Post-hoc Utterance Refining for Faithful Knowledge Grounded Conversations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing knowledge grounded conversation (KGC) models suffer from entity-level hallucination, leading to inaccurate and unfaithful utterances despite having access to ground truth knowledge.
Innovation Solution
A post-hoc utterance refining method called REM (Refining by Entity Mining) is introduced, which uses entity mining to extract named entities from source knowledge and regenerate utterances to improve faithfulness, allowing for plug-and-play refinement of unfaithful utterances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a knowledge grounded conversation (KGC) model generates utterances based on source knowledge, then the conversation provides informative content, but entity-level hallucination occurs causing misinformation
Solution Approach 1:
The system implements a feedback mechanism by calculating a source-faithfulness score for generated utterances and using this score to determine whether refinement is needed. The faithfulness score acts as feedback that guides the entity mining and utterance regeneration process, allowing the system to iteratively improve faithfulness while maintaining information quality.
Solution Approach 2:
The patent introduces an intermediary entity mining process that extracts named entities from source knowledge before generating utterances. This intermediary step ensures that only entities present in the source knowledge are used in generated utterances, thereby preventing entity-level hallucination while maintaining information accuracy.
2Reliability
If entity mining is performed to extract named entities from source knowledge, then faithfulness to source knowledge improves, but computational complexity increases
Solution Approach 1:
The system performs entity mining as a preliminary action before utterance generation. By extracting and storing named entities from source knowledge in advance, the system prepares a curated set of valid entities that can be directly used during generation, avoiding the need for complex real-time verification and reducing overall processing complexity.
Solution Approach 2:
The patent segments the KGC process into distinct modules: entity mining, faithfulness scoring, and utterance regeneration. This segmentation allows each component to be optimized independently and enables selective application of refinement based on faithfulness scores, reducing unnecessary computational overhead.
3Reliability
If utterances are regenerated based on mined entities, then entity-level hallucination is reduced, but generation time increases
Solution Approach 1:
The system applies partial refinement by only regenerating utterances that have low source-faithfulness scores. Utterances with high faithfulness scores are kept as-is, avoiding unnecessary regeneration and reducing overall generation time while still correcting unfaithful utterances.
Solution Approach 2:
The patent changes the parameter of faithfulness scoring threshold to control the balance between faithfulness and generation time. By adjusting the threshold, the system can selectively refine only the most problematic utterances, optimizing the trade-off between reliability and efficiency.
Data Source
AI summary
Disclosed is a device and method for post-hoc utterance refining for faithful knowledge grounded conversation. The post-hoc utterance refining method is performed by a computing device including at least one processor, and includes generating, by a knowledge grounded conversation (KGC) model, utterance corresponding to knowledge and dialogue history; determining whether to refine the utterance based on source-faithfulness score indicating a degree to which the utterance reflects the knowledge; extracting a named entity from the knowledge; and regenerating the utterance based on the knowledge, the utterance, and the named entity.