Related Work and Citation Text Generation: A Survey

• A related-work section should tell a coherent story linking cited works to the current paper, rather than concatenate generic summaries.The paper cites Li and Ouyang (2024) on this writing goal.Source: §1 · p. 1
• RWG combines subtasks such as retrieval-augmented generation, long-document understanding, and query-focused multi-document summarization, making it a test bed for NLP models.NLP researchers are described as natural domain experts for evaluating related-work sections.Source: §1 · p. 1
• Rapid publication makes it time-consuming for researchers to keep up, motivating concise automatic literature reviews that summarize papers in their feeds.Source: §1 · p. 1
• RWG approaches have progressed from rule-based and extractive methods to sentence-level and then section-level abstractive generation, with renewed interest following LLM advances.The survey reviews this history alongside task definitions and ongoing challenges.Source: §1 · p. 1
• Most RWG works are not directly comparable because task definitions and input representations differ; there is no standard benchmark, and unreleased models or outputs impede comparison.The survey also discusses ethical concerns, including plagiarism, non-factual statements, and effects of fully automatic RWG on scientific thinking and writing.Source: §1 · pp. 1–2
• Table 1 compares task definitions across extractive, single-citation, and full-section abstractive RWG approaches, with symbols marking selected task variants.The caption identifies variants including multi-sentence citations, full-text features, human editing of citation groupings, and LLM prompts.Source: Table 1 · p. 2

2 Task Definition

• RWG task definitions vary with summarization approaches, including the system inputs available and whether generation targets a citation or a larger section.Source: §2 · p. 2

2.1 Extractive Related Work Generation

• Extractive RWG selects salient cited-paper sentences; some approaches mine other citing papers, while ordering is either supplied or predicted.Hoang and Kan (2010), Wang et al. (2018), Chen and Zhuge (2019), and Wang et al. (2019) assume an ordering; Hu and Wan (2014) and Deng et al. (2021) predict one.Source: §2.1 · p. 2

2.2 Abstractive Related Work Generation

• Abstractive RWG includes two task forms: generating single-citation texts or generating paragraphs or full related-work sections.Source: §2.2 · p. 2

2.2.1 Citation Text Generation

• Single-citation generation arose under neural models’ input-length limits, using citation context and cited-paper abstracts rather than full papers; works vary in citation length and number of cited papers.AbuRa’ed et al. (2020), Xing et al. (2020), Ge et al. (2021), Luu et al. (2021), and Gu and Hahnloser (2023) use single sentences; some later works allow multiple sentences, shorter spans, or multiple cited papers.Source: §2.2.1 · pp. 2–3

2.2.2 Section-Level Generation

• Section-level methods generate paragraphs with multiple citations; later work organizes citations into paragraphs and adds transitions, while LLM context windows enable full-section generation in one pass.Chen et al. (2021, 2022) used target-paragraph organization; Liu et al. (2023), Li and Ouyang (2024), and Martin-Boyle et al. (2024) address citation organization and transitions.Source: §2.2.2 · p. 3

3 Overview of Approaches

• Most RWG works skip retrieval by using the gold cited-paper list; extractive approaches concatenate selected sentences, while abstractive approaches generate summaries, often without explicit salience modeling.The survey describes these as common patterns within approach categories, not an exhaustive account of every method.Source: §3 · p. 3

3.1 Representing Cited Papers

• Cited-paper abstracts commonly proxy for full text; Cited Text Spans (CTS) instead provide citation-specific evidence for extract-then-abstract generation, while some section-level systems encode citation graphs.Li et al. (2023) argue abstracts may not contain enough information to ground a target citation; graph approaches range from pruning salient-sentence search to adding network information as model input or an LLM-generated relationship description.Source: §3.1 · pp. 3–4

3.2 The Importance of Citation Context

• Target-paper context signals salient topics and restricts RWG; conditioning on it improves coherence, and readers preferred citations generated with the full context embedding the target citation.Extractive systems also use context as a query to score cited-paper sentences; some early approaches omitted target context, whereas later works used it.Source: §3.2 · p. 4

3.3 Applying Citation Analysis

• Citation-function, intent, and sentiment labels have been used to improve RWG; LLM-generated descriptions of citation functions address ambiguity in simple label sets.Li and Ouyang (2024) found a strong correlation between reference-type citation proportions and human preference scores, and concluded readers prefer integrative RWS supported by reference-type citations; Martin-Boyle et al. (2024) found human-written and human-assisted RWS had more cited papers per sentence than pure machine-generated RWS.Source: §3.3 · pp. 4–5

3.4 Human-Assisted Generation

• Human input helps address readers’ sensitivity to paper organization and emphasis, which SOTA LLMs do not yet manage without guidance.Approaches use user keywords or human-written summaries to guide generation, or let users edit predicted paper groupings before generation.Source: §3.4 · p. 5

4 Datasets

• Despite twenty published RWG works, there is no standard benchmark dataset because studies define the task differently and adapt datasets to their own definitions.Table 3 lists common dataset sources; Appendix Table 6 gives work-level dataset details.Source: §4 · p. 5, Table 3 · p. 6

The ACL Anthology Network (AAN) Corpus

• Common sources include AAN, its SciSummNet subset, Delve, S2ORC, and CORWA, with resources varying in citation annotations, paper coverage, and availability.AAN provides sentences citing each paper; SciSummNet adds human-validated citation sentences and summaries, while CORWA labels citations as dominant or reference.Source: The ACL Anthology Network (AAN) Corpus · p. 5

4.2 Discussion

• Missing cited papers can disrupt section-level RWS flow, while most RWG datasets draw on NLP papers and almost none on work outside computer science.The survey attributes NLP’s prevalence partly to extractable RWS sections, researchers’ domain expertise, and access to human judges.Source: §4.2 · p. 6
• LLM-based RWG must account for possible training-data overlap, so recent approaches explicitly target recent papers.Source: §4.2 · p. 6

5.1 Baselines

• Extractive RWG works commonly use LEAD, MEAD, LexRank, and TexRank, while abstractive works use sequence-to-sequence baselines such as PTGEN, BertSumAbs, and Longformer Encoder-Decoder.These general-purpose baselines are relatively easy to replicate, unlike specialized RWG approaches, which are rarely reused as baselines.Source: §5.1 · p. 6

5.2 Metrics

• Almost all RWG works use ROUGE for automatic evaluation, and Luu et al. (2021) additionally use BLEU.Source: §5.2 · p. 6
• Human evaluations commonly rate fluency, coherence, and relevance or informativeness, but samples are relatively few because expert judging is difficult and time-consuming.Most works evaluate at least 15 samples with three judges per sample, using five-point Likert scales.Source: §5.2 · pp. 6–7

6.1 Lack of Comparability

• No standard RWG benchmarks exist, and task definitions, datasets, and evaluation practices vary, limiting direct comparison across systems.Works differ in what counts as a citation or related work section, in source-corpus annotation, and in evaluation criteria.Source: §6.1 · p. 7
• Single-sentence citation assumptions can leak target-citation information into context when human citations span multiple sentences.Li et al. (2022) argue that real citations can exceed one sentence or discuss multiple papers, making narrower assumptions unrealistic.Source: §6.1 · p. 7

6.2 Common Limitations

• Only four of twenty surveyed works predict citation ordering or grouping; readers disliked grouping errors, and organization differed with human-assisted grouping.Li and Ouyang (2024) report reader reactions; Martin-Boyle et al. (2024) report differences in organization.Source: §6.2 · pp. 7–8
• Modern abstractive systems often underprovide transitions and connections among cited works, leaving transition sentences and multi-paper citations open problems.Li and Ouyang (2024) report that readers wanted more transitions; Martin-Boyle et al. (2024) find generated sections do not connect cited papers enough.Source: §6.2 · p. 8
• Existing RWG systems assume the cited-paper list is known, while reviewers’ missing-citation questions illustrate why retrieval could be useful.Li and Ouyang (2024) report judges wanted systems that could surface additional relevant papers; the survey identifies RAG as a promising direction.Source: §6.2 · p. 8

6.3 Ethical Concerns

• Extractive RWG copies cited-paper sentences, while abstractive RWG can still produce plagiarized or hallucinated text.Extractive outputs are less likely to contain factual errors about cited papers, whereas abstractive systems may copy input text and factual consistency remains an active concern.Source: §6.3 · p. 8
• There is no scientific-community consensus on whether submitting automatically generated related work raises academic-dishonesty concerns, including after editing.The survey urges caution for RWG researchers and users.Source: §6.3 · p. 8
• Because writing a literature review can be a learning and thinking experience, the survey urges human-in-the-loop RWG frameworks.It points to Gu and Hahnloser (2023), Li and Ouyang (2024), and especially Martin-Boyle et al. (2024).Source: §6.3 · p. 8

A Appendix

• The appendix summarizes problem formulations for extractive and neural network-based RWG.The extractive formulations share a generation target of extracted-sentence sequences; the neural RWG table defines context as surrounding sentences or paragraphs.Source: Table 4 · p. 14, Table 5 · p. 14
• The appendix includes summary tables for RWG datasets and prior approaches.Source: Table 6 · p. 15, Table 7 · p. 16
• The appendix summarizes evaluation methods for extractive and abstractive RWG and perspectives for human evaluation.Source: Table 8 · p. 16, Table 9 · p. 17, Table 10 · p. 18