[Paper Review] WikiChat: Stopping the Hallucination of Large Language Model Chatbots by Few-Shot Grounding on Wikipedia

Sina J. Semnani, Violet Z. Yao+2 more|arXiv (Cornell University)|May 23, 2023

Topic ModelingComputer Science43 references9 citations

TL;DR

WikiChat is a few-shot, Wikipedia-grounded chatbot pipeline that combines retrieval with LLM generation and per-claim fact-checking, distilled to a smaller model for latency, cost, and privacy gains.

ABSTRACT

This paper presents the first few-shot LLM-based chatbot that almost never hallucinates and has high conversationality and low latency. WikiChat is grounded on the English Wikipedia, the largest curated free-text corpus. WikiChat generates a response from an LLM, retains only the grounded facts, and combines them with additional information it retrieves from the corpus to form factual and engaging responses. We distill WikiChat based on GPT-4 into a 7B-parameter LLaMA model with minimal loss of quality, to significantly improve its latency, cost and privacy, and facilitate research and deployment. Using a novel hybrid human-and-LLM evaluation methodology, we show that our best system achieves 97.3% factual accuracy in simulated conversations. It significantly outperforms all retrieval-based and LLM-based baselines, and by 3.9%, 38.6% and 51.0% on head, tail and recent knowledge compared to GPT-4. Compared to previous state-of-the-art retrieval-based chatbots, WikiChat is also significantly more informative and engaging, just like an LLM. WikiChat achieves 97.9% factual accuracy in conversations with human users about recent topics, 55.0% better than GPT-4, while receiving significantly higher user ratings and more favorable comments.

Motivation & Objective

• Improve factuality in open-domain chatbots by grounding LLM outputs in a trusted corpus (Wikipedia).
• Achieve high conversationality and low latency through a 7-stage retrieval-and-generation pipeline.
• Demonstrate distillation of a multi-stage system into a smaller model without sacrificing quality.
• Provide an evaluation methodology combining simulated and real-user data with human-LLM assessments.

Proposed method

• Stage 1: generate a query from user utterance and retrieve Wikipedia passages with temporal re-ranking.
• Stage 2: extract and summarize relevant passages into bullet points for grounding.
• Stage 3: prompt LLM to generate a response from conversation history and bullet points.
• Stage 4: decompose LLM response into claims and retrieve evidence for each claim.
• Stage 5: use a chain-of-thought prompt to classify each claim as supported, refuted, or indeterminate, discarding unsupported claims.
• Stage 6: draft a final response from grounded bullets and history; Stage 7: refine the draft using feedback on relevance, naturalness, non-repetition, and temporal correctness.

Experimental results

Research questions

• RQ1 Can a few-shot LLM grounded in a trusted corpus produce factual and engaging responses with low hallucination rates?
• RQ2 How does a 7-stage retrieval-grounded pipeline compare to retrieval-only and pure LLM baselines in factuality, conversationality, and latency?
• RQ3 Does distilling WikiChat into a smaller model preserve factuality and conversationality while reducing latency and cost?
• RQ4 What evaluation methodology best assesses factuality and conversationality for knowledge-grounded chatbots?

Key findings

• WikiChat GPT-4 achieves 97.3% factual accuracy in simulated conversations and 97.9% in real-user conversations.
• WikiChat variants outperform Atlas (state-of-the-art retrieval-based model) in factuality and match conversationality with LLMs.
• Distilling WikiChat G4 into a 7B LLaMA model yields 91.1% factual accuracy with 3.2x lower end-to-end latency than the teacher.
• WikiChat's factual accuracy advantage is strongest on tail and recent knowledge compared to GPT-4.
• One-third of claims in final responses are rejected by the fact-checking stage, highlighting the importance of claim-level verification.