PaperlayerView exact revision on arXiv

Source-linked AI summary

WikiContradict: A Benchmark for Evaluating LLMs on Real-World Knowledge Conflicts from Wikipedia

Yufang Hou, Alessandra Pascale, Javier Carnerero-Cano, Tigran Tchrakian, Radu Marinescu, Elizabeth Daly, Inkit Padhi, Prasanna Sattigeri

arXiv:2406.13805v1cs.CLcs.AIcs.LG

TL;DR

Research has not clearly established how LLMs handle conflicts between retrieved passages of equal trustworthiness, particularly when those conflicts come from Wikipedia. This paper introduces and evaluates WikiContradict, finding that models struggle to represent such conflicts, especially implicit ones, and proposing an automated evaluator.

  • Problem

    It remains unclear how LLMs handle conflicts between retrieved passages from a trusted source, while prior synthetic datasets emphasize surface-level contradictions rather than real-world complexity.

    Source: §1 · pp. 1–2

  • Method

    The authors introduce WikiContradict, a 253-instance human-annotated Wikipedia benchmark, and evaluate LLMs across internal-knowledge, single-passage RAG, and two-contradictory-passage QA settings.

    Source: §1 · p. 2

  • Results

    All evaluated models struggled to reflect conflicts between two passages, especially implicit conflicts; WikiContradictEval achieved an F-score of 0.8 on the human-evaluation set.

    Source: §1 · p. 2

  • Takeaways & Limitations

    WikiContradict offers a resource for examining and tracking LLM progress in handling real-world inter-context conflicts.

    Source: §1 · p. 2

    The benchmark is limited to English, Wikipedia-tagged contradictions, and text-text conflicts, so its coverage does not extend to other languages or modalities.

    Source: §6 · pp. 8–9

Abstract

from arXiv · show

Retrieval-augmented generation (RAG) has emerged as a promising solution to mitigate the limitations of large language models (LLMs), such as hallucinations and outdated information. However, it remains unclear how LLMs handle knowledge conflicts arising from different augmented retrieved passages, especially when these passages originate from the same source and have equal trustworthiness. In this work, we conduct a comprehensive evaluation of LLM-generated answers to questions that have varying answers based on contradictory passages from Wikipedia, a dataset widely regarded as a high-quality pre-training resource for most LLMs. Specifically, we introduce WikiContradict, a benchmark consisting of 253 high-quality, human-annotated instances designed to assess LLM performance when augmented with retrieved passages containing real-world knowledge conflicts. We benchmark a diverse range of both closed and open-source LLMs under different QA scenarios, including RAG with a single passage, and RAG with 2 contradictory passages. Through rigorous human evaluations on a subset of WikiContradict instances involving 5 LLMs and over 3,500 judgements, we shed light on the behaviour and limitations of these models. For instance, when provided with two passages containing contradictory facts, all models struggle to generate answers that accurately reflect the conflicting nature of the context, especially for implicit conflicts requiring reasoning. Since human evaluation is costly, we also introduce an automated model that estimates LLM performance using a strong open-source language model, achieving an F-score of 0.8. Using this automated metric, we evaluate more than 1,500 answers from seven LLMs across all WikiContradict instances. To facilitate future work, we release WikiContradict on: https://ibm.biz/wikicontradict.

1 Introduction

The introduction identifies a gap in evaluating real-world inter-context conflicts: prior work often uses synthetic, surface-level contradictions. WikiContradict addresses this with Wikipedia-based instances and evaluates how LLMs respond to conflicting passages.

1 Introduction

  • Prior synthetic conflict datasets tend to emphasize explicit, surface-level contradictions and miss the complexity of real-world conflicts.Examples include entity substitution and model-generated counterfactual evidence.Source: §1 · p. 2
  • WikiContradict contains 253 human-annotated instances of real-world conflicts from equally credible passages originating from Wikipedia.The instances cover contradiction types identified by Wikipedia editors; 36% require implicit reasoning.Source: §1 · p. 2
  • The evaluation covers internal-knowledge QA, RAG with one passage, and RAG with two contradictory passages, with human assessment of model responses.The final human-evaluation set contains 1,200 samples from five LLMs, 48 instances, and five prompt templates.Source: §1 · p. 2
  • All evaluated models, including GPT-4, struggled to reflect conflicts between two passages, especially when the contradiction required implicit reasoning.Prompting models to attend to contradictions improved performance, largely on explicit conflicts; Llama-3-70b-instruct rose from 10.4% to 43.8%.Source: §1 · p. 2
  • WikiContradictEval uses few-shot prompting to have Llama-3-70b-instruct judge responses and achieves an F-score of 0.8 on the human-evaluation set.Source: §1 · p. 2
  • The authors present WikiContradict as a resource for examining and tracking LLM progress on real-world inter-context conflicts.Source: §1 · pp. 2–3

2 Related work

Related work distinguishes inter-context conflicts from other knowledge conflicts and notes that synthetic datasets do not capture the complexity of real-world conflicts. The paper focuses on questions with multiple equally trustworthy answers rather than binary factuality judgments.

2 Related work

  • The paper focuses on inter-context conflicts, where retrieved information can disagree, and aims to study such conflicts in real-world settings rather than synthetic ones.Prior work on knowledge conflicts has also examined intra-memory and context-memory conflicts.Source: §2 · p. 3
  • Prior synthetic conflict datasets are described as failing to represent the complexity of real-world conflicts.The authors aim to study inter-context conflicts using contradictions extracted and annotated from Wikipedia.Source: §2 · p. 3
  • Existing factuality benchmarks commonly evaluate short- or long-form answers as true-or-false facts, while this work examines questions with multiple equally trustworthy answers.Source: §2 · p. 3

3 WikiContradict

WikiContradict is built from Wikipedia maintenance tags and annotated into a QA benchmark of real-world conflicts. Its instances capture varied conflict types, and its evaluation protocol tests LLM responses across internal-knowledge and retrieval settings.

3 WikiContradict

  • WikiContradict is a QA benchmark of 253 human-annotated instances covering real-world knowledge conflicts.Source: §3 · p. 3

3.1 Data collection and processing

  • The data collection focuses on Wikipedia’s inconsistent, self-contradictory, and contradict-other tags, drawing on tagged articles and editors’ comments.The first two tags mark contradictions within an article; contradict-other flags a conflict between articles. Around 1,200 articles were collected.Source: §3.1 · pp. 3–4

3.2 Data annotation

  • Annotators verify tags, extract and edit conflicting passages, explain and categorize each contradiction, then write at least one question eliciting different answers.Edits can resolve anaphora; implicit conflicts require reasoning and take longer to annotate.Source: §3.2 · p. 4

3.3 Data statistics

  • The benchmark contains 253 instances, with approximately 61% wh-questions and 36% implicit contradictions.Each instance includes a question, two contradictory passages, and two answers, one derived from each passage.Source: §3.3 · p. 5, Table 1 · p. 5

3.4 Evaluation

  • Five prompt templates test internal knowledge, RAG with one passage, and RAG with two contradictory passages.Templates 4 and 5 assess responses to two passages that can yield different answers.Source: §3.4 · p. 5
  • For two-passage prompts, a correct answer must represent both contradictory answers without preferring one or merging them as if simultaneously true.Responses are categorized as correct, partially correct, or incorrect; the table breaks results down by model, prompt, and explicit or implicit conflict type.Source: §3.4 · pp. 5–6, Table 2 · p. 6

4 Human evaluation: LLMs are struggling on WikiContradict

Human evaluations show that LLMs often mishandle conflicting passages, though explicit conflict prompts improve answers, especially for explicit rather than implicit conflicts. The evaluation also finds weaknesses in contradiction detection and in following context when it conflicts with a model’s internal knowledge.

4 Human evaluation: LLMs are struggling on WikiContradict

  • The main human-evaluation dataset contained 1,200 samples from 48 instances after disagreement resolution, drawn from an initial 1,375 responses across 55 instances.Seven ambiguous instances were excluded. The results break out 30 explicit and 18 implicit conflicts; supplementary prompt variants tested position swaps, contradiction detection, and consistent contexts, with one annotator per response.Source: §4 · pp. 6–7
  • When context contained no conflicting information, models generally answered correctly in most instances, but GPT-4 often relied on internal knowledge and achieved 87.8% accuracy, below Mixtral-8X7b-inst and Llama3-3-70b-inst.Source: §4 · p. 7
  • With two conflicting passages, models often rely on one; explicit conflict prompts raised Llama-3-70b-inst correct answers from 10.4% to 43.8%.The study notes that GPT-4 improved from 6.3% to 10.4%, the smallest reported improvement.Source: §4 · p. 7
  • For Llama-3-70b-inst, prompting lifted accuracy from 13.3% to 60.0% on explicit conflicts but only 5.9% to 17.6% on implicit conflicts.The authors report that improvements are particularly pronounced when conflicts are explicitly stated.Source: §4 · p. 7
  • Contradiction detection exceeded correct-answer generation; GPT-4 and Llama-3-70b-instruct detected conflicts in about 88% and 77%, but confirmed only 44% and 48% of consistent cases.On consistent contexts, correct-answer rates were 85%–87%; the two models sometimes invented reasons the passages conflicted.Source: §4 · p. 7

5 Automatic evaluation: WikiContradictEval

WikiContradictEval uses few-shot LLM judging to assess answers to contradictory-context questions, enabling evaluation across all 253 benchmark instances. The judge estimates Mistral-7b-inst as the top performer, but may overestimate its correctness; explicit conflicts are generally answered more accurately than implicit ones.

5 Automatic evaluation: WikiContradictEval

  • 82.5 F-score, 73.4 precision, and 94.0 recall were achieved by GPT-4 in identifying correct responses; Llama-3-70b-inst had the best open-source F-score at 80.4.The judge models were evaluated on 240 prompt-template-5 responses from five testing LLMs.Source: §5 · p. 8
  • The Llama-3-70b-inst judge generally aligned with human judgments, but overestimated Mistral-7b-instruct by accepting answers that supplied both conflicting answers.Mistral-7b-instruct often attempted to reconcile the conflict by giving both answers.Source: §5 · p. 8

6 Limitation and future work

  • With Llama-3-70b judging 253 WikiContradict instances under prompt template 5, Mistral-7b-inst and Llama-3-70b-inst led at 50.6% and 45.8%; all models except Flan-ul2 did better on explicit than implicit contradictions.The authors suggest the judge may overestimate Mistral-7b-inst’s performance, indicating potential evaluation bias.Source: §5 · p. 8

6 Limitation and future work

The benchmark is limited to English instances and text-text contradictions, and its reliance on Wikipedia tags may skew which contradiction types it covers.

6 Limitation and future work

  • The benchmark contains English-language instances only, so its results may not generalize to other languages.The paper identifies the nuance of contradicting statements as a reason for this boundary.Source: §6 · p. 8
  • Reliance on Wikipedia tags may skew the benchmark toward specific types of contradictory statements.Source: §6 · p. 8
  • The benchmark covers text-text contradictions, while future work aims to include cross-modal cases such as text that differs from an image.Source: §6 · pp. 8–9

7 Conclusion

WikiContradict provides a benchmark for studying real-world conflicts between Wikipedia passages, and its evaluation indicates that LLMs often struggle to manage and reason over those conflicts. The appendix also documents the annotation workflow used to construct the benchmark.

7 Conclusion

  • LLMs often struggle to correctly identify and manage real-world inter-context conflicts in WikiContradict.The benchmark contains 253 human-annotated instances of contradictions identified by Wikipedia editors.Source: §7 · p. 9

A Annotation Guidelines

  • Annotations capture conflicting passages, resolve anaphora, and create questions whose answers vary with the passage used.Annotators also classify the contradictions using a predefined taxonomy.Source: §7 · p. 9, §1 · pp. 15–16

1. Setup the Label Studio environment

  • The setup instructions install Label Studio locally, create a project, and import annotation tasks for the annotators.The project may be named “WikipediaContradict” or another name; the instructions specify importing 170 tasks.Source: §1 · p. 13
  • The annotation interface records original passages separately from optional stand-alone versions and asks annotators to mark whether the passages are the same.The interface also provides fields for each passage’s contradiction reason and, when possible, its contradicted span.Source: §1 · p. 15

2. Task annotation

The task-annotation instructions guide annotators from checking Wikipedia inconsistency tags to preparing passages, labeling contradiction attributes, and writing questions with passage-specific answers.

2. Task annotation

  • Annotators verify each inconsistency tag against the article and, when needed, consult its revision history to assess whether it is valid.The instructions describe checking editor comments and investigating the edit history when the reason is not specified in the tag.Source: §2 · pp. 17–18
  • Annotators copy the conflicting information into separate passage fields and make stand-alone versions by resolving anaphora or expressing table content in text.Original passage text is retained without citation marks or inconsistency tags; decontextualization is applied to separate stand-alone fields.Source: §2 · pp. 18–20
  • The contradiction reason is recorded as what Passage 1 states versus what Passage 2 states, with exact contradicted spans added when identifiable.A span may be omitted when the contradiction spans multiple sentences.Source: §2 · p. 21
  • Type I classifies contradiction semantics at entity, noun-phrase, event/relation, and discourse levels.The examples distinguish phrase-level and discourse-level contradictions, including entity, non-entity noun phrase, and event/relation cases.Source: §2 · pp. 22–24
  • Types II and III label whether information comes from text or an infobox/table and whether the passages come from the same article.The instructions say most inconsistencies are within one article, while some occur across articles.Source: §2 · p. 25
  • Type IV distinguishes explicit contradictions from implicit ones that require reasoning to recognize.The example contrasts “monotypic” with “species to species,” requiring commonsense reasoning about subspecies.Source: §2 · p. 26
  • Annotators may record attributes outside the taxonomy in an additional comment, such as contradictions across language versions.They then write at least one question that elicits different answers from the two passages and record annotation confidence.Source: §2 · pp. 26–27

3. Export annotations

  • Completed annotation tasks are exported from the task pool in JSON format and renamed using the specified filename template.The example filename is wikipediaContradict_10_<your name>.json.Source: §3 · p. 28

B Details on the Prompt Templates 5.1 - 5.4 and Human Evaluation Results

The appendix supplies prompt examples and expands the human-evaluation rubric, including how it handles multiple conflicting gold answers and responses that do not express their conflict coherently.

B Details on the Prompt Templates 5.1 - 5.4 and Human Evaluation Results

  • Prompt 5.1 asks for a concise answer reflecting the context’s comprehensive, potentially contradictory information, while Prompt 5.2 asks whether context conflicts could yield different answers.Prompt 5.1 presents Passage 2 before Passage 1; Prompt 5.2 requests a short answer and explanation.Source: B Details on the Prompt Templates 5.1… · p. 30
  • Prompt 5.4 supplies two modified passages as context for the question.The appendix excerpt provides only the context portion of the template.Source: B Details on the Prompt Templates 5.1… · p. 30

C Further Details on the Judge LLM Prompt

  • The judge marks an answer correct when it matches all listed answers and handles their contradictory nature without preferring one answer.Combining incompatible answers as simultaneously true is not sufficient for correctness.Source: C Further Details on the Judge LLM… · p. 33
  • A response is partially correct if it gives only one gold answer, prefers one among multiple correct answers, or adds an answer absent from the gold list.A response that matches none of the listed answers is incorrect.Source: C Further Details on the Judge LLM… · p. 33
  • Numerical answers generally must be exact unless the ground truth explicitly includes an approximation, and entity answers are expected to use complete or commonly recognized names.The judge prompt also permits irrelevant hallucinated or outdated information when it does not significantly affect the primary answer.Source: C Further Details on the Judge LLM… · p. 33
  • The examples distinguish a single matching age answer from an answer that lists both conflicting ages as simultaneously true.The former is partially correct; the latter is incorrect because one person cannot have two ages at the same time.Source: C Further Details on the Judge LLM… · p. 33
  • The judge accepts answers that attribute the two ages to different sources or present them as alternatives without asserting a preference.Both example treatments are labeled correct.Source: C Further Details on the Judge LLM… · p. 34

D Additional Experimental Settings

The appendix specifies model and decoding settings, describes the benchmark’s JSON fields, and notes that one annotation component is omitted from the dataset.

D Additional Experimental Settings

  • The appendix lists five models for human evaluation and six judge models for automatic evaluation, with token and decoding settings for the tested models.The listed open-source models use greedy search, while the GPT models use the Chat Completions API with temperature 0 and a 250-token maximum.Source: D Additional Experimental Settings · p. 36

E Data Format

  • WikiContradict is stored in JSON with 253 instances and fields for article and passage information, validity, contradiction taxonomy, questions, and passage-specific answers.The format includes metadata such as the tag date and reason, plus annotation results and decontextualized passage text.Source: E Data Format · pp. 36–37

F Examples of WikiContradict

G Datasheets for WikiContradict

The appendix provides dataset documentation and intended uses following the Datasheets for Datasets framework.

G.1 Dataset Documentation and Intended Uses

This appendix documents WikiContradict’s purpose, composition, collection, processing, uses, distribution, and maintenance. The benchmark contains 253 human-annotated instances built from Wikipedia’s tagged content inconsistencies.

G.1 Dataset Documentation and Intended Uses

  • WikiContradict was created to support research assessing LLMs on retrieved passages containing real-world knowledge conflicts.Source: §G.1 · p. 38

G.2 Composition

  • The dataset contains 253 instances, each built around questions and conflicting passages with corresponding answers; some records may lack selected fields.The documentation says the dataset contains all possible instances and notes that manual annotation inevitably introduces noise.Source: §G.2 · pp. 38–39

G.3 Collection process

  • Annotators checked Wikipedia’s inconsistency tags against article content and, when needed, editor comments, edit histories, and talk pages.The tagged articles came from Wikipedia’s content-issues category; collection covered around 1,200 articles.Source: §G.3 · p. 39

G.4 Preprocessing/cleaning/labeling

  • Annotators made passages stand-alone by resolving anaphoric references, and the dataset retains the raw passages.Python was used for data cleaning, with scripts to be shared in the codebase.Source: §G.4 · p. 40

G.5 Uses

  • The dataset was used to assess LLMs on real-world conflicts and may support prompt augmentation or fine-tuning for conflicting sources.The documentation states that the public Wikipedia data presents minimal risk for harm.Source: §G.5 · pp. 40–41

G.6 Distribution

  • The dataset and metadata were planned for public distribution, with an MIT license specified for WikiContradict.The documentation lists a website for distribution and states the dataset and DOI would be published after acceptance.Source: §G.6 · p. 41

G.7 Maintenance

  • Yufang Hou, Alessandra Pascale, Javier Carnerero-Cano, and Tigran Tchrakian are named as dataset maintainers.Source: §G.7 · p. 42

How can the owner/curator/manager of the dataset be contacted (e.g., email address)?

The appendix describes dataset updates and maintenance, responsibility for rights violations, the license, and potential biases in the data and annotation.

How can the owner/curator/manager of the dataset be contacted (e.g., email address)?

  • Dataset updates or extensions were to be released on the website and GitHub, and all versions were to remain supported.The documentation invites those wishing to build on the dataset to contact the authors.Source: How can the owner/curator/manager of… · p. 42

G.8 Responsibility

  • The authors state that they bear responsibility in case of rights violations.Source: §G.8 · p. 42

H Explicit License

  • WikiContradict is distributed under an MIT license, which requires including the copyright and permission notices in copies or substantial portions.Source: H Explicit License · p. 42

I Ethics Statement

  • Wikipedia sourcing biases the data toward its content, and annotator subjectivity may affect which techniques are identified.The authors prepared 17-page annotation guidelines and instructed annotators not to rely on personal feelings about topics.Source: I Ethics Statement · p. 43
Loading 2406.13805v1…