Source-linked AI summary
$Q^{2}$: Evaluating Factual Consistency in Knowledge-Grounded Dialogues via Question Generation and Question Answering
Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, Omri Abend
TL;DR
Knowledge-grounded dialogue models can generate fluent responses that conflict with their grounding knowledge, while existing factuality metrics are not well suited to open-ended dialogue. The paper introduces Q2, a reference-free metric that uses question generation, question answering, and NLI to compare response and knowledge answers, and finds higher correlations with human judgments across three datasets. Q2 remains limited by errors involving chit-chat and unresolved pronouns.
Problem
Knowledge-grounded dialogue systems often produce factually inconsistent or hallucinated content, while existing metrics face dialogue-specific content and reference-dependence limitations.
Method
Q2 generates questions from informative response spans, answers them using grounding knowledge, and compares the answer spans with NLI without requiring reference responses.
Results
Q2 achieves significantly higher correlations with human judgments than other reference-free metrics across Wizard of Wikipedia, Topical-Chat, and Dialogue NLI.
Takeaways & Limitations
Q2 provides an evaluation framework for factual consistency in knowledge-grounded dialogue and is reported to remain effective across multiple dialogue benchmarks.
Takeaways & Limitations
Q2 can generate errors for chit-chat and responses containing pronouns that require resolving entities from dialogue history.
Abstract
from arXiv · showhide
Neural knowledge-grounded generative models for dialogue often produce content that is factually inconsistent with the knowledge they rely on, making them unreliable and limiting their applicability. Inspired by recent work on evaluating factual consistency in abstractive summarization, we propose an automatic evaluation metric for factual consistency in knowledge-grounded dialogue using automatic question generation and question answering. Our metric, denoted $Q^2$, compares answer spans using natural language inference (NLI), instead of token-based matching as done in previous work. To foster proper evaluation, we curate a novel dataset of dialogue system outputs for the Wizard-of-Wikipedia dataset, manually annotated for factual consistency. We perform a thorough meta-evaluation of $Q^2$ against other metrics using this dataset and two others, where it consistently shows higher correlation with human judgements.
1 Introduction
Knowledge-grounded dialogue systems can generate fluent but factually inconsistent or hallucinated responses, while existing factuality metrics are poorly suited to open-ended dialogue. Q2 addresses this gap with reference-free question generation, question answering, and NLI, and correlates better with human judgments across three benchmarks.
- Knowledge-grounded dialogue models often produce fluent outputs that conflict with or hallucinate beyond their grounding knowledge.Such inconsistencies can mislead users and limit system applicability.
- Existing factual-consistency metrics are less suitable for dialogue because responses mix factual content with history references, opinions, questions, and chit-chat.Many also require gold human references, whereas dialogue is open-ended.
- Q2 evaluates factual consistency without gold reference responses by pairing automatic question generation and question answering.It generates questions from informative response spans, answers them using grounding knowledge, and compares the resulting answer spans.
- Q2 compares response and knowledge answer spans with natural language inference rather than token-based matching, improving robustness to lexical variation.This modular design is intended to handle the mixture of utterance types characteristic of dialogue.
- Q2 obtains significantly higher correlations with human judgments than other reference-free metrics across Wizard of Wikipedia, Topical-Chat, and Dialogue NLI.The authors evaluate it using a manually annotated Wizard-of-Wikipedia response dataset and the two additional benchmarks.
2 Evaluating Factual Consistency
Q2 scores response factual consistency by turning informative response spans into questions, answering them from grounding knowledge, and comparing the paired answers. NLI handles lexical variation, while filtering and fallback procedures address unreliable questions and answers.
- Q2 takes dialogue history, grounding knowledge, and a generated response, then scores how consistently the response reflects its grounding source.Its central assumption is that factual questions answerable from the response should receive similar answers from the knowledge.
- Question Generation: Q2 identifies named entities and noun phrases as informative response spans and generates questions whose answers should be those spans.It uses a question-generation system over each selected span and response.
- Question Answering: The QA stage answers each generated question from the grounding knowledge, including the possibility that no answer exists for hallucinated content.The resulting knowledge answer span is paired with the original response answer span.
- Final Scores and Filtering: Q2 averages answer-pair scores into response-level scores and then system-level scores, after filtering questions that fail answer validation or target personal opinions.When no valid questions remain, an end-to-end NLI fallback assigns 1 for entailment, 0 for contradiction, and 0.5 for neutral.
3 Evaluation Benchmarks
The evaluation uses three knowledge-grounded dialogue benchmarks with different grounding settings: Wikipedia-based dialogue, human-human conversations with varied sources, and persona or dialogue-history consistency. These settings test Q2 beyond a single dialogue domain.
- Wizard of Wikipedia: Wizard of Wikipedia contains dialogues where a bot responds knowledgeably using a relevant Wikipedia sentence as grounding.Because it lacks explicit factual-consistency annotations, the authors construct a separately annotated response dataset from it.
- Topical-Chat: Topical-Chat is a human-human knowledge-grounded dataset accompanied by Wikipedia pages, Washington Post articles, and Reddit fun-facts.The authors examine Q2’s correlation with human judgments of whether responses use knowledge.
- Dialogue NLI: Dialogue NLI pairs persona descriptions or dialogue-history utterances with subsequent utterances labeled entailment, neutral, or contradiction.It tests whether Q2 measures consistency when grounding is a persona sentence or prior dialogue history.
4 Dataset Creation and Annotation
The authors construct a manually annotated dataset of Wizard-of-Wikipedia responses from two dialogue systems, balancing consistent and inconsistent examples and validating annotation reliability.
- Dataset construction: Responses from MemNet and dodecaDialogue were sampled from the Wizard-of-Wikipedia validation set for factual-consistency annotation.Both systems generated responses using beam search with a beam size of 10.
- Annotation procedure: 300 factually inconsistent and 300 factually consistent responses were annotated, while chit-chat without grounding references was excluded.Annotators selected challenging inconsistent examples and skipped irrelevant general chit-chat among consistent responses.
- Annotation quality: Annotator agreement reached a Fleiss’ kappa of 0.853, indicating high agreement on factual consistency labels.The agreement check used a sample of 100 responses labeled by each annotator.
5 Experiments and Results
Experiments across three dialogue benchmarks evaluate Q2 against reference-free baselines at system and response levels. Q2 consistently correlates better with human judgments, separates consistent from inconsistent responses more effectively, generalizes across grounding domains, and remains interpretable, though chit-chat and unresolved dialogue-history references create failure cases.
- Wizard of Wikipedia: Q2 assigns higher scores to consistent outputs than inconsistent outputs, while random samples fall between them.This pattern holds across the evaluated baselines, with Q2 better separating general responses from inconsistent ones.
- Wizard of Wikipedia: 77.3% accuracy at threshold 0.5 makes Q2 more accurate than Q2 without NLI and end-to-end NLI for binary response classification.The threshold was selected for demonstration rather than tuned on separate development and test sets.
- System-level evaluation: Q2 achieves a 0.9798 average system-level correlation with human judgments, exceeding end-to-end NLI, overlap, BERTScore, and BLEU.The corresponding baseline correlations are 0.9216, 0.878, 0.8467, and 0.3051, respectively.
- Cross-benchmark evaluation: Q2 obtains significantly higher correlation than baselines on the Topical-Chat Uses Knowledge category, with p < 0.001.The NLI component produces larger correlation gains there than in the Wizard-of-Wikipedia experiments.
- Cross-benchmark evaluation: Across Wizard of Wikipedia, Topical-Chat, and Dialogue NLI, Q2 generalizes zero-shot without reference responses and outperforms end-to-end NLI baselines.The grounding sources include non-Wikipedia articles and facts, while Dialogue NLI concerns persona and self-consistency.
- Interpretability: Q2 outputs questions and answer spans that explain its score and can highlight potentially inconsistent response text.A qualitative example assigns score 0 when generated questions cannot be answered from the grounding knowledge.
- Error analysis: Q2 can fail when it generates questions for chit-chat or when pronouns require resolving entities from dialogue history.These cases reduce question quality or make faithful responses appear inconsistent.
6 Related Work
Prior evaluation metrics often overlook factual consistency, while grounded dialogue adds irrelevant utterance types and lacks suitable references. Q2 builds on question-generation and question-answering evaluation, using NLI comparison, and is reported to correlate better with human judgments than competing summary and dialogue metrics.
- Automatic dialogue evaluation is difficult because token-matching metrics correlate weakly or not at all with human judgments, while supervised methods require substantial annotation.
- QA-based metrics evaluate whether source information appears in generated text, whereas QG-QA methods use answer similarity to identify factual inconsistencies in summaries.
- Q2 extends QG-QA evaluation to knowledge-grounded dialogue and compares answer spans with NLI rather than token matching, allowing lexical variability.
- QG-QA methods showed higher correlations with human judgments of factual consistency than other automatic evaluation methods for abstractive summaries.
7 Conclusion and Future Work
The paper presents Q2, a reference-free metric combining question generation, question answering, and NLI for factual-consistency evaluation in knowledge-grounded dialogue. Experiments using annotated Wizard-of-Wikipedia responses and two additional datasets report strong results, while future work targets utterance-type separation and broader applications.
- Q2 combines question generation, question answering, and NLI without requiring reference responses to evaluate factual consistency in knowledge-grounded dialogue.
- The authors compiled and manually annotated Wizard-of-Wikipedia responses, then evaluated Q2 on that dataset, Topical-Chat, and DialogueNLI against multiple baselines.
- Future work proposes mapping response spans to chit-chat, persona, and factual types so each can be evaluated against its appropriate source of truth.
- The authors also propose applying Q2 to tasks such as automated fact-checking and using its evaluation signal to improve factual consistency in generation models.
A Ablation Study
The ablation study examines question-decoding strategy and filtering of personal-statement questions, reporting effects on Q2 scores and question coverage. Greedy decoding generally raises scores, while removing personal-statement filtering lowers scores but increases coverage.
- The ablations measure Q2 scores together with question coverage, defined as responses receiving at least one valid question without the end-to-end NLI fallback.
- Greedy decoding generates one question per answer candidate instead of selecting the top n questions with beam search.
- Greedy decoding increases scores on all data splits except MemNet consistent responses, while top-n decoding improves Q2 question coverage.
- Removing the filter for questions about personal statements lowers scores across all data splits while naturally increasing question coverage.
B Computing Infrastructure
The computing-infrastructure report records the experimental CPU setup and runtime for each data split. Each experiment used four CPUs, and processing a split of 150 responses took approximately 1.5–2 hours.
- Each experiment ran on 4 CPUs.
- Processing each data split of 150 responses took approximately 1.5–2 hours, and the authors planned a more efficient version of Q2.
C Additional Experiments
Additional experiments test whether Q² responds to adversarial knowledge changes and whether response length acts as a simple surface cue. Q² is expected to score randomly selected knowledge very low, while response-length comparisons find no strong differences.
- Random Knowledge: Q² is tested with knowledge randomly selected from another turn or another dialogue, where scores are expected to be extremely low.The first variant preserves the dialogue but changes the turn; the second uses knowledge from a different dialogue.
- Response Length: No strong differences in average character or token counts distinguish consistent from inconsistent responses.This result holds for dodeca outputs, with similar results reported for MemNet.
D Additional Graphs
The additional graphs show response-score distributions for Q², its token-matching variant, and the overlap baseline across consistent and inconsistent data. The surrounding evaluation materials define the dialogue grounding and annotation categories used to construct these comparisons.
- Additional Graphs: Figures 4–6 compare response-level score distributions for Q², Q² with token matching, and the overlap baseline on consistent and inconsistent data.Each figure separates the distributions for inconsistent and consistent responses.
- Dialogue Setting: Each dialogue turn provides a relevant Wikipedia sentence alongside the conversation context, with the knowledge integrated into the dialogue.This establishes the grounding setting evaluated by the figures.
- Annotation Criteria: Inconsistent responses include information absent from the knowledge or subtle changes that produce information not expressed by it.Annotators were instructed to identify both wholly unsupported information and semantic alterations of the knowledge.
- Annotation Criteria: Annotators checked clarity and coherence, selected responses that used the given knowledge, and ignored uninformative chit-chat.The instructions also required annotators to focus on the information provided to the bot rather than background knowledge.