基于预训练语义编码的判断句答案推理
Judgment Answer Inference Bsaed on Pre-trained Semantic Encoding
李飞 1王颜颜 2王超 2黄友志2
作者信息
- 1. 中国科学技术大学 计算机科学与技术学院,安徽 合肥 230026
- 2. 科大国创云网科技有限公司,安徽 合肥 230088
- 折叠
摘要
目前大规模文本问答依赖句子表征从候选文本中检索答案,但是忽略了有些答案需要进一步推理,无法直接从文中获取,比如判断句.为解决此类问题,一个面向大规模文本的判断句答案生成方法被提出.首先在语义编码器中通过对大规模文本进行预训练获取语义编码器,对问题、依据进行语义编码;其次在答案生成器中基于对比学习构造正负样本进行数据增强;之后在答案依据获取器中通过使用Faiss实现问题和大规模文本的快速表征与匹配.在最终的判断句问答中,准确率高达96.58%,验证了该方法的有效性.
Abstract
Currently,large-scale text question answering relies on sentence representation to retrieve answers from candidate texts,but it ig-nores that some answers require further reasoning and cannot be obtained directly from the text,such as judgment sentences.To solve such problems,a judgment sentence answer generation method for large-scale text is proposed.Firstly,in the semantic encoder,the semantic en-coder is obtained by continuing to pre-train large-scale texts,and the questions and cues are semantically encoded.Sceondly,in the answer generator module,positive and negative samples are constructed based on contrastive learning for data enhancement.Then fast characteriza-tion and matching of questions and large-scale text is achieved by using Faiss in the answer basis obtainer.The accuracy of the final judgment sentence question and answer is as high as 96.58%,which verifies the effectiveness of this method.
关键词
智能问答/开放域问答/问答分类器/语义编码Key words
intelligent question answering/open domain Q&A/Q&A classifier/semantic coding引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据