结合注意力和双向LSTM的开放域问句分类研究  

Research on the Open-domain Question Classification Based on Attention and Bi-directional LSTM

作　　者：苏雪峰 SU Xuefeng(School of Modern Logistics,Shanxi Vocational University of Engineering Science and Technology,Jinzhong 030619)

机构地区：[1]山西工程科技职业大学现代物流学院,晋中030619

出　　处：《办公自动化》2023年第3期6-10,共5页Office Informatization

基　　金：2019年山西省哲学社会科学规划课题(2019B453)。

摘　　要：针对现有端对端模型没有对问句局部信息进行显式建模以及模型可解释性差等方面的不足,本文提出面向开放域的基于注意力机制和双向LSTM的问句分类方法。该法一方面使用注意力机制捕捉问句的局部信息;另一方面将注意力机制视为一种模型内置的自解释机制,将其与双向LSTM结合完成对问句局部和全局信息的建模。在TREC、MSQC、Baidu-Zhidao、Baidu-Search四个公开的开放域问句分类数据集上的实验结果表明,本文提出的方法在分类性上优于现有的基准方法,而且该方法的注意力机制能捕捉到问句分类的关键局部信息,提高模型的可解释性,为下游任务提供除类别以外的关键信息。The existing end-to-end model did not model explicitly the local information of the question sentence and the model interpretability was poor. To solve these problems, we proposed a question classification method for open-domain based on attention mechanism and bi-directional LSTM. This method regarded the attention as a built-in self-explanatory mechanism to capture the local information of the question, and captured the global information with bi-directional LSTM. The experiments on TREC, MSQC, Baidu-Zhi, Baidu-Search show that the proposed method is more superiority than the state-of-art methods in classification performance. Moreover, this method can capture the key local information of the question with the help of attention mechanism, which promotes the interpretability of the model in further and provides the key information for downstream tasks besides class information.

关 键 词：开放域 问句分类 注意力机制 可解释性 双向长短期记忆(LSTM)网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]