An LLM-Assisted Framework for Response Selection in Low-Resource Language Dialogue: A Case Study on Taiwan Indigenous Languages

Dialogue tasks in low-resource languages face numerous challenges. Directly applying Large Language Models (LLMs) for response generation often leads to issues such as grammatical errors or hallucinations. Dialogue Response Selection (DRS), which selects responses from a predefined corpus, offers a more reliable alternative by reducing such risks. However, in extremely low-resource scenarios, limited retrieval capabilities and the lack of training data present significant challenges for practical implementation. To address this, this study proposes a two-stage DRS-based framework for dialogue systems in Taiwanese Indigenous languages. The first stage uses On-the-fly Translation to convert user inputs from low-resource languages into Chinese, leveraging a high-resource language as a pivot to improve retrieval accuracy. In the second stage, response candidates are retrieved based on the translated input and ranked by an LLM serving as the response selector, eliminating the need for additional model fine-tuning. To support evaluation under low-resource constraints, we also construct a simulated user input dataset and propose LaRQ (LLM-guided Response Quality Evaluation), an automated metric shown to align closely with human judgments. Overall, our method improves LaRQ over the no-pivot baseline by about 0.3 on average and up to 0.44 across models, demonstrating the effectiveness of pivot-language translation and LLM-based selection in enhancing dialogue quality for low-resource languages.