive summarization approaches such as difficulty in generating summaries that reflect a global coverage of the document's content, are coherent, non-redundant and read naturally like human-written summaries.

This category includes challenges such as lack of high-quality, large-scale, and labeled training data for training supervised models for summarization. A related challenge is the cost and time required to manually annotate this large-scale training data.

This category includes challenges such as difficulty in properly analyzing the impact of pre-training models and auxiliary tasks for improving performance on downstream summarization task, and difficulty in training the models with limited labeled data or selecting a subset of efficient samples that are representative of the data distribution and can lower the computation costs of training.

This category includes challenges such as limitations of ROUGE, lack of metrics that balance lexical and semantic similarity of the document-summary pair, lack of metrics that include faithfulness and factuality into the summary quality evaluation, and have a high correlation with human judgments.