Automatic Related Work Generation: A Meta Study

1 INTRODUCTION

• Related work sections are repetitive, stand-alone literature reviews, motivating systems that automatically draft them after the main paper is written.Authors use these sections to compare their contributions with prior work and establish novelty.Source: §1 · p. 1
• The task takes a target scientific document and cited papers as input, framing related work generation as scientific multi-document summarization.Scientific articles require combining different arguments, viewpoints, and technical concepts across studies.Source: §1 · pp. 1–2
• Accurate generation also requires citation and discourse analysis because the system must model relationships and interactions among multiple documents.Existing studies had not yet incorporated this broader view beyond central summarization.Source: §1 · p. 2
• The task received little attention for over 10 years before recent neural systems, and its formulation, datasets, and evaluation remain difficult to compare across studies.The survey therefore compares prior work across problem formulation, dataset collection, methods, results, evaluation, future work, and related tasks.Source: §1 · p. 2

2 TERMINOLOGIES

• A Related Work Section is a stand-alone literature review section, typically found in NLP papers and called a literature review in some other fields.Source: §2 · p. 2
• Related Work Generation means automatically generating related work sections, an activity also called related work summarization in early literature.Source: §2 · p. 2
• The Target Paper is the paper whose related work section will be generated and is also called the citing paper in citation analysis.Source: §2 · p. 2
• A Cited Paper is a paper referenced by the target paper and is also called a reference paper.Source: §2 · p. 2

3.1.1 First Work of Related Work Generation.

• Hoang and Kan introduced related work generation as topic-biased multi-document summarization, with RoWoS producing human scores above 3.5/5 but relying on only 20 articles and a pre-built topic tree.The task decomposes into finding relevant documents, identifying salient aspects, and generating a topic-biased summary.Source: §3.1.1 · p. 3

3.1.2 Extractive Related Work Generation with Topic Modeling.

• Chen and Zhuge used keyword-based Steiner-tree coverage and greedy sentence selection, with ROUGE showing RWS-Cit outperforming their replication of RoWoS.Their setting also incorporates papers that cite the target paper’s cited papers.Source: §3.1.2 · p. 4

3.1.3 Identifying Cited Text Span for Extractive Related Work Generation.

• Extractive systems expanded topic modeling with cited-text-span selection, citation sentences from co-citing papers, and neural sentence extraction and reordering in SERGE.SERGE automatically constructs training data with ROUGE-based retrieval, then combines BERT, keyword features, and MobileBERT.Source: §3.1.4 · p. 4, §3.1.3 · p. 4

3.1.4 Sentence Extraction and Reordering.

• SERGE avoids requiring citation context during training and inference, which the authors claim prevents delays when new publications are cited.This design is presented as a consequence of not directly requiring citation context.Source: §3.1.4 · p. 4

3.2 Abstractive Related Work Generation

• Recent work shifts toward bottom-up abstractive generation, treating citation text generation and document-relationship explanation as related-work-generation variants.These approaches generate fine-grained tokens or sentences as building blocks for complete sections.Source: §3.2 · p. 4

3.2.1 Neural Network-Based Abstractive Related Work Generation.

• AbuRa’ed et al. applied Pointer-Generator Networks and Transformers to citation sentences, introducing abstractive summarization to related work generation without a new architecture.The input is the concatenation of a cited paper’s title and abstract.Source: §3.2.1 · pp. 4–5

3.2.2 Abstractive Citation Text Generation Given Contexts.

• Xing et al. modeled citation contexts with PTGEN-Cross, while Ge et al. added citation-network, salience, and citation-function signals through joint multi-task learning.Xing et al. restrict the setting to single-citation sentences and two input sources; Ge et al. inherit the single-citation limitation.Source: §3.2.2 · p. 5, §3.2.3 · p. 5

3.2.4 Capturing Relations Between Scientific Papers.

• Chen et al. generated paragraphs containing multiple citations from multiple abstracts, whereas Luu et al. [75] explained two-document relationships using SciGPT2, retrieval, entity enrichment, and candidate ranking.Chen et al. use a relation-aware generator and address a setting closer to real related-work sections; Luu et al. [75] face a 512- or 1024-token context-window bottleneck and do not cover sentences citing multiple documents.Source: §3.2.4 · p. 5, §3.2.5 · p. 6

4 TASK DEFINITION AND PROBLEM FORMULATION

• The task combines a target scientific document and cited papers to produce a contextually related, topic-biased summary through document selection, salient-aspect identification, and generation.Source: §4 · p. 6
• The loose task definition leaves format, length, writing style, target-paper scope, and cited-paper set unspecified, allowing multiple valid interpretations.Source: §4 · p. 6

4.1 Extractive Related Work Generation

• Extractive systems vary in their inputs, including target-paper sections, cited papers, co-citing papers, citation sentences, and occasionally a pre-built topic tree.Source: §4 · p. 6, §4.1 · pp. 6–7
• Extractive outputs are sequences of selected sentences concatenated into one paragraph, lacking the paragraph divisions and salient structure typical of NLP related-work sections.Source: §4.1 · p. 7, Table 1 · p. 7

4.2 Abstractive Related Work Generation

• Abstractive systems usually restrict the task to citation-sentence generation because producing long related-work sequences is more challenging.Source: §4.2 · p. 7
• Most abstractive formulations use one cited paper at a time, so their outputs describe a single cited work rather than relationships among multiple cited papers.Source: §4.2 · p. 8
• Chen et al. uniquely generate an entire paragraph containing two or more citations, making their formulation closer to practical related-work generation.Source: §4.2 · p. 8
• Overall, abstractive approaches proxy full cited papers with abstracts and simplify related-work generation into single-paper biased summarization.Source: §4.2 · p. 8, Table 2 · p. 8

5 DATASETS AND PREPROCESSING

• Online academic papers, especially open-access papers, make related-work datasets inexpensive to collect and provide abundant training material.Source: §5 · p. 8

5.1.1 ACL Anthology Network.

• S2ORC provides 81.1M English papers, structured full text for 8.1M open-access papers, and automatic annotations for citations, figures, and tables across many disciplines.Source: §5.1.2 · p. 9

5.2 Overview of Datasets Used in Related Work Generation

• Prior datasets mostly focus on computational linguistics because researchers find unfamiliar academic domains challenging to read.Source: §5.2 · p. 9

5.2.1 Extractive Approaches.

• Extractive datasets range from manually annotated small collections, such as RWSData’s 20 articles, to automatically labeled corpora derived from citation sentences and salient cited-paper sentences.Source: §5.2.1 · pp. 9–10

5.2.2 Abstractive Approaches.

• Abstractive datasets use title–abstract–citation pairs, citation contexts, citation-function labels, and large corpora from ACL Anthology, MAG, OAG, S2ORC, and Delve.Source: §5.2.2 · pp. 10–11

5.3 Discussion

• Abstractive datasets are significantly larger than extractive datasets because neural approaches require more training data.Source: §5.3 · p. 11
• Despite abundant papers and released datasets, no widely recognized off-the-shelf dataset exists because studies require incompatible input formats and dataset sizes.Source: §5.3 · p. 11

6 APPROACHES

• The surveyed literature divides into earlier extractive systems and later neural abstractive systems, reflecting the broader evolution of summarization methods.Source: §6 · p. 11, Table 4 · p. 12

6.1.1 Heuristic-based Approach.

• ReWoS uses topic-tree-guided heuristics, template-generated indicative sentences, TF × ISF extraction, contextual windows, and node-position weighting.Source: §6.1.1 · pp. 11–12

6.1.2 Extracting Sentences Based on Topic Modeling.

• Hu and Wan remove the need for manual topic trees by combining PLSA topic information, SVR sentence-importance scores, and global optimization.Source: §6.1.2 · pp. 12–13

6.1.3 Identifying Cited Text Spans.

• Other extractive systems model document relationships with graph-based Steiner-tree selection or combine topic modeling with cited-text-span classification and KL-divergence-based greedy selection.Source: §6.1.2 · p. 13, §6.1.3 · p. 13

6.1.4 Sentence Extraction and Reordering.

• Deng et al. extract salient sentences with BERT and keyword scores, then reorder the top three using MobileBERT-based next-sentence prediction under a Markov assumption.Source: §6.1.4 · pp. 13–14

6.2.1 Neural Network for Abstractive Citation Sentence Generation.

• Abstractive citation generation progresses from sequence-to-sequence models to pointer-generator systems using citation context, cross-attention, citation functions, citation networks, and salient cited-paper sentences.Source: §6.2.1 · p. 14, §6.2.2 · p. 14, §6.2.3 · pp. 14–15

6.2.4 Relation-Aware Related Work Generation.

• Chen et al. generate complete related-work paragraphs by encoding multiple cited abstracts, modeling document relations, and decoding with a pointer-generator-like module.Source: §6.2.4 · p. 15

6.2.5 Fine-tuning Pretraiend Language Model.

• Luu et al. [75] compare SciGPT2-based generation with retrieval, enrich representations using entities or TF-IDF terms, and rank candidates by entity matching to combat hallucinations.Source: §6.2.5 · p. 15

7.1 Evaluation methods

• ROUGE is widely used and fast to compute, but overlap-based scoring misses semantic similarity and can penalize valid summaries containing relevant phrases absent from the reference.All prior works for related work generation use ROUGE [75]; Luu et al. [75] additionally consider BLEU.Source: §7.1.1 · p. 16, §7.4 · p. 20

7.2 Extractive Related Work Generation Performance Evaluation

• Hoang and Kan’s RoWoS system reports ROUGE-1, ROUGE-2, ROUGE-S4, and ROUGE-SU4 recall scores of 0.698, 0.183, 0.218, and 0.298, respectively.Its human scores are 3.6, 3.4, 3.4, and 3.6 for correctness, novelty, fluency, and usefulness.Source: §7.2 · p. 17
• Other surveyed systems use distinct datasets, baselines, metrics, and human protocols, with reported results spanning extractive and abstractive approaches.Examples include ARWG, QueryTopicSum, SERGE, BACO, RRG, SciGen, and the Pointer-Generator and Transformer systems.Source: §7.2 · pp. 18–19, §7.3 · p. 19

7.3 Abstractive Related Work Generation Performance Evaluation

• Abstractive systems report varied ROUGE results, including PTGEN-Cross’s 0.263 ROUGE-1, 0.075 ROUGE-2, and 0.205 ROUGE-L F1 scores.PTGEN-Cross’s human ratings are 3.79 for readability, 2.77 for content, 2.85 for coherence, and 2.91 overall.Source: §7.3 · p. 19

7.4 Summary

• Extractive systems generally achieve higher ROUGE scores than abstractive systems, although differing task definitions and datasets prevent direct comparison across studies.RRG’s paragraph-level output is particularly low-scoring among abstractive approaches.Source: §7.4 · p. 20
• Human evaluation commonly assesses fluency or readability, coherence with the citing context, and correctness or relevance, but lacks a standardized procedure.Studies vary in sample size, evaluator count, and rating format, including 5-point, 3-point, and yes/no judgments.Source: §7.4 · p. 20

8.1 Overview of Future Works Mentioned

• Hoang and Kan identify reliance on a pre-existing topic tree as a limitation and suggest generating topic trees automatically.The survey also connects automated scholarly-work organization with future digital-library and reference-management applications.Source: §8.1 · pp. 20–21
• Future systems are encouraged to use richer evidence, including citation sentences, broader paper content, co-cited papers, and finer-grained citation purpose.These proposals aim to model relationships among papers and produce more coherent related work.Source: §8.1 · p. 21
• Several studies propose more abstractive and knowledge-informed methods, including entity extraction, knowledge bases, stronger neural models, and paper or abstract generation.Ge et al. specifically propose adapting their framework to Transformer [111] models and refining citation-function categories.Source: §8.1 · p. 21
• Luu et al. [75] propose filtering or altering citation sentences, reusing SciGPT2, curating better training data, and improving factuality and cited-paper modeling.They also report that SciGen often generates vague or generic sentences and that some citation sentences are unsuitable as explanations.Source: §8.1 · p. 21

8.2 Discussion

• Unresolved priorities include reducing vague citation text, improving training-set quality, ensuring factual accuracy, and better modeling of cited papers.These issues remain after several earlier future-work proposals were addressed by subsequent studies.Source: §8.2 · p. 21, §8.1 · p. 21

9.1 Automatic Summarization of Scientific Papers

• Scientific paper summarization includes single-document and multi-document settings, with citation-based summaries offering focused information but inheriting potential citation-source bias.Related work generation differs because its target is specifically the related work section, although outputs can overlap substantially.Source: §9.1 · p. 22
• The survey presents automatic related work generation as a topic-biased multi-document summarization task conditioned on a target paper and cited papers.This task is adjacent to citation-based scientific summarization but does not always share the same generation target.Source: §9.1 · p. 22

9.2.2 Integrative v.s. Descriptive Literature Review.

• Integrative literature reviews emphasize research results, critiques, high-level summaries, and paraphrases, whereas descriptive reviews provide more methods, results, and cut-pasted information.Integrative reviews cite results and conclusions more often; descriptive reviews cite abstracts and introductions more often.Source: §9.2.2 · p. 23, §9.2.1 · p. 23
• The integrative-versus-descriptive distinction reflects writing variation but is coarse because it operates at the document level rather than the paragraph or sentence level.The underlying study analyzes only 20 information-studies papers, and patterns may vary across domains.Source: §9.2.2 · p. 23

9.2.3 Deconstructing Human Literature Reviews.

• Jaidka et al. model literature reviews through document structure, rhetorical structure, and summarization strategies using templates and discourse markers.Their framework represents integrative reviews hierarchically and descriptive reviews more flatly, with different proportions of meta-summary, meta-critique, and methods-and-results elements.Source: §9.2.3 · pp. 23–24
• The template-based system identifies 110 discourse expressions in 1,298 cases and improves over MEAD on ROUGE and human comprehensibility, readability, and usefulness.Its information-selection strategies reflect transformations such as copy-paste and paraphrase.Source: §9.2.3 · p. 24
• Template-based literature-review generation remains limited by the small 30-paper study and manually engineered regular-expression features, motivating stronger neural replacements.The survey nevertheless regards the document-level generation intuition as valuable.Source: §9.2.3 · p. 24

9.3 Citation Analysis

• Citation analysis covers citation function, citation intent, and citation sentiment, and Table 8 summarizes datasets proposed by the surveyed works.Source: §9.3 · p. 24, Table 8 · p. 25

9.3.1 Citation Function and Intent.

• Teufel et al. introduced citation-function annotation with weakness, contrast, agreement or usage, and neutral categories over 548 citations from 360 computational-linguistics articles.Their coarse four-label scheme achieved higher macro-F1 because the original labels were unbalanced.Source: §9.3.1 · p. 24
• Later studies broadened citation-function modeling through richer features, domain-specific labels, resource roles, and larger datasets, including SciRes and algorithm-focused corpora.Approaches included ensemble classifiers, self-training, class balancing, and multitask BERT-based classification and resource ranking.Source: §9.3.1 · pp. 24–26
• SciCite classifies 11020 citation sentences from 6627 papers into Background, Method, and ResultComparision using multitask BiLSTM models with auxiliary citation tasks.The method also improved performance on the ACL-ARC citation dataset.Source: §9.3.1 · p. 26
• MultiCite replaces single-sentence single-intent analysis with a multi-sentence multi-intent framework and evaluates multilabel classification and question-answering models.Its labels include Background, Motivation, Uses, Extends, Similarities, Differences, and Future Work.Source: §9.3.1 · p. 26

9.3.2 Citation Sentiment.

• Using context sentences, Athar and Teufel annotated 1741 citations and identified significantly more citation sentiment labels than Athar, mainly because they labeled more implicit citations such as “this method” and “the method”.They considered two sentences each before and after an explicit citation reference.Source: §9.3.2 · p. 26

Discourse Analysis

• Argumentative Zoning segments scientific articles into discourse zones such as Background, Aim, Contrast, Basis, and Own for classification, summarization, and citation analysis.Teufel et al. and Teufel and Moens use a manually annotated corpus and a naive Bayesian classifier.Source: §9.4 · p. 27
• Medline abstract studies formulate rhetorical-section identification as sequence labeling over author-provided section headings, using CRFs on automatically collected corpora.The categories include objective, methods, results, conclusions, background, and related sections.Source: §9.4 · p. 27
• CoreSC and AZ-II annotate full chemistry papers with layered scientific concepts, properties, concept identities, and fine-grained discourse labels.Liakata et al. also compare their label distributions and recognize CoreSC automatically with SVM and CRF models.Source: §9.4 · pp. 27–28
• Biology-focused schemes label clause-level discourse types including Goal, Fact, Result, Hypothesis, Method, Problem, and Implication, linking rhetorical structure to tense.The work targets experimental biology papers rather than general scientific text.Source: §9.4 · p. 28
• The INTACT corpus combines discourse segments, evidence fragments linked to experimental sub-figures, and paragraph-level experiment types, enabling several downstream tagging tasks.Subsequent work improves these tasks with attention-based RNNs, rules, transfer learning, LSTMs, CNNs, and multimodal Transformers.Source: §9.4 · p. 28

9.5 Cited Text Span

• Cited Text Spans (CTS) identify the specific reference-paper sentences cited by each citation in a target paper and support automatic CTS retrieval.AbuRa’ed et al. also collect papers that cite each referenced paper by extending Hoang and Kan ’s RWSData dataset.Source: §9.5.1 · p. 28

9.5.1 Explicit Annotation of Cited Text Span.

• CTS annotation has low agreement, averaging 0.52 pairwise agreement, because citation networks do not record CTS and the spans are naturally ambiguous.The refined dataset removes controversial annotations, while retrieval precision remains unsatisfactory.Source: §9.5.1 · p. 28
• The CL-Scisumm shared task links citing sentences to sentence fragments or short spans, assigns facet types, and generates summaries from the cited text spans.ScisummNet expands the resource to 1000 papers, and the task also advocates combining CTS with cited-paper abstracts.Source: §9.5.2 · p. 28

10.1 Extractive Related Work Generation

• Extractive systems select sentences from target and cited papers using topic models, producing topic-biased and comprehensive sections without generating new text.Their design therefore emphasizes sentence selection and coherence rather than neural generation.Source: §10.1 · p. 29
• High ROUGE from natural extracted sentences does not prevent low human-rated coherence and fluency, while copying prior sentences may create intellectual-property concerns.Naive aggregation fails because related work is not simply a collection of independent cited-paper summaries.Source: §10.1 · p. 29

10.2 Abstractive Related Work Generation

• Abstractive systems use pointer-generator networks and Transformers to fuse multiple-document representations through attention mechanisms.The survey identifies attention-based neural models as a likely future direction for related work generation.Source: §10.2 · p. 29
• Future systems should combine manual and automatic data collection, exploit citation networks and salience signals, use stronger multi-document encoders, and select inputs appropriately.The survey cites RRG and external supervision as examples of these directions.Source: §10.2 · p. 29
• The field lacks standardized task definitions, reusable datasets, released code, and persuasive cross-study comparisons, motivating shared tasks such as CL-Scisumm [50].Most studies collect new datasets and evaluate under different assumptions and metrics.Source: §10.2 · p. 29, §10.3 · p. 30

10.3 General Discussions

• Current pipelines commonly simplify the problem by treating sentences as the basic unit and abstracts as representations of cited papers, despite multi-citation sentences and full-text evidence.Efficient encoders for full-text papers remain a necessary direction.Source: §10.3 · p. 30
• Transformer systems can generate fluent but hallucinated citation text, while coherent long-form related work remains underdeveloped and factuality is difficult to evaluate.Knowledge-base supervision is suggested, but available resources and practical methods remain limited.Source: §10.3 · p. 30
• The surveyed evidence is concentrated in computational linguistics, making expansion to natural and social-science domains challenging but necessary.Source: §10.3 · p. 31

• The meta-study compares limited prior work across problem formulation, dataset collection, approaches, results, evaluation, and future work.Source: §11 · p. 31
• Combining manual and automatic data collection, external supervision, sophisticated multi-document encoders, and input filtering improves related work generation systems.Source: §11 · p. 31
• The field lacks focused studies, standardized task definitions and datasets, and methods that represent full papers without sentence-level assumptions.Source: §11 · p. 31
• Future work must address factuality and hallucination, knowledge-driven approaches, full related work construction, and expansion beyond computational linguistics.Source: §11 · p. 31