融合字段类型与文本匹配的中文问句解析
CHINESE QUESTION PARSING BASED ON FIELD TYPES AND TEXT MATCHING
纪相存 1李大林 2彭晓东2
作者信息
- 1. 中国科学院国家空间科学中心 北京 101499;中国科学院大学 北京 100049
- 2. 中国科学院国家空间科学中心 北京 101499
- 折叠
摘要
自然语言转SQL语句技术可以帮助用户使用数据库,而WikiSQL数据集对表格内容的保护一定程度上限制了模型的使用,基于此,提出一种融合字段类型与文本匹配的中文问句解析方法.基于SQL结构分解问句解析任务,通过字段类型相关的分隔符将表结构信息结合到RoBERTa编码器输入中,并使用结合编辑距离与语义词典的文本匹配来使模型更加鲁棒.在中文数据集TableQA进行测试,该方法取得了最好的效果,正确率达到 93.44％.
Abstract
Translating natural language questions to SQL statements can help more users to obtain what they want from the database.The protection of table content by the English dataset WikiSQL limits the migration and use of the model to a certain extent.In order to solve this problem,this paper proposes a Chinese question parsing methods combining the field types with text matching.The task was decomposed based on the SQL structure.The table structure information was combined into the input of the Roberta encoder through the column separators related to the field types.The text matching method combining the edit distance and semantic dictionary was used to make the model more robust.This method was tested on the more difficult Chinese dataset TableQA.The accuracy rate was up to 93.44％and the result verified that the method was efficient.
关键词
自然语言转SQL语句/表结构信息/SQL结构/文本匹配Key words
Natural language to SQL statements/Table structure/SQL structure/Text matching引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据