Source-linked AI summary
Towards Generating Citation Sentences for Multiple References with Intent Control
Jia-Yan Wu, Alexander Te-Wei Shieh, Shih-Ju Hsu, Yun-Nung Chen
TL;DR
Citation-generation systems were limited to single citations and lacked intent control, despite writers’ need to discuss multiple studies with different intents. The paper introduces CiteMI and an intent-conditioned Fusion-in-Decoder approach; it reports improved automatic evaluation when intent labels are added.
Problem
Prior citation-generation systems focused on one cited document and did not provide controllable citation intent.
Source: §1 · p. 1Method
The paper collects CiteMI for multiple cited documents and incorporates predicted citation intents into a Fusion-in-Decoder generation model.
Source: §1 · pp. 1–2, §3 · p. 2Results
Takeaways & Limitations
The study demonstrates the feasibility of generating citation text from multiple abstracts while supporting intent control.
Source: §7 · p. 4CiteMI covers computer science and natural language processing, and some citation intents were predicted rather than accurately labeled.
Source: §6 · p. 4
Abstract
from arXiv · showhide
Machine-generated citation sentences can aid automated scientific literature review and assist article writing. Current methods in generating citation text were limited to single citation generation using the citing document and a cited document as input. However, in real-world situations, writers often summarize several studies in one sentence or discuss relevant information across the entire paragraph. In addition, multiple citation intents have been previously identified, implying that writers may need control over the intents of generated sentences to cover different scenarios. Therefore, this work focuses on generating multiple citations and releasing a newly collected dataset named CiteMI to drive the future research. We first build a novel generation model with the Fusion-in-Decoder approach to cope with multiple long inputs. Second, we incorporate the predicted citation intents into training for intent control. The experiments demonstrate that the proposed approaches provide much more comprehensive features for generating citation sentences.
1 Introduction
Citation generation research had focused on one cited document and did not control citation intent, despite writers needing multiple studies and context-dependent intents. The paper introduces CiteMI and uses intent-conditioned Fusion-in-Decoder generation to address these gaps.
1 Introduction
- Existing systems generated one sentence for one citation without controllable intent, while writers may need different intents for the same cited work depending on context.Citation intents include background, method, and result comparison.Source: §1 · p. 1
- The paper introduces CiteMI for multiple cited documents, predicts citation intents, and incorporates those intents into generation.The intent prediction is trained on prior single-citation data and the newly collected multi-citation corpus.Source: §1 · pp. 1–2
- Fusion-in-Decoder is chosen to process long inputs from multiple documents while using pretrained contextual language models such as T5 for generation.FiD had originally been used for generative open-domain question answering.Source: §1 · p. 2
- The work frames multiple cited documents and intent control as two unmet features of citation sentence generation.It presents CiteMI as a dataset addressing both problems and FiD as the model approach.Source: §1 · p. 2
2 Related Work
Prior work studied single-document citation generation, citation-intent classification, and controllable extreme summarization. These lines of work provide related datasets and methods but do not describe the multiple-reference, intent-controlled task addressed here.
2.1 Citation Text Generation
- Xing et al. (2020) generated citation text from citing context and a cited abstract, using explicit/implicit citation labels and a PG-net with cross-attention.Luu et al. (2020) also studied one source and one cited document, using GPT-2; generation models did better in human evaluation while retrieval scored better automatically.Source: §2.1 · p. 2
2.2 Citation Intents
- Cohan et al. (2019) introduced SciCite, with 11,020 crowdsourced citation intents categorized as Background, Method, or Result comparison.Other work used multi-task prediction with section-title and citation-worthiness tasks, while Pride and Knoth (2020) collected six intent types from authors.Source: §2.2 · p. 2
2.3 Extreme Summarization
- SciTLDR’s CATTS approach uses control codes to generate a title or TLDR from a paper’s abstract, introduction, and conclusion.Its short TLDRs resemble background citations but are not conditioned on the citing document.Source: §2.3 · p. 2
3 Task Description and Data
CiteMI defines citation generation from a citing paper, multiple consecutively cited papers, and their citation intents. Its data combines multi-reference ACL examples with single-reference sources and includes both provided and predicted intent labels.
3 Task Description and Data
- CiteMI maps a citing document’s title and abstract, cited documents’ titles and abstracts, and their intents to the corresponding citation sentences.Consecutively cited sentences are explicit citation sentences separated by one sentence without an explicit citation; the sentences were extracted with handcrafted regular expressions.Source: §3 · p. 2
- The multi-reference examples came from ACL Anthology proceedings published during 2015–2020, while single-reference examples were added from Xing et al. (2020) and SciCite.The SciCite papers’ abstracts were crawled because they were absent from the original release.Source: §3 · pp. 2–3
- Table 1 illustrates the task using citing and cited abstracts plus intent as inputs, and shows FiD generations with and without intent.The table caption describes an example rather than benchmark results.Source: Table 1 · p. 3
- SciCite examples use provided intent labels; a BERT-based model trained on SciCite predicted labels for the rest of CiteMI, with 84% testing accuracy on SciCite.The dataset contains 56,955 single-citation examples and 4,189 multiple-citation examples.Source: §3 · p. 3
4 Methods and Experiments
The experiments compare retrieval baselines with a T5-based Fusion-in-Decoder generator. FiD encodes cited documents separately, adds intent information, and scales its computation linearly with the number of cited documents.
4.1 Retrieval Baseline
- The oracle retrieval baseline selects the cited-abstract sentence most cosine-similar to the target citation, using mean BERT token embeddings.A second baseline applies the same sentence comparison to the citing abstract instead.Source: §4.1 · p. 3
4.2 Generative Models
- FiD augments T5 by encoding the citing abstract with each cited document, prepending an intent code, and letting the decoder attend over encoder blocks.The target citation text replaces the relevant citation with a cited-document marker; supportive and not_supportive are labels for Result comparison.Source: §4.2 · pp. 3–4
- FiD computation grows linearly with the number of cited documents, unlike the quadratic growth described for using a single encoder block.Generation experiments used 80% of CiteMI for training, 10% for validation, and 10% for testing.Source: §4.2 · p. 4
5 Results and Analysis
Adding intent labels improved FiD’s CiteMI performance across every evaluated automatic metric. Round-trip intent prediction was more accurate when intent was specified, while abstracts alone still provided some clues about underlying intent.
5 Results and Analysis
- Adding intent labels unanimously improved FiD’s CiteMI results across all evaluated metrics.The evaluation included BLEU, ROUGE, Meteor, and BLEURT.Source: §5 · p. 4
- 86.39% micro-average intent accuracy with specified intent fell to 70.14% when intent was omitted.This round-trip prediction evaluated intent expressed in generated citation text.Source: §5 · p. 4
- Abstracts without intent labels still revealed some clues about the underlying intent.The authors also report that specifying intent could effectively change the intent expressed by generated text.Source: §5 · p. 4
6 Future Work
The authors propose expanding CiteMI beyond computer science and NLP and improving its intent labels. They also flag possible hallucinations from abstractive generation and the need to examine generated content for correctness.
6 Future Work
- CiteMI covers computer science and NLP, so broader domains such as medicine remain outside its current data scope.The authors call for multi-domain data to build a more comprehensive model.Source: §6 · p. 4
- Some citation intents were predicted, and the authors call for more accurately labeled intent data.They also suggest further augmentation of CiteMI.Source: §6 · p. 4
- Abstractive generation may produce hallucinated or unfaithful content, making correctness checks important.The authors propose including methods to examine the correctness of generated results.Source: §6 · p. 4
7 Conclusion
The paper introduces CiteMI and uses predicted intent labels with Fusion-in-Decoder to generate citation sentences from multiple abstracts. Experiments show intent labels improve automatic evaluation metrics, while the authors expect more accurate labels and more multi-citation examples to further enhance quality.
7 Conclusion
- CiteMI contains single- and multiple-citation examples in computer science, and predicted intent labels provide control over generated content.The authors use Fusion-in-Decoder to handle long inputs containing multiple abstracts and report feasibility of realizing both features.Source: §7 · p. 4
- Adding intent labels improved the generation model’s performance across various automatic evaluation metrics.The authors expect more accurate intent labels and more multiple-citation examples to further enhance overall quality.Source: §7 · p. 4