PaperlayerView exact revision on arXiv

Source-linked AI summary

Retrieval Augmentation Reduces Hallucination in Conversation

Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, Jason Weston

arXiv:2104.07567v1cs.CLcs.AI

TL;DR

State-of-the-art dialogue models can produce fluent but factually incorrect responses, motivating retrieval for knowledge-grounded conversation. The paper studies retrieval-augmented retrievers, rankers, and encoder-decoders, finding state-of-the-art results, substantially reduced hallucination, and stronger generalization beyond training data. These gains are reported while maintaining conversational ability, although the experiments use Wikipedia and leave broader knowledge sources and long-term memory for future work.

  • Problem

    Dialogue models often generate plausible but factually incorrect statements, while knowledge-grounded dialogue requires retrieval over complex multi-turn context.

  • Method

    The paper studies retrieval-augmented dialogue architectures combining retrievers, rankers, and encoder-decoders, including several new retrieval variants.

  • Results

    The best models achieved state-of-the-art results on two knowledge-grounded conversational tasks and reduced hallucinated responses by over 60%.

  • Takeaways & Limitations

    Retrieval-augmented generation substantially reduces dialogue hallucination, generalizes to out-of-distribution data, and maintains conversational ability.

  • Takeaways & Limitations

    The experiments use Wikipedia as the knowledge source, while broader document choices and long-term memory augmentation remain future directions.

Abstract

from arXiv · show

Despite showing increasingly human-like conversational abilities, state-of-the-art dialogue models often suffer from factual incorrectness and hallucination of knowledge (Roller et al., 2020). In this work we explore the use of neural-retrieval-in-the-loop architectures - recently shown to be effective in open-domain QA (Lewis et al., 2020b; Izacard and Grave, 2020) - for knowledge-grounded dialogue, a task that is arguably more challenging as it requires querying based on complex multi-turn dialogue context and generating conversationally coherent responses. We study various types of architectures with multiple components - retrievers, rankers, and encoder-decoders - with the goal of maximizing knowledgeability while retaining conversational ability. We demonstrate that our best models obtain state-of-the-art performance on two knowledge-grounded conversational tasks. The models exhibit open-domain conversational capabilities, generalize effectively to scenarios not within the training data, and, as verified by human evaluations, substantially reduce the well-known problem of knowledge hallucination in state-of-the-art chatbots.

1 Introduction

Large dialogue models are fluent but can generate plausible factual errors, especially across similar entities or when one token changes correctness. The paper applies retrieval-augmented architectures to dialogue and reports improved factuality while preserving conversational ability.

  • Large language models store knowledge implicitly in billions of parameters, supporting fluent and somewhat knowledgeable open-domain dialogue.
  • They nevertheless hallucinate plausible-looking factual errors, including mixed-up facts between similar entities and single-token mistakes.
  • Retrieval-augmented generation encodes dialogue context together with passages retrieved from a large unstructured document set, but multi-turn dialogue makes retrieval harder than single-fact question answering.
  • The paper studies retrievers, rankers, and encoder-decoders, including Poly-encoder scoring, iterative retrieval, end-to-end retrievers, and turn-based retrieval.
  • Over 60% fewer hallucinated responses were observed, while knowledgeability gains over baseline reached 70% in-distribution and 85% out-of-distribution.
  • The best models achieved state-of-the-art results on Wizard of Wikipedia and CMU Document Grounded Conversations.

2 Related Work

Related work spans hallucination studies, neural retrieval for open-domain question answering, and knowledge-grounded dialogue. Unlike many dialogue datasets that provide gold knowledge passages, this paper considers learning retrieval from a large unstructured collection.

  • Hallucination has been studied in summarization, machine translation, news generation, and dialogue, but dialogue hallucination remained unresolved in prior work.
  • Neural open-domain question answering evolved from vector-space retrievers toward end-to-end generation models with neural retrieval components.
  • Recent neural retrieval approaches over unstructured text overtook methods exploiting graph structure in knowledge sources and became an attractive alternative for dialogue.
  • Many knowledge-grounded dialogue datasets provide a gold paragraph or passage, whereas this work requires models to retrieve knowledge from a large unstructured set.
  • Retrieval augmentation has also been explored for task-oriented dialogue, language-model pre-training, memory, and nearest-neighbor cache models.

3 Model Architectures

The paper extends retrieval-augmented generation to knowledge-grounded dialogue by combining retrievers, rankers, and encoder-decoders. It develops architectures for richer retrieval, iterative retrieval, turn-aware conditioning, and retriever-enhanced FiD.

  • 3 Model Architectures: The architecture notation represents dialogue contexts, labels, retrieved documents, retrieval probabilities, and generator probabilities.The paper uses q(x_i) and d(z_j) for context and document representations, p_η(z_j|x_i) for retrieval, and p_θ for generation.
  • 3.1 RAG and FiD: Neural retrieval scales to millions of documents and supplies passages that augment dialogue contexts before generation.RAG uses DPR representations and FAISS maximum inner product search to retrieve relevant passages; BART provides the base sequence-to-sequence model.
  • 3.3 Improving Retrieval: Poly-encoders and ColBERT increase context-candidate interaction to improve document ranking within retrieval-augmented architectures.Poly-encoders use context codes for candidate-sensitive scoring, while ColBERT uses token-level maxsim comparisons and late-stage re-ranking.
  • 3.3.2 Iterative Retrieval: Iterative retrieval performs two retrieval-generation rounds, with the second retrieval conditioned on the first generated output.ReGReT can share one model across both rounds or use separate models, and training accounts for both stages.
  • 3.4.1 Conditioning on Dialogue Turns: RAG-Turn marginalizes separately over dialogue turns before jointly marginalizing, addressing cumbersome long dialogue contexts.The method retrieves documents for each turn and can group earlier turns jointly while treating the most recent T* turns independently.
  • 3.4.2 Improving FiD: FiD fuses independently encoded document and context representations before decoding, but its retriever is not trained within the model.FiD-RAG combines FiD with a DPR-based retriever trained using RAG to adapt retrieval to knowledge-grounded dialogue.

4 Experiments

The experiments evaluate retrieval-augmented dialogue across two knowledge-grounded datasets, using automatic metrics and human evaluations. Retrieval improves knowledge use and reduces hallucination while generally preserving conversational ability, though evaluation reveals important architectural trade-offs.

  • Experimental setup: Experiments use Wizard of Wikipedia and CMU Document Grounded Conversations, with modified seen and unseen splits for evaluating knowledge-grounded dialogue.The CMU_DoG unseen split holds out conversations about two of its 30 movies; results focus on the modified splits.
  • Metrics: The evaluation combines perplexity, unigram F1, BLEU-4, ROUGE-L, Knowledge F1, Rare F1, and human judgments.Knowledge F1 measures overlap with knowledge used by human annotators, whereas Rare F1 restricts overlap to infrequent reference words.
  • Retrieval effectiveness: Retrieval-augmented BART-Large with RAG-Token DPR outperforms non-retrieval BART-Large on both datasets, and retrieval substantially improves performance overall.The comparison uses dialogue context while retrieving five documents from the full Wikipedia collection.
  • Knowledge use: Knowledge F1 shows a substantially larger gap between retrieval and non-retrieval models than regular F1, indicating that retrieval gains reflect greater use of relevant knowledge.The authors interpret these results as confirming that the models appropriately utilize knowledge.
  • Human evaluation: Human evaluations show that hallucination rates drop dramatically for retrieval-augmented models while knowledgeability rises sharply.The paper reports that its best models reduce hallucinated responses by over 60%.
  • Conversational ability: Factuality generally does not sacrifice conversational ability, although some models show slight declines in engagingness when grounding responses heavily in retrieved knowledge.Regular F1 remains similar across retrieval and non-retrieval variants, while Knowledge F1 differs more strongly.

4.3 Does retrieval help generalization to unseen distributions?

Retrieval augmentation improves generalization to unseen topics and distributions. On WoW, retrieval-augmented models show smaller performance losses than non-retrieval baselines.

  • Retrieval-free models suffer performance declines when evaluated on unseen topics and distributions.This pattern appears across automated metrics for WoW and modified CMU_DoG unseen splits.
  • 16%, 5%, and 8% are the corresponding RAG DPR-Poly drops in PPL, F1, and Knowledge F1.Retrieval-augmented models generally decrease performance less, particularly on Knowledge F1.
  • Retrieval-augmented models achieve new state-of-the-art results on the Wizard of Wikipedia Test Unseen split.

4.4 How should generation be augmented?

Generation augmentation choices strongly affect the balance between conversational quality and factual knowledge use. Across retrieval, reranking, and generator configurations, models using full dialogue context and neural retrieval generally outperform simpler alternatives.

  • RAG-Turn architectures: Retrieval over the whole dialogue context substantially outperforms retrieval based only on the most recent turn across all metrics.
  • RAG-Turn architectures: RAG-Turn models balance the trade-off between RAG-Sequence’s knowledge use and RAG-Token’s conversational performance.RAG-Sequence has higher Knowledge F1 but lower regular F1, while RAG-Turn methods bridge these strengths.
  • Retriever and architecture choices: 33% and 41% Knowledge F1 gains are achieved by FiD-RAG-Poly with BART on the seen and unseen splits, respectively.FiD-RAG with T5 yields corresponding gains of 37% and 25%.
  • Retriever and architecture choices: Neural retrieval outperforms TFIDF retrieval, whose fixed bag-of-words representations are less adaptable to the generator.The paper attributes the advantage to richer Transformer representations and end-to-end adaptation between retriever and generator.
  • Retriever and architecture choices: Adding a Poly-encoder reranker to DPR produces the best automated metrics on both Wizard of Wikipedia validation splits.ColBERT and PolyFAISS also perform strongly, while DPR already provides a strong retrieval base.
  • Generator architectures and size: BART-Large and T5 perform comparably when retrieval is held constant, whereas BlenderBot-400m performs worse.Perplexity is not directly comparable across generator architectures because their dictionaries differ.
  • Generator architectures and size: Larger models reduce perplexity but leave generation statistics roughly constant; for BlenderBot, increasing size decreases Knowledge F1.

4.7 Is a neural model trained for retrieval necessary?

The paper tests whether retrieval and generation can share an encoder instead of relying on a separate retrieval module. Shared-encoder models remain effective on restricted knowledge sources, but larger and noisier sources reduce retrieval precision and performance.

  • Shared-encoder retrieval: BREAD uses a shared encoder for retrieval and generation, replacing a separate retrieval module.The design treats the generator’s context representations as quasi-retrievers.
  • Knowledge-source scale: On a 3k-document, 500k-token Wikipedia subset, BREAD achieves performance similar to its DPR-retrieval counterpart.
  • Knowledge-source scale: Scaling BREAD to 11 million documents and 1 billion tokens slightly reduces performance, but it still retrieves relevant information and beats no retrieval.
  • Knowledge-source scale: As the knowledge source becomes larger and more unstructured, BREAD has difficulty identifying the relevant sentence within passages.
  • Retriever-less models: TREAD improves WoW Valid Seen F1 from 19.3 to 22.1 and Knowledge F1 from 14.6 to 24.1 relative to no retrieval.Its RAG and FiD counterparts nevertheless achieve better Knowledge F1 and perplexity.

4.8 Additional Relevant Ablations

Additional ablations examine decoding, retriever pre-training, knowledge-source selection, and the number of retrieved documents. They show that decoding and retrieval breadth involve important trade-offs between fluency, dialogue quality, and factual grounding.

  • Decoding strategies: Beam-blocking the dialogue context hurts performance, likely because it can block discussion of named entities from earlier turns.
  • Decoding strategies: Beam search yields the highest F1 scores, while nucleus sampling can retain comparable ROUGE-L and F1 yet suffer lower Knowledge F1.The authors associate the Knowledge F1 difference with increased hallucination under nucleus sampling.
  • Pre-training: Retriever pre-training produces minimal performance gains, and DPR-Poly reranker pre-training yields no noticeable downstream improvement.The results suggest that a retriever already in a good state can work under fine-tuning.
  • Knowledge sources: Restricting retrieval to nearly 3,000 documents that fully cover Wizard of Wikipedia topics produces noticeable performance gains.The smaller source makes the retrieval task easier.
  • Number of retrieved documents: For RAG-Token and FiD-RAG, retrieving more documents improves perplexity and F1 but substantially lowers Knowledge F1 and increases human-annotated hallucination.
  • Number of retrieved documents: For RAG-Sequence, more documents improve perplexity while maintaining Knowledge F1 and BLEU, but F1 drops beyond one document.The authors hypothesize that considering more generations increases effective beam size while reducing dialogue-like engagement.
  • Number of retrieved documents: The authors use 25 reranked documents to limit computational overhead because increasing the reranked count does not substantially improve performance.

5 Conclusion

The paper finds knowledge hallucination in conversational agents across language-model sizes and training data, while retrieval augmentation significantly reduces it. It also identifies broader knowledge sources and retriever-generator designs as future directions.

  • Knowledge hallucination occurs independently of language-model size or training data.
  • Retrieval-augmented generation significantly reduces hallucination in conversational agents.
  • The study uses Wikipedia as its only knowledge source, leaving other documents and long-term-memory augmentation for future exploration.

A Retriever Performance

Retriever performance is evaluated by whether retrieved passages contain the gold knowledge sentence. The evaluation reports top-1 and top-5 retrieval success.

  • The retrieval evaluation covers the retrievers considered in the study.
  • Retriever performance is measured by whether the top retrieved document or top five contain the gold knowledge sentence.
  • The gold knowledge sentence is the sentence used in Wizard of Wikipedia, and retrieval results are reported in Table 17.

B RAG Turn Further Explorations

RAG-Turn is examined across different effective numbers of dialogue context turns. Increasing the context-turn setting generally raises perplexity without improving Knowledge F1 consistently.

  • RAG-Turn is evaluated using different values of T*, the effective number of context turns considered.
  • Perplexity generally increases as the effective context-turn setting changes.
  • Knowledge F1 remains roughly stable, with marginal increases or decreases depending on the model.

C Automated Metrics and Human Evaluation

The study compares automated metrics with human evaluations across the reported model set and retrieval architectures. Knowledge F1 and Rare F1 align more closely with perceived knowledge use and reduced hallucination than standard F1.

  • Improvements in PPL, Knowledge F1, and Rare F1 correlate with greater perceived knowledge use and reduced hallucination.
  • Standard F1 shows relatively low correlation with all considered human-evaluation criteria.
  • The architecture comparisons include retrieval-augmented and non-retrieval models evaluated on Wizard of Wikipedia and CMU Document Grounded Conversations.
  • Figure 2 visualizes Pearson correlations between automated metrics and human evaluations on Wizard of Wikipedia Valid Unseen data.
Loading 2104.07567v1…