Source-linked AI summary
Verifiable Generation with Subsentence-Level Fine-Grained Citations
Shuyang Cao, Lu Wang
TL;DR
Prior work largely targets sentence-level citations, leaving less precise indication of which parts of a sentence sources support. This paper introduces SCiFi, a 10K-example benchmark of Wikipedia paragraphs with subsentence citations, and evaluates LLMs and document-processing strategies. Experiments find that complete source-document context improves citation quality, and highlight citation-rich training as an important factor.
Problem
Prior work mainly targets sentence-level citations, which do not precisely locate the parts of a sentence supported by sources.
Source: §1 · p. 1Method
The paper collects SCiFi, a dataset of 10K Wikipedia paragraphs with subsentence-level citations, and evaluates LLMs using document-processing strategies and citation-rich training samples.
Source: §1 · p. 2Results
Takeaways & Limitations
Subsentence-level citations provide more precise localization of supported content, which the paper argues is critical for transparency and trustworthiness.
Source: §1 · p. 1Attribution quality is assessed with an entailment model designed for short passages, applied to extracted facts and long source documents; specialized long-document entailment models could improve evaluation accuracy.
Source: Limitations · p. 5
Abstract
from arXiv · showhide
Verifiable generation requires large language models (LLMs) to cite source documents supporting their outputs, thereby improve output transparency and trustworthiness. Yet, previous work mainly targets the generation of sentence-level citations, lacking specificity about which parts of a sentence are backed by the cited sources. This work studies verifiable generation with subsentence-level fine-grained citations for more precise location of generated content supported by the cited sources. We first present a dataset, SCiFi, comprising 10K Wikipedia paragraphs with subsentence-level citations. Each paragraph is paired with a set of candidate source documents for citation and a query that triggers the generation of the paragraph content. On SCiFi, we evaluate the performance of state-of-the-art LLMs and strategies for processing long documents designed for these models. Our experiment results reveals key factors that could enhance the quality of citations, including the expansion of the source documents' context accessible to the models and the implementation of specialized model tuning.
1 Introduction
Sentence-level citations do not show which parts of a sentence each source supports, making verification less efficient. The paper studies subsentence-level citations and introduces SCIFI to evaluate models producing them.
1 Introduction
- Existing verifiable-generation work typically cites at sentence level, leaving readers to infer which sentence portions each source supports.The introduction links this ambiguity to less efficient verification and reduced understanding of supporting-document scope.Source: §1 · p. 1
- Subsentence-level citations more precisely locate sourced information, supporting easier assessment of accuracy than sentence-level citations.The introduction also notes that finer-grained citations are more prevalent on frequently viewed Wikipedia pages.Source: §1 · pp. 1–2
- SCIFI contains 10K Wikipedia paragraphs with fine-grained citations, candidate source documents, and queries based on paragraph content.The dataset is intended to support training and evaluation of verifiable generation with finer-grained citations.Source: §1 · p. 2
- Complete source-document context improves citation quality, larger models improve answer quality but not citation quality, and fine-grained citation generation requires supervised fine-tuning.The benchmark covers GPT, Llama2, Vicuna, and Mistral models and evaluates citation behavior, citation quality, and answer quality.Source: §1 · p. 2
2 Related Work
Prior work addresses citation generation and fact-based evaluation, but largely leaves subsentence-level citation grounding unaddressed. This work evaluates citation quality and answer coverage using fact-based evaluation.
Verifiable Generation
- Prior verifiable-generation work uses sentence-level citations, leaving the exact supported portions of outputs unspecified [Liu et al., 2023a; Schuster et al., 2023].Earlier approaches include fine-tuning on demonstrations or human-verified model samples, prompting, better-entailing source retrieval, and output refinement.Source: §2 · p. 2
Fact-based Evaluation
- The paper applies fact-based evaluation to both answer coverage and citation quality, extending prior uses for factual entailment and precision.It decomposes outputs into facts and pairs citations with facts using heuristics.Source: §2 · p. 2
3 SCIFI
SCIFI is built from Wikipedia paragraphs with dense citations and is designed for subsentence-level citation research. The section reports that fine-grained citations are prevalent, especially in more-viewed articles.
3 SCIFI
- 85% of 100K randomly selected Wikipedia articles feature subsentence-level citations, and higher-pageview bins show a greater tendency to use them.Articles were sorted into 20 pageview bins using total views over five years.Source: §3 · p. 3, Figure 2 · p. 3
Data Collection
- SCIFI uses cited sources from the host Wikipedia page as candidate documents and focuses on downloadable sources whose text can be extracted uniformly.Citation positions and metadata for all cited-source types are preserved, including sources such as images and PDFs.Source: §3 · p. 3
- SCIFI samples 10K paragraphs from the top quarter of paragraphs with at least three citations ranked by citation density, then prompts GPT-4 to generate a query for each.The queries target paragraph content; samples include the reference paragraph, query, and cited source documents, with 9K training and 1K test samples.Source: §3 · p. 3
Statistics
- Compared with other citation datasets, SCIFI has high citation density, moderate target lengths, and subsentence-level citations rather than only sentence-level citations.The section points to Table 1 for dataset statistics and comparison.Source: §3 · p. 3, Table 1 · p. 3
4 Experiment Setups
The experiments test LLMs generating paragraph-length answers from queries and candidate documents, using strategies to manage lengthy source texts. Evaluation measures citation behavior, citation quality, and answer coverage.
4 Experiment Setups
- Models receive the query, positive sources cited by the reference, and five randomly selected negative sources, shuffled before generation.The full available source collection is too lengthy to input at once.Source: §4 · pp. 3–4
Models
- The document strategies are truncation, source-document summaries, and a two-stage process that selects documents from summaries before writing with selected original texts.The two-stage strategy reduces the document pool for generation to allow more complete context.Source: §4 · p. 4
Evaluation Metrics
- The evaluation measures citation density, citation entailment against generated facts, and reference-fact coverage.Outputs are decomposed into facts and paired with citations using heuristics; reference-fact entailment scores are aggregated as recall.Source: §4 · p. 4
4 Experiment Setups
- More complete source-document context yields better citation quality across document-reading strategies, as summarized by Table 2.Summaries expose major content, while truncation exposes leading content and can prevent accurate connections between generated content and sources.Source: Table 2 · p. 4
5 Results
The results associate more complete source-document context with better citation quality, while larger models improve answer quality without guaranteeing better citations. Fine-grained citation frequency increases after supervised fine-tuning, but citation quality gains vary across models.
5 Results
- More complete source context improves citation quality, while two-stage generation offers the most complete context but depends on strong instruction-following.Summaries convey major content, whereas truncation exposes only leading content and consistently yields worse citation quality.Source: §5 · pp. 4–5
- Larger models increase reference-fact coverage across backbones and strategies, but do not guarantee better citation quality.The results suggest backbone pretraining designs place varying emphasis on citation and attribution capabilities.Source: §5 · p. 5
- Supervised fine-tuning substantially increases fine-grained citation generation, but its effect on citation quality varies across models.Citation density stays stable across document-reading strategies for the same backbone models.Source: §5 · p. 5, Table 3 · p. 5
- Correctly linking sentence parts to supporting documents requires more specialized training beyond encouraging fine-grained citation behavior.The authors identify built-in citation or attribution mechanisms during pretraining as a possible future direction.Source: §5 · p. 5
- The two-stage strategy was tested only with OpenAI GPT families because other models could not consistently follow its output format.Source: §5 · p. 5
6 Conclusions
The paper studies subsentence-level fine-grained citations and introduces SCIFI, a benchmark built from Wikipedia paragraphs. Experiments on SCIFI highlight source context and citation-rich training as factors associated with citation quality.
6 Conclusions
- SCIFI contains 10K citation-rich Wikipedia paragraphs paired with queries and candidate cited sources.Source: §6 · p. 5
- SCIFI supports training and evaluating models on subsentence-level citation generation.Source: §6 · p. 5
- Experiments with state-of-the-art LLMs and document-processing strategies highlight more complete source context and citation-rich training as factors that can advance verifiable generation.Source: §6 · p. 5
Limitations
The study's document-processing strategies are not exhaustive, and its attribution evaluation applies a short-passage entailment model to long source documents.
Limitations
- The three tested document-processing strategies leave room for more sophisticated approaches and techniques that expand model input limits.The authors note these approaches could potentially improve performance on the benchmark.Source: Limitations · p. 5
- The reported findings concern performance within the benchmark study, where expanded input limits could potentially lead to higher performance.Source: Limitations · p. 5
- Attribution quality is evaluated with an entailment model tailored to short passages but applied to long source documents.The authors state that entailment models specialized for long documents could provide more accurate evaluation.Source: Limitations · p. 5
Ethical Considerations
The benchmark supports evaluation of subsentence-level citations that point users to sources for different parts of generated sentences. Its texts and candidate sources have a defined scope, and unreliable candidates may carry risks.
Ethical Considerations
- Subsentence-level citations can provide precise pointers from different parts of output sentences to supporting sources.Source: Ethical Considerations · p. 5
- The dataset uses formally written Wikipedia text and candidate supporting documents from reliable online sources.Source: Ethical Considerations · p. 5
- With unreliable candidate sources, a high-performing model might cite documents containing fake facts and further propagate incorrect information.Benchmark users should consider candidate-source reliability when assessing LLM-based applications.Source: Ethical Considerations · p. 6
A Details of Data Collection
The appendix details SCiFi’s Wikipedia and cited-source collection, topic coverage, citation statistics, and query generation.
A.1 Wikipedia Article and Cited Source Collection
- SCiFi paragraphs come from Wikipedia; downloadable cited sources were retrieved from Internet Archive and parsed with Trafilatura, and the dataset is CC BY 4.0.The passages also note that a sample is shown in Table 13.Source: §A.1 · p. 8
A.2 Topic Distribution
- SCiFi pages cover diverse topics, with their top 10 topics summarized in Table 4.Source: §A.2 · p. 8, Table 4 · p. 8
A.3 Additional Statistics
- SCiFi contains 39,292 sentences, 30.1% with multiple citations; 30.2% of subsentence citations are not located around punctuation.Source: §A.3 · p. 8
A.4 Prompt for Query Generation
- GPT-4 generates one query for each paragraph in SCiFi.Source: §A.4 · p. 8
B Additional Results
The appendix reports that an oracle setup offers little help from removing irrelevant sources, and describes the prompting and fine-tuning procedures.
B Additional Results
- In the oracle setup, models maintain high answer and citation quality, while removing irrelevant source documents offers minimal help.Table 6 reports the oracle-setup evaluation results.Source: B Additional Results · p. 8, Table 6 · p. 9
C.1 Model Prompting
- In-context experiments use two-shot examples; Truncated and Summary prompts are provided, with source-document summaries generated by GPT-3.5 using a 16K context length.Source: §C.1 · p. 8
- The Two-stage strategy selects documents using the current answer and prior selections, then generates the next answer sentence from the selected documents’ original text.Source: §C.1 · p. 8
C.1 Model Prompting; C.2 Supervised Fine-tuning
- Models are prompted once because of API cost; Llama2 fine-tuning uses LoRA for 3 epochs at 5e-5 with batch size 16, rank 32, and α=64.Each tuned model takes 4 hours on 2 Nvidia A40 GPUs.Source: §C.1 · p. 8, §C.2 · pp. 8–9
C.3 Evaluation Metrics
The evaluation decomposes generated sentences into facts, maps citations to those facts, and scores their entailment against source-document chunks.
C.3 Evaluation Metrics
- GPT-3.5 decomposes each output sentence into facts, which are used to evaluate citation and answer quality.Source: §C.3 · p. 9
- Each fact is mapped back to a sentence segment, and generated citations are ranked by distance, favoring citations after the segment when distances tie.Source: §C.3 · p. 9
- Source documents are split into 256-token chunks, and the highest entailment score across chunks becomes the fact’s final score.Source: §C.3 · p. 9