RHO ($ρ$): Reducing Hallucination in Open-domain Dialogues with Knowledge Grounding

Ziwei Ji, Zihan Liu, Nayeon Lee, Tiezheng Yu, Bryan Wilie, Min Zeng, Pascale Fung

TL;DR

RHO ($\rho$) is a novel framework designed to tackle the persistent problem of "hallucinations"—factual errors—in knowledge-grounded dialogue systems. By integrating structured Knowledge Graph (KG) embeddings directly into the model's architecture and using a specialized re-ranking mechanism, RHO significantly improves the truthfulness of AI responses, achieving a 17.54% boost in faithfulness metrics over previous state-of-the-art models.

What Problem Does This Paper Address?

Large language models (LLMs) are remarkably fluent, but they have a habit of "making things up." In the context of Knowledge-Grounded Dialogue (KGD), where a system is supposed to use external facts to inform its responses, these hallucinations are particularly problematic.

The authors categorize these errors into two types:

  1. Intrinsic Hallucinations: The response directly contradicts the provided knowledge.
  2. Extrinsic Hallucinations: The response includes information that cannot be verified by the provided source (see Figure 1 for an example).

The core challenge is the heterogeneity between unstructured dialogue text and structured knowledge graphs. Current models struggle to fuse these two different types of information effectively, leading to responses that sound plausible but are factually untethered.

An example of intrinsic and extrinsic hallucinations in dialogue

Methods

The RHO framework (illustrated in Figure 2) introduces three key mechanisms to ground the dialogue in reality:

  1. Local Knowledge Grounding (LKG): The system identifies entity and relation mentions within the dialogue. It then fetches their specific embeddings from a Knowledge Graph (trained via TransE) and fuses them directly into the model's token-level representations. This ensures the model "knows" exactly which real-world entity it is talking about.
  2. Global Knowledge Grounding (GKG): To enable multi-hop reasoning, RHO maintains a memory bank of the entire relevant KG sub-graph. Using an attention mechanism, the model learns to weigh different triples in the graph based on their relevance to the current conversation context.
  3. Response Re-ranking (RR): Instead of just picking the most likely sentence the model generates, RHO generates several candidates and uses a conversational reasoning model to re-rank them. This re-ranker checks if a candidate response can be "reasoned backward" to a valid path in the Knowledge Graph. If the path doesn't exist, the response is penalized as a likely hallucination.
Figure 2: The overview of the proposed RHO framework. The input follows a knowledgeable task guidance tem- plate, including meta-knowledge and dialogue history (§3.2). To facilitate knowledge grounding, we first employ TransE to gain KG embeddings. For local knowledge grounding, we adapt entity/relation linking to recognize men- tions in the dialogue context. The corresponding KG embeddings are locally fused into textual embeddings (§3.3). For global knowledge grounding, we aggregate the entire knowledge sub-graph in a memory bank via the attention mechanism so that the textual embeddings receive all context-related knowledge with emphasis (§3.4). After that, the encoder-decoder model generates several candidate responses. During post-processing, a re-ranker trained by traversal over a knowledge sub-graph conditioned on the dialogue context selects the most faithful response as the final output (§3.5).

Datasets

The researchers evaluated RHO using OpenDialKG, currently the only publicly available English corpus for open-ended dialogues that includes explicit Knowledge Graph path annotations. They filtered the dataset to keep only samples with these annotations, splitting it into standard training, validation, and testing sets (8:1:1 ratio).

Key Insights and Ideas

  • The Power of Dual Grounding: Local grounding provides precision for specific entities, while global grounding provides the "big picture" context needed to connect disparate facts across multiple steps (multi-hop reasoning).
  • Verification via Re-ranking: The authors treat faithfulness as a search problem. By checking if a generated response corresponds to a legitimate traversal of a factual graph, they create a robust filter for nonsense.
  • The Reference Paradox: Interestingly, the authors found a slight trade-off between traditional metrics like BLEU (which measures similarity to a reference text) and faithfulness metrics. This is because the "ground truth" human references in existing datasets often contain their own extrinsic hallucinations. RHO is sometimes "too factual" for the original dataset's noisy references!

Key Results

RHO outperformed strong baselines, including GPT-2, BART, and KG-BART, across several dimensions:

  • Faithfulness: Improved by 17.54% on the FeQA metric and 9.33% on QuestEval compared to the previous state-of-the-art (GPT2+NPH).
  • Entity Coverage: Achieved a 33.97% improvement in F1 score for entity coverage, meaning it uses the provided facts much more accurately.
  • Human Evaluation: In blind tests, humans rated RHO as significantly more faithful (81.67%) than the SOTA baseline (72.67%), specifically noting a massive reduction in extrinsic hallucinations.
Automatic evaluation results showing RHO's superiority

Limitations and Future Work

Despite its success, the authors acknowledge several hurdles:

  • Black Box Nature: While RHO is more grounded, the underlying neural networks remain "black boxes," meaning 100% faithfulness cannot be mathematically guaranteed.
  • Flexibility vs. Faithfulness: The model's intense focus on factual grounding can sometimes make it less flexible or diverse in its conversational style.
  • Reasoning Depth: The model detects statistical regularities but doesn't yet truly "understand" qualitative abstractions like causality or complex hierarchies.

Future work aims to combine RHO's re-ranking logic with massive foundation models like ChatGPT to automate the labor-intensive human feedback loops currently used to keep LLMs honest.

Summary updated

Abstract

Dialogue systems can leverage large pre-trained language models and knowledge to generate fluent and informative responses. However, these models are still prone to produce hallucinated responses not supported by the input source, which greatly hinders their application. The heterogeneity between external knowledge and dialogue context challenges representation learning and source integration, and further contributes to unfaithfulness. To handle this challenge and generate more faithful responses, this paper presents RHO ($ρ$) utilizing the representations of linked entities and relation predicates from a knowledge graph (KG). We propose (1) local knowledge grounding to combine textual embeddings with the corresponding KG embeddings; and (2) global knowledge grounding to equip RHO with multi-hop reasoning abilities via the attention mechanism. In addition, we devise a response re-ranking technique based on walks over KG sub-graphs for better conversational reasoning. Experimental results on OpenDialKG show that our approach significantly outperforms state-of-the-art methods on both automatic and human evaluation by a large margin, especially in hallucination reduction (17.54% in FeQA).
BibTeX
@inproceedings{ji2023rho,
  author = {Ziwei Ji and Zihan Liu and Nayeon Lee and Tiezheng Yu and Bryan Wilie and Min Zeng and Pascale Fung},
  title = {{RHO ($ρ$): Reducing Hallucination in Open-domain Dialogues with Knowledge Grounding}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
  doi = {10.18653/v1/2023.findings-acl.275},
}
Figure