Detecting Hallucination and Coverage Errors in Retrieval Augmented Generation for Controversial Topics

TL;DR

This work introduces the NPOV Response Task, a retrieval augmented generation framework aimed at controversial topics and governed by Wikipedia’s Neutral Point of View. It combines a deterministic knowledge retrieval from Britannica ProCon with a soft-prompt-tuned large language model to generate multi-perspective introductions and verbatim argument verbalizations, while studying hallucination and coverage errors. The authors propose three error-detection approaches—ROUGE, salience, and LLM-based classifiers—and show that classifiers trained on synthetic errors achieve high ROC AUC (up to 95.3% for hallucinations and 90.5% for coverage errors) on unambiguous cases, with reasonable performance (84.0% and 85.2%) even without labeled training data. The results highlight the potential of data-efficient, multi-perspective, and neutral content generation, while also acknowledging limitations related to annotation, topic diversity, and computational costs, guiding future improvements in controlled, retrieval-augmented dialogue systems.

Abstract

We explore a strategy to handle controversial topics in LLM-based chatbots based on Wikipedia's Neutral Point of View (NPOV) principle: acknowledge the absence of a single true answer and surface multiple perspectives. We frame this as retrieval augmented generation, where perspectives are retrieved from a knowledge base and the LLM is tasked with generating a fluent and faithful response from the given perspectives. As a starting point, we use a deterministic retrieval system and then focus on common LLM failure modes that arise during this approach to text generation, namely hallucination and coverage errors. We propose and evaluate three methods to detect such errors based on (1) word-overlap, (2) salience, and (3) LLM-based classifiers. Our results demonstrate that LLM-based classifiers, even when trained only on synthetic errors, achieve high error detection performance, with ROC AUC scores of 95.3% for hallucination and 90.5% for coverage error detection on unambiguous error cases. We show that when no training data is available, our other methods still yield good results on hallucination (84.0%) and coverage error (85.2%) detection.

Figures (4)

• Figure 1: Example NPOV response to a user query on a senstive topic (left) based on pro and con perspectives with two arguments each, as obtained from a knowledge base (right). Arguments taken from procon. Our error detection methods focus on the NPOV main response.
• Figure 2: Top: salience map from input argument content words (rows) to model response content words (columns). Bottom: individual word scores for contribution (input words; left) and attribution (response words; right). The purple highlighted words are hallucinated in the model response.
• Figure 3: Task format for the NPOV Response Task.
• Figure 4: Task format for LLM-based error classifiers.