LLM-Enhanced Chinese NL2SQL Translation Task Under Resource-Limited Condition
Xichonglang Xiao, Hao Xu · 2024
With the development of big data and artificial intelligence technology, data has become an important production factor and core resource. There is also a growing demand for real-time accurate, convenient and fast data analysis in the military field.NL2SQL technology can be embedded into various intelligent platforms, data analysis tools or intelligent assistants, realizing seamless dialogue between users and data, responding instantly to complex analytical requests, and providing timely and accurate support of decision-making for commanders. However, the Chinese NL2SQL task has a low success rate compared to the English task due to the base model’s bias in converting Chinese entities, difficulty in linking schemas, and unclear thought chain. In this paper, based on the llama3-8b model, we explored the NL2SQL technique under resource-constrained conditions.By analyzing the characteristics of the model’s erroneous output results under regular prompts,target refinements of Chinese prompts are designed.And through supervised fine-tuning, we have made the model’s accuracy rate on the dev set increase from the initial 46% to 70%.