A Preliminary Attempt to Generate a Sichuan Dialect Handbook by LLMs

Against the background that structured resources are rare and manually compiled dictionaries are inefficient, this paper puts forward an automatic construction framework for Sichuan dialect handbook based on the LLM. We propose an end-to-end pipeline consisting of corpus generation, vocabulary filtering, and entry explanation generation. Corpus generation is guided by scenario prompt templates, which instruct the DeepSeek language model to generate excellent dialect dialogues. And candidate vocabulary is filtered by word rule extraction. Finally, corresponding explanations are automatically generated. The experimental results show that the obtained handbook contains about 15,270 entries of dialects in total. And in evaluation we use three labels (Generally Accurate, Incorrect, Uncertain) to annotate all our words. The results showing 38.9% labeled as Generally Accurate, 48.1% as Uncertain, and 12.9% as Incorrect. This approach provides a practical technical solution for the digitization of dialects and the development of low-resource language corpora.