BiCon-Gate: Consistency-Gated De-colloquialisation for Dialogue Fact-Checking
BiCon-Gate is a new method that boosts automated fact-checking in conversations by using a technique called consistency-gated de-colloquialisation. It creates conservative rewritten options for claims through a step-by-step de-colloquialisation, which combines light surface normalization with focused coreference resolution. A special consistency gate, which is aware of the dialogue’s meaning, picks a rewritten option only if it aligns with the context; if not, it sticks with the original claim. This careful selection process makes fact-checking more reliable and enhances both evidence gathering and verification. When tested on the DialFact benchmark, BiCon-Gate outperformed other methods, especially excelling in SUPPORTS and beating a competitive one-shot LLM rewrite approach.
Key facts
- BiCon-Gate is a consistency-gated de-colloquialisation method for dialogue fact-checking.
- It generates conservative rewrite candidates via staged de-colloquialisation.
- The method combines lightweight surface normalisation with scoped in-claim coreference resolution.
- A semantics-aware consistency gate selects the rewrite candidate only when semantically supported by the dialogue context.
- The gated selection stabilises downstream fact-checking.
- It yields gains in both evidence retrieval and fact verification.
- On the DialFact benchmark, it improves retrieval and verification with strong gains on SUPPORTS.
- It outperforms competitive baselines including a decoder-based one-shot LLM rewrite.
Entities
—