Experimental evaluation demonstrates enhanced dialogue quality in Taiwanese Indigenous languages, indicating viable conversational AI deployment without fine-tuning.
Key Points
To develop a two-stage dialogue response selection framework that mitigates hallucinations and grammatical errors in low-resource language dialogues by leveraging high-resource pivot translation.
Implemented a two-stage pipeline using on-the-fly translation of low-resource user queries into Chinese as a pivot language for candidate retrieval.
Employed an LLM as a zero-shot response selector to rank candidate responses without additional model fine-tuning.
Constructed a simulated user input dataset and introduced LaRQ (LLM-guided Response Quality Evaluation) to automate response assessment.
The pivot-based framework improved the LaRQ evaluation score by an average of approximately 0.3 over the no-pivot baseline.
Observed performance gains reached up to 0.44 across tested model architectures.