PaperlayerView exact revision on arXiv

Source-linked AI summary

Attribute First, then Generate: Locally-attributable Grounded Text Generation

Aviv Slobodkin, Eran Hirsch, Arie Cattan, Tal Schuster, Ido Dagan

arXiv:2403.17104v3cs.CL

TL;DR

Coarse document- or paragraph-level citations make fact-checking laborious. Attribute First, then Generate selects and plans source spans before generating sentences from them, using those spans as fine-grained attributions. The approach matches or outperforms baselines on task metrics while producing much shorter attributions and shortening verification time.

  • Problem

    Existing attributed generation often cites entire documents or paragraphs, requiring assessors to sift through irrelevant details during verification.

    Source: §1 · p. 1

  • Method

    Attribute First, then Generate selects relevant source details, organizes them into sentence plans, and generates sentences conditioned on those details.

    Source: §1 · pp. 1–2

  • Results

    The pipeline matches or outperforms baselines on task-specific metrics, yields orders-of-magnitude shorter attributions, and has little impact on attribution accuracy.

    Source: §7 · p. 9

  • Takeaways & Limitations

    The approach shortens fact-checking time by almost 50%.

    Source: §7 · p. 9

    The multi-step approach uses more computing resources and runs slower than direct end-to-end generation because it requires several model calls.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Recent efforts to address hallucinations in Large Language Models (LLMs) have focused on attributed text generation, which supplements generated texts with citations of supporting sources for post-generation fact-checking and corrections. Yet, these citations often point to entire documents or paragraphs, burdening users with extensive verification work. In this paper, we introduce a locally-attributable text generation approach, prioritizing concise attributions. Our method, named "Attribute First, then Generate", breaks down the conventional end-to-end generation process into three intuitive steps: content selection, sentence planning, and sequential sentence generation. By initially identifying relevant source segments ("select first") and then conditioning the generation process on them ("then generate"), we ensure these segments also act as the output's fine-grained attributions ("select" becomes "attribute"). Tested on Multi-document Summarization and Long-form Question-answering, our method not only yields more concise citations than the baselines but also maintains - and in some cases enhances - both generation quality and attribution accuracy. Furthermore, it significantly reduces the time required for fact verification by human assessors.

1 Introduction

Coarse citations to whole documents or paragraphs make verification laborious, motivating locally attributable generation with concise, complete source spans. The proposed pipeline selects and plans source content before generating sentences, yielding shorter citations while maintaining or sometimes improving quality and attribution accuracy.

1 Introduction

  • Whole-document or paragraph citations force assessors to sift through irrelevant details, making fact-checking time-consuming and less effective.Attributed generation is intended to support factuality verification and help detect and address factual errors in model outputs.Source: §1 · p. 1
  • The paper formalizes locally-attributable generation as precise, concise snippets that fully cover generated facts, rather than broad document-level citations.Its Attribute First, then Generate approach chooses relevant source details before planning and generating sentences from them.Source: §1 · pp. 1–2
  • The approach is tested on multi-document summarization and long-form question answering with both fine-tuning and few-shot prompting strategies.Content selection is adapted to each task’s needs, such as salience for summarization or query relevance for question answering.Source: §1 · p. 2
  • The approach sometimes surpasses baselines on generation quality, maintains high attribution accuracy, and reduces fact-checking effort by over 50%.The paper also reports significantly shorter citations.Source: §1 · p. 2

2 Locally-attributable Grounded Text Generation

Locally-attributable grounded generation produces task-specific text from a document set while attaching sentence-level citations to supporting source spans. These citations should cover each sentence’s information while targeting the smallest useful units in both output and input.

2 Locally-attributable Grounded Text Generation

  • Given documents D, the task produces sentences that fulfill an underlying task and carry citations to supporting text in D.The documents may come from a retrieval system or be supplied by the user.Source: §2 · p. 2
  • Attributions should support as much of the output as possible while reaching the highest granularity in both source and generated text.Source: §2 · p. 3
  • Output granularity assigns each generated sentence its own attribution set, while input granularity targets specific, concise source spans.A sentence’s citations collectively must fully back its information; spans may be non-consecutive.Source: §2 · p. 3

3 Modeling

The modeling framework separates grounded generation into content selection, sentence planning, and sentence-by-sentence generation. It is applied through in-context prompting and fine-tuning, with task-specific content selection for summarization and question answering.

3.1 Attribute First, then Generate

  • The framework’s three stages are content selection, sentence planning, and sentence-by-sentence generation, tested on MDS and LFQA with task-specific selection.For summarization, selection targets salient content; for question answering, it targets query-relevant content.Source: §3.1 · p. 3
  • Selected source spans serve as attributions; planning groups them into sentence-fit clusters, and generation uses each cluster with preceding sentences.Planning simplifies generation by conditioning each iteration on one cluster, while preserving sentence-level attribution.Source: §3.1 · pp. 3–4

3.2 In-context Learning

  • The in-context strategy prompts models to select verbatim salient spans for MDS or query-answering spans for LFQA, cluster them, then generate one sentence per cluster using prior sentences.Identified spans are located in the documents by string matching, and undetectable spans are omitted.Source: §3.2 · p. 4
  • A Chain-of-Thought variant combines planning and sentence generation, having the model plan highlights for each sentence before producing and refining the output.Each output sentence is linked to the highlights from its planning stage.Source: §3.2 · p. 4

3.3 Fine-tuning

  • Fine-tuning trains components to copy selected source segments and arrange highlighted spans into ordered clusters, using constrained decoding for lexical fidelity.Generation uses special tokens to distinguish highlighted documents, the prefix, and, for LFQA, the query.Source: §3.3 · pp. 4–5
  • A joint fine-tuning strategy combines selection and planning by mapping the document set directly to ordered highlight clusters.Constrained decoding is used to ensure exact copying of source spans.Source: §3.3 · p. 5

4 Experimental Setup

The setup adapts source-output alignments to train the framework’s phases, compares in-context and fine-tuned models with baselines, and evaluates output and attribution quality using automated metrics and human judgments.

4.1 Datasets

  • For MDS, alignments come from DUC, TAC, and Multi-News, while LFQA uses human-annotated response-citation alignments filtered for verifiability.Source-output alignments supply phase-specific targets: source spans for selection, grouped spans for planning, and output sentences for generation.Source: §4.1 · p. 5

4.2 Models and Baselines

  • The study tests Gemini in-context learning and fine-tunes PRIMERA for MDS and LFQA, comparing against end-to-end task prompts and ALCE for in-context models.PRIMERA was selected for its MDS performance and its development-set advantage over Long-T5 on LFQA; the Gemini API was used for ICL.Source: §4.2 · p. 5, §4.3 · p. 5

4.3 Evaluation

  • Automated evaluation uses ROUGE-L and BERTScore for task output, AUTOAIS for attribution quality, cited-text token length for conciseness, and non-attributed sentences per instance.AUTOAIS uses TRUE in an NLI framework and averages sentences the classifier predicts as attributed.Source: §4.3 · pp. 5–6
  • Human evaluation rates coherence on a five-point Likert scale, LFQA helpfulness, and sentence-level attribution support; fluency is evaluated only by humans.Attribution judgments score full support as 1 and otherwise 0, averaged across sentences; TRUE was found to correlate better with human judgments than TrueTeacher.Source: §4.3 · p. 6

5 Results and Analyses

The approach produces much shorter attributions while largely maintaining attribution accuracy and output quality, and annotators verify its citations faster. Prefixes aid cohesion, with their contribution increasing as outputs progress.

5 Results and Analyses

  • ATTR. FIRSTCoT citations were 45 times shorter than ALCE’s, and annotators took roughly half as long to assess them.The human evaluation compared ALCE with the ICL ATTR. FIRSTCoT variant.Source: §5 · pp. 6–7
  • The method surpassed ALCE on attribution accuracy in LFQA, while ALCE scored higher in MDS but had twice as many unattributed sentences.In MDS, around 42% of 109 sentences judged unsupported were partially supported, with other attribution details often near the highlighted spans.Source: §5 · p. 7
  • MDS summary quality changed minimally, while LFQA improved on ROUGEL and BERTSCORE; human evaluation also assessed fluency and LFQA helpfulness.In-context variants tended to produce higher-quality outputs than fine-tuned models.Source: §5 · p. 7
  • Attributions were orders of magnitude shorter, with little impact on attribution accuracy or output quality.The authors note that further improvements to the individual components remain future work.Source: §5 · p. 7

5.1 Do Models Benefit from Seeing the Prefix?

  • About 30% of 113 analyzed sentences were more cohesive with the prefix: 23% improved in referential linkage and 7% in discourse connectors.The analysis examined 20 instances per task, starting from the second sentence.Source: §5.1 · p. 8
  • The prefix’s contribution to cohesion increased as output progressed; Figure 3 reports the percentage of more cohesive sentences by upper sentence-index limit.Source: Figure 3 · p. 7, §5.1 · p. 8

6 Related Work

Prior attribution methods vary in granularity, while post-hoc matching faces a combinatorial challenge. Text planning decomposes generation, and this work contrasts its approach with extractive and Chain-of-Thought alternatives.

6 Related Work

  • LFQA methods range from response-level URLs [Thoppilan et al., 2022] to sentence-level document citations [Gao et al., 2023b] and fine-grained SEMQA spans [Schuster et al., 2023], whose semi-extractive design limits MDS abstraction.Source: §6 · p. 8
  • Post-hoc attribution retrieves documents and matches answer parts to text spans, creating a combinatorial challenge [Bohnet et al., 2023; Gao et al., 2023a].Source: §6 · p. 8
  • Text-planning adds a planning stage before generation [Gehrmann et al., 2018; Narayan et al., 2021, 2023; Zhao et al., 2020]; unlike Ernst et al. (2022), this work supports fluid, multi-proposition sentences.Source: §6 · p. 8
  • Unlike text-planning, Chain-of-Thought typically combines planning and prediction in one generation call; this work explores both approaches.Source: §6 · pp. 8–9

7 Conclusion

Attribute First, then Generate separates generation into multiple steps so attribution emerges as a by-product. Evaluations found task-metric performance matched or exceeded baselines, with much shorter attributions and nearly 50% less fact-checking time.

7 Conclusion

  • The decomposed pipeline matched or outperformed existing baselines on task-specific metrics in automatic and manual evaluations.The authors report fine-grained attributions in both generated text and references.Source: §7 · p. 9
  • Attributions were orders of magnitude shorter, cutting human fact-checking time by almost 50% with little impact on attribution accuracy.Source: §7 · p. 9
  • The scheme separates generation into multiple steps, making attribution a by-product of the generation process.Source: §7 · p. 9
  • The work focused on one implementation of content selection and generation, and the authors encourage exploring other configurations.Source: §7 · p. 9

Limitations

The approach tends to use more computing resources and run more slowly than direct end-to-end generation. Its interlinked components may also propagate errors, though individual-component robustness can partly alleviate this risk.

Limitations

  • Multiple model calls across the approach’s steps tend to increase computing use and slow it relative to direct end-to-end generation.Source: Limitations · p. 9
  • Interlinked components can potentially propagate errors, although individual-component robustness partly alleviates the risk.Later stages are designed to filter non-essential information when content selection returns irrelevant fragments.Source: Limitations · p. 9

Ethics Statement

The human evaluation used recruited and compensated Mechanical Turk annotators. The statement also specifies dataset licensing and the limited use of AI assistants.

Ethics Statement

  • MTurk annotators were invited based on prior high performance, onboarded with guidelines and feedback, and paid about $12 per hour with reimbursements for overruns exceeding 5%.Source: Ethics Statement · p. 9
  • DUC and TAC data were acquired under NIST guidelines and not published; Multi-News and the human evaluation annotations are limited to non-commercial use.Liu et al. (2023) annotation data is published under the MIT license.Source: Ethics Statement · p. 9
  • AI assistants were used exclusively to improve the grammatical structuring of the text.Source: Ethics Statement · p. 9

A Datasets Preprocessing

The section describes the source datasets and filtering used for LFQA, the alignment datasets used for MDS, and dataset-size and evaluation-metric tables.

A Datasets Preprocessing

  • Table A.1 reports dataset sizes for training, development, testing, and hyperparameter tuning.Source: Table A.1 · p. 13
  • Table A.2 reports average Kendall-Tau and Spearman rank correlations and 95% confidence intervals for evaluation metrics against human judgments.Its caption says “prob.” denotes variants averaging the model’s probabilities of label 1.Source: Table A.2 · p. 13
  • LFQA examples with unsupported statements, perceived utility below 4, or evidence excerpts that could not be aligned to source documents were removed.Support was determined by full support from at least one citation or the union of citations; retained utility ratings were 4 or 5.Source: A Datasets Preprocessing · p. 13
  • MDS training uses alignments extracted from DUC and TAC, while development and test use alignments from Multi-News.The Multi-News set contains 100 summaries averaging three source articles each, with annotator-identified aligning spans.Source: A Datasets Preprocessing · p. 13

B Meta-Evaluation for AUTOAIS

The authors compare NLI models and probability-averaging variants against human annotations to select an automatic attribution evaluator. TRUE aligns more closely with human judgments than TrueTeacher, while averaging entailment probabilities correlates worse for both models.

B Meta-Evaluation for AUTOAIS

  • TRUE aligns more closely with human judgment than TrueTeacher, and averaging entailment probabilities correlates significantly worse for both models.The comparison uses Kendall-Tau and Spearman rank correlations against human annotations, with bootstrapping over 1,000 samples of 150 instances.Source: B Meta-Evaluation for AUTOAIS · p. 13

C Implementation Details

This section documents the ICL and fine-tuned implementations of the pipeline, its baselines, and evaluation procedures. It also describes human assessment and supplementary comparisons, including post-generation attribution, latency, and token counts.

C.1 Content Selection

  • ICL content selection extracts verbatim salient spans for MDS or query-relevant spans for LFQA, while fine-tuned selection returns source highlights separated by markup.The fine-tuned model concatenates documents and trims them to fit a 4,096-token context; ICL selection uses task-specific word limits.Source: §C.1 · p. 14

C.2 Generation Planning

  • Generation planning groups selected highlights into clusters intended to co-occur in output sentences, using either generative clustering or a joint source-to-clusters model.For ICL, highlights are enumerated and the LFQA query is included; the fine-tuned generative clustering model takes highlights as input, while the joint variant takes source documents.Source: §C.2 · pp. 14–15

C.3 Attributed Fused Output

  • Attributed fused output generates one sentence from each highlight cluster, iteratively conditioning on the response prefix and attaching citations based on that sentence’s highlights.The ICL chain-of-thought variant combines planning and fusion by iteratively clustering and fusing highlighted source text.Source: §C.3 · p. 15

C.4 End-to-end Generation; C.5 ALCE Variant

  • The end-to-end baseline generates a full summary or answer from the documents, while the ALCE baseline jointly generates text with inline citations to supporting source documents.The ICL end-to-end baseline also includes zero-shot evaluation.Source: §C.4 · p. 15, §C.5 · p. 15

C.6 Constrained Decoding

  • Fine-tuned selection and planning use constrained decoding to copy source n-grams, with minimum highlight lengths and counts and a maximum of two highlights per cluster.The minimum highlight length is three words; minimum counts are 30 for MDS and 5 for LFQA.Source: §C.6 · pp. 15–16

D Full-sentence AUTOAIS

F Human Study; G Annotation Interface

  • Attribution annotations showed fair inter-annotator agreement: Fleiss’ Kappa was 0.37, computed on 101 sentences and their attributions from 23 instances across both models, each annotated by three workers.Partial Support was treated as non-attributed; “Citation Has Support but also Refutes Statement” was labeled attributed.Source: F Human Study · p. 16, G Annotation Interface · pp. 16, 18

E Fine-grained Attribution Examples; H Partially-attributing Examples; J Latency and Total Number of Input and Output Tokens; K Post-generation Attribution

  • Supplementary materials show fine-grained attribution and partial-support examples, prefix-versus-no-prefix predictions, and average token and runtime measurements; post-generation attribution produced longer, less accurate attributions.The post-generation variant adapted RARR to retrieve supporting paragraphs after generating an unattributed output; the tables provide examples and measurements but no values in the cited passages.Source: E Fine-grained Attribution Examples · p. 16, Table D.1 · p. 17, K Post-generation Attribution · p. 18
Loading 2403.17104v3…