基于多特征融合的ATT-BiLSTM-CNN危化品问句分类  

ATT-BiLSTM-CNN Hazardous Chemicals Question Classification Based on Multi-feature Fusion

作　　者：胥心心 朱全银 孙纪舟 王文川 XU Xinxin;ZHU Quanyin;SUN Jizhou;WANG Wenchuan(School of Computer and Software Engineering,Huaiyin Institute of Technology,Huai'an 223001)

机构地区：[1]淮阴工学院计算机与软件工程学院,淮安223001

出　　处：《计算机与数字工程》2023年第8期1738-1744,共7页Computer & Digital Engineering

基　　金：青年科学基金项目“多源数据修复问题的关键技术研究”(编号:62002131)资助。

摘　　要：针对危化品问句文本较短、语义理解困难以及噪声大特点,提出一种基于多特征融合的ATT-BiLSTM-CNN问句分类模型。针对危化品问句文本特征,构建危化品专业词汇表用于Bert模型训练,利用已训练好的Bert模型处理问句文本可以增强危化品问句文本的语义表达和语义理解。其次,利用双层BiLSTM+注意力机制和卷积神经网络共同提取文本特征,可以更好地提取全局特征和局部特征以及增强关键特征。最后,将提取到的两部分特征融合后作为softmax层的输入,计算分类概率得到问句类别。实验结果表明该模型与传统问句分类算法相比F1值、召回率和准确率分别提升了2.75%~10.58%、2.53%~11.11%、3.38%~10.63%,在危化品领域的问句分类任务中取得了良好的效果。Aiming at the characteristics of short text,difficulty in semantic understanding and high noise of dangerous chemical question sentences,an ATT-BiLSTM-CNN question classification model based on multi-feature fusion is proposed.According to the characteristics of the question text of dangerous chemicals,a professional vocabulary of dangerous chemicals is constructed for Bert model training.The trained Bert model is used to process the question text can enhance the semantic expression and semantic understanding of the dangerous chemicals question text.Secondly,using the double-layer BiLSTM+Attention mechanism and convolutional neural network to extract text features together can better extract global features and local features and enhance key features.Finally,the two extracted features are fused as the input of the softmax layer,and the classification probability is calculated to obtain the question category.The experimental results show that compared with the traditional question classification algorithm,the F1 value,recall rate and accuracy rate of the model are increased by 2.75%~10.58%,2.53%~11.11%,3.38%~10.63%,respectively.Good results have been obtained in the task of question classification in the field of hazardous chemicals.

关 键 词：危化品问句分类 Bert 双层BiLSTM 注意力机制 卷积神经网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]