PaperlayerView exact revision on arXiv

Source-linked AI summary

Attributed Question Answering: Evaluation and Modeling for Attributed Large Language Models

Bernd Bohnet, Vinh Q. Tran, Pat Verga, Roee Aharoni, Daniel Andor, Livio Baldini Soares, Massimiliano Ciaramita, Jacob Eisenstein, Kuzman Ganchev, Jonathan Herzig, Kai Hui, Tom Kwiatkowski, Ji Ma, Jianmo Ni, Lierni Sestorain Saralegui, Tal Schuster, William W. Cohen, Michael Collins, Dipanjan Das, Donald Metzler, Slav Petrov, Kellie Webster

arXiv:2212.08037v2cs.CL

TL;DR

The paper addresses how to evaluate attribution and how current systems perform, given the potential of LLMs for information-seeking. It defines Attributed QA, benchmarks architectures using human ratings and an automatic metric, and finds retrieve-then-read systems strongest while noting an AutoAIS caveat for reranked variants.

  • Problem

    Although LLMs show promise for information-seeking, the field has limited understanding of how to attribute their generated text and evaluate attribution.

    Source: §1 · p. 1, §3.3 · p. 4

  • Method

    The paper defines Attributed QA as answering questions with supporting evidence and evaluates a broad set of architectures using human ratings and AutoAIS.

    Source: §3 · p. 2, §2.2 · p. 3

  • Results

    Retrieve-then-read approaches achieve the strongest performance, though they require full use of a traditional training set; AutoAIS correlates well with human judgment at the system level.

    Source: §5.3 · p. 7, §7 · p. 10

  • Takeaways & Limitations

    AutoAIS offers promise as a development metric when human ratings are infeasible, while post-hoc attribution of LLM-generated answers remains challenging.

    Source: §5.4 · p. 8, §7 · p. 10

    Reranked variants are outliers to the AutoAIS-human correlation, consistent with instance-level AutoAIS being noisier than system-level AutoAIS.

    Source: §5.5 · p. 10

Abstract

from arXiv · show

Large language models (LLMs) have shown impressive results while requiring little or no direct supervision. Further, there is mounting evidence that LLMs may have potential in information-seeking scenarios. We believe the ability of an LLM to attribute the text that it generates is likely to be crucial in this setting. We formulate and study Attributed QA as a key first step in the development of attributed LLMs. We propose a reproducible evaluation framework for the task and benchmark a broad set of architectures. We take human annotations as a gold standard and show that a correlated automatic metric is suitable for development. Our experimental work gives concrete answers to two key questions (How to measure attribution?, and How well do current state-of-the-art methods perform on attribution?), and give some hints as to how to address a third (How to build LLMs with attribution?).

1 Introduction

LLMs show promise for information-seeking tasks, but the field has limited understanding of how to attribute their outputs and evaluate attribution. The paper identifies attribution measurement and current system performance as open questions.

1 Introduction

  • LLMs have shown promise in information-seeking tasks, from simple and long-form question answering to information-seeking dialogue, despite the difficulty of constructing labeled datasets.That difficulty extends from simple QA to multi-faceted QA and interactive dialogue.Source: §1 · pp. 1–2
  • Attribution is likely crucial in information-seeking scenarios, yet understanding of the challenge and how to make progress remains limited.The paper asks how to measure attribution and how well current state-of-the-art methods perform.Source: §1 · p. 2

1. How to measure attribution?

  • The paper identifies attribution measurement as an open question even for simple question answering.Source: §1 · p. 2

3. How to build LLMs with attribution?

The paper studies Attributed QA as a first step toward attributed LLMs, where systems answer questions and point to supporting evidence. It proposes a human-grounded evaluation framework and benchmarks varied systems to inform future design.

3. How to build LLMs with attribution?

  • Attributed QA takes a question as input and returns an answer with a pointer into a fixed corpus to evidence supporting that answer.Source: §3 · p. 2
  • Attributed QA is a comparatively simple information-seeking task that may inform work on building attributed LLMs, while also being useful to study in its own right.The paper notes potential benefits of system-provided attribution for developers and users.Source: §3 · p. 2
  • The paper proposes a reproducible evaluation framework using human annotations as the gold standard and finds AutoAIS correlates strongly with human judgment for development.Source: §3 · p. 2
  • The paper systematically analyzes systems built from state-of-the-art components across architectures and supervision levels, and releases scored outputs for further exploration.Source: §3 · p. 2
  • The work gives concrete answers about measuring attribution and current system performance, while offering hints for building attributed LLMs.Source: §3 · p. 2

2 Related Work

Prior QA work spans reading comprehension, open-domain retrieval pipelines, and closed-book models, while attribution work includes retrieval-based systems and information-seeking demonstrations. The paper situates its study against these approaches and identifies limits in prior attribution evaluation.

2.1 Question Answering Tasks

  • Open-domain QA systems answer queries from a fixed evidence corpus, with retrieve-then-read pipelines retrieving passages before reading them to infer an answer.Dense methods jointly optimize passage retrieval and answer prediction and have typically used less training signal than pipeline approaches.Source: §2.1 · p. 3
  • Closed-book QA showed that T5 can answer without a corpus at inference, relying on information retained in model parameters, but raises evaluation questions.Source: §2.1 · p. 3
  • Natural Questions provides question, short-answer, and supporting-paragraph examples, but only one long answer is annotated per question; KILT adds supporting paragraphs, averaging 1.57 per question-answer pair.Source: §2.1 · p. 3

2.2 LLMs with Attribution

  • Prior attribution work includes post-editing LLM text against retrieved web content and GopherCite, which combines search evidence with human evaluation.Source: §2.2 · p. 3
  • GopherCite evaluates only one hybrid system, and 192 of 307 questions were discarded after raters skipped items, making its reported 80% accuracy hard to interpret.Source: §2.2 · p. 3
  • Recent information-seeking dialogue demonstrations condition LLM outputs on retrieved content, motivating a systematic study of architecture choices using AIS principles.Source: §2.2 · p. 3

3 Attributed Question Answering

Attributed QA models a question-answer pair with a supporting source, and evaluates whether that answer is interpretable and fully supported. The discussion explains why attribution and human ratings matter, while noting evaluation challenges and complexities in extending the task to broader LLM outputs.

3.1 Task Definition

  • Attributed QA takes a question as input and returns an answer string paired with a member of a fixed set of evidence units, such as corpus paragraphs.Source: §3.1 · p. 4

3.2 Evaluation

  • Human evaluation marks an answer attributable only when its information is interpretable and fully supported by the cited source.Source: §3.2 · p. 4
  • Human test accuracy is the proportion of examples for which a majority of raters judge the system output attributable.Source: §3.2 · p. 4

3.3 Discussion

  • The set of valid answer-and-source labels can be large because multiple passages may support an answer and questions can have diverse valid answers.Answer diversity can reflect differing strings, opinions, or interpretations, making QA evaluation challenging.Source: §3.3 · p. 4
  • Attribution lets users and developers inspect sources for trustworthiness and nuance without requiring difficult factuality judgments for complex questions.Source: §3.3 · p. 4
  • Closed-book QA omits attribution and relies on curated gold answers, which can be difficult for questions with diverse answers and may undercount performance.The paper says the scale of possible undercounting cannot be estimated without large-scale human evaluations.Source: §3.3 · p. 5
  • The large set of human labels gathered in the paper may support development of automatic evaluations correlated with human ratings or trained on reference answer-attribution pairs.Source: §3.3 · p. 5
  • Extending Attributed QA to broader LLM outputs raises questions about which statements need attribution, moving beyond question-answer pairs, and multi-source support.The authors hypothesize that progress on Attributed QA will extend naturally to more complex tasks.Source: §3.3 · p. 5

4 Approaches to Attributed QA

The section organizes systems into retrieve-then-read, post-hoc retrieval, and LLM-as-retriever architectures, then distinguishes them by supervision and specifies representative best systems and variants.

4.1 Architectures

  • RTR retrieves k passages from the question, then generates an answer from selected passages and attributes it to one; experiments pair BM25, GTR, and FID.BM25 is used for sparse retrieval, GTR for dense retrieval, and FID for answer generation.Source: §4.1 · pp. 5–6
  • Post-hoc systems generate an answer without retrieval, then retrieve passages using the question-answer pair and select a passage containing the generated answer.For k > 1, the highest-scoring passage containing the answer is selected as attribution.Source: §4.1 · p. 6
  • LLM-as-retriever systems generate an answer and pointer without sparse or dense retrieval; this paper generates a Wikipedia URL, then selects a supporting paragraph.Generating a paragraph pointer directly is noted as a possible extension, not investigated here.Source: §4.1 · p. 6

4.2 Supervision

  • NQ-64 systems use 64 randomly chosen Natural Questions examples, while NQ-full systems assume the full training set; RTR uses full-set supervision for GTR and FID.Post-hoc systems using NQ-full select 64 similar examples by BM25 for prompting.Source: §4.2 · p. 6

4.3 Best Systems

  • The best RTR retrieves 50 passages with GTR, reranks them with NQ-reranker, trains FID on 50 passages, but answers from and attributes to the top one.This approach is described as close to Izacard et al. (2022), with a final passage-selection step for attribution evaluation.Source: §4.3 · p. 6
  • The best post-hoc system pairs 540B PaLM with 64 BM25-selected NQ examples and GTR retrieval; the low-resource version uses BM25 and 64 examples without fine-tuning.Both form retrieval queries by concatenating the question and generated answer; the GTR variant selects a passage containing the answer string.Source: §4.3 · p. 6
  • The best LLM-as-retriever fine-tunes 540B PaLM to generate answers and Wikipedia URLs, then attributes answers to the highest-BM25-scoring Wikipedia paragraph.Fine-tuning data includes generated questions over decontextualized Wikipedia sentences and Natural Questions training examples.Source: §4.3 · p. 6
  • AutoAIS-reranked RTR and post-hoc variants select the highest-AutoAIS-scoring passage among the top 50; only AIS is evaluated because AutoAIS is a system component.The authors encourage distinguishing automatic metrics used as system components from their use for evaluation.Source: §4.3 · pp. 6–7

5 Experiments

The experiments evaluate Attributed QA on Natural Questions using human AIS ratings, AutoAIS, and Exact Match, then compare system architectures and analyze agreement between metrics. RTR systems lead overall, while metric comparisons show that AutoAIS tracks human AIS more closely than EM at the system level.

5.1 Datasets

  • The evaluation uses short-answer questions from the Natural Questions validation set and an attribution corpus derived from a 2021-10-13 Wikipedia snapshot.Pyserini was used to extract paragraphs from Wikipedia pages.Source: §5.1 · p. 7

5.2 Evaluation Metrics

  • Human AIS is the gold standard, based on majority votes from five raters over 1,000 randomly chosen questions; AutoAIS scores entailment, while EM ignores attribution.AutoAIS uses an 11B-parameter T5 checkpoint, and scores of at least 0.5 indicate valid answer-attribution predictions.Source: §5.2 · p. 7

5.3 System Results

  • Best RTR significantly outperforms the best non-RTR system (p ≪10−5, t = 4.55), despite using T5 XL with 3B parameters versus PaLM with 540B.RTR approaches require substantial explicit supervision and are likely to depend strongly on retrieval accuracy.Source: Table 1 · p. 7, §5.3 · p. 7
  • Best Post-hoc achieves relatively high EM with prompted answer generation, but post-hoc systems require large LLMs and lag RTR on attribution.Best Low Resource performs competitively with Best Post-hoc on AIS and AutoAIS; LLM-as-retriever is competitive with low-resource post-hoc attribution but also requires large models.Source: §5.3 · pp. 7–8

5.4 Ablations

  • Dense retrieval leads sparse retrieval in both architectures; RTR benefits from T = 50 training and matching P = A = 50, while NQ-full adds 10 Exact Match points but little AIS.Selecting an attribution passage among 50 retrieved passages tends to outperform using only the top passage in post-hoc systems.Source: §5.4 · p. 8

5.5 Correlation between AIS and EM/AutoAIS

  • AIS and EM have a modest system-level correlation of 0.45; disagreements include inexact or stale reference answers and multiple valid answers.The authors point to limitations of reference-answer corpora and string-matching evaluation.Source: §5.5 · p. 8, Figure 2 · p. 9
  • AIS and AutoAIS have a strong system-level correlation of 0.96, but individual AutoAIS scores are noisier and reranked variants fall below the linear fit.AutoAIS is suitable as an aggregate development metric, not as a system component; individual scores should not be read too closely.Source: §5.5 · pp. 8, 10, Figure 3 · p. 9

6 Future Directions

The authors outline modeling, evaluation, and task directions for extending their study of Attributed QA. They identify retrieval and training challenges, ways to improve automatic evaluation, and questions about broader task coverage.

Modeling

  • Retrieve-then-read systems perform strongly but often need substantial training data and resources; future modeling work should examine post-hoc retrieval challenges and training signals for end-to-end systems.AutoAIS could provide a noisy training signal; instruction tuning and chain-of-thought prompting are also proposed for improving answer quality.Source: §6 · p. 10

Evaluation

  • Human ratings could support either a cache of previously scored predictions or a learned attribution metric, with the latter also usable alongside AutoAIS as a noisy training signal.The proposed cache would extend KILT (Petroni et al., 2021); learning an improved metric draws on prior work (Sellam et al., 2020; Rei et al., 2020).Source: §6 · p. 10

Tasks

  • The study centers on Natural Questions, and its best system uses the full training set, leaving transfer to other datasets and settings for future evaluation.Suggested extensions include multilingual or multimodal attribution and long-form QA.Source: §6 · p. 10

7 Conclusion

The paper defines Attributed QA and evaluates a broad range of system designs, using human ratings as the gold standard. Retrieve-then-read performs best but requires a traditional training set, while post-hoc attribution remains challenging.

Conclusion

  • Retrieve-then-read systems achieve the strongest performance, but require full use of a traditional training set; post-hoc attribution remains challenging.The paper presents post-hoc attribution as a viable architecture for future work.Source: §7 · p. 10
  • Attributed QA pairs an answer with an attribution pointer into a fixed corpus, and the paper benchmarks system architectures using AIS principles.The conclusion frames this task as a step toward developing attributed large language models.Source: §7 · p. 10
  • Human ratings are the evaluation gold standard, while AutoAIS correlates well with human judgment at system level and may serve as a development metric or noisy training signal.The paper also notes that AutoAIS is promising when human rating is infeasible.Source: §7 · p. 10

8 Ethical Considerations

The paper identifies factuality assessment as difficult beyond simple claims and presents attribution as a way for users to inspect sources and judge trustworthiness and answer scope. Its study is English-only and mostly uses resource-intensive approaches.

Ethical Considerations

  • Judging factuality is especially difficult for claims beyond the simplest ones, and source attribution lets users assess trustworthiness and answer scope.The authors identify factuality as the main ethical consideration of the work.Source: §8 · p. 10
  • Attributed QA is explored only in English and mostly with resource-intensive approaches that may not be accessible to many.The authors note that AIS principles, system outputs, and ratings are publicly available to encourage further work.Source: §8 · pp. 10–11
  • The authors have released the AIS principles, system outputs, and ratings to support future work on lower-resource and end-to-end attribution solutions.Source: §8 · p. 11

9 Contributions

The contribution statement credits technical work, infrastructure, evaluation components, corpus development, and writing across the paper’s contributors.

Contributions

  • Bernd Bohnet, Vinh Q. Tran, and Pat Verga led technical work, including model implementation, experiments, analysis, and improvements; Kellie Webster acted as TL.Source: §9 · p. 11
  • Livio Baldini Soares built BM25 retrieval infrastructure, Ji Ma and Jianmo Ni contributed dense retrieval and reranking, and Kai Hui helped with PaLM usage.Daniel Andor and Kuzman Ganchev ran components enabling the LLM-as-retriever model.Source: §9 · p. 11
  • Bernd Bohnet managed human-rating collection; Roee Aharoni and Jonathan Herzig trained NLI evaluation models, and the team also contributed to AutoAIS, the attribution corpus, and evaluation scripts.The passage credits Roee Aharoni with open-sourcing the model on Hugging Face, Massimiliano Ciaramita with conceiving the AutoAIS prompt, and Lierni Sestorain Saralegui with exploring variants.Source: §9 · p. 11
  • Michael Collins and Kellie Webster were primary writers; William W. Cohen helped with related work, Jacob Eisenstein contributed statistical analysis, and several researchers developed the work’s direction and provided feedback.Direction and feedback are credited to William W. Cohen, Michael Collins, Dipanjan Das, Don Metzler, Slav Petrov, and Kellie Webster.Source: §9 · p. 11

A Examples

The section presents three classes of system predictions that AIS regards as valid but Exact Match scores as incorrect.

A Examples

Loading 2212.08037v2…