基于样本构造和孪生胶囊网络的医学意图识别
Medical intent detection based on sample construction and siamese capsule network
投稿时间:2022-04-18  修订日期:2022-04-18
DOI:
中文关键词: 意图识别  孪生胶囊网络  医学信息  语义相似度
英文关键词: intent detection  siamese capsule network  medical information  semantic similarity
基金项目:国家重点研发计划资助项目(2020YFC1522900);中央高校基本科研业务费专项资金资助项目(CZZ21001);中南民族大学研究生学术创新基金资助项目(3212022SYCXJJ337)
作者单位邮编
龚建全 中南民族大学 计算机科学学院 
王德军 中南民族大学 计算机科学学院 430074
孟博 中南民族大学 计算机科学学院 
摘要点击次数: 149
全文下载次数: 0
中文摘要:
      意图识别是自然语言理解中的重要任务,为构建医疗领域对话系统奠定了基础.然而医学问句特征稀疏、种类易混淆,导致识别率不高.针对上述问题,提出一种结合BERT和胶囊网络的S-BCN模型,并将意图识别任务转化为语义相似度任务进行处理.首先将意图分类数据集构造成问句对样本和问句意图对样本进行分阶段训练,一阶段通过将问句对样本输入BERT层生成词向量矩阵,然后通过胶囊网络层提取出局部语义特征,得到问句的向量表示,再通过计算向量的余弦距离得到问句对的语义相似度;二阶段将问句意图对样本输入模型再次进行训练,最后通过打分模块得到问句的意图类别.该方法在中文医疗信息处理挑战榜CBLUE(Chinese Biomedical Language Understanding Evaluation)的意图分类数据集上测试,其准确率达到86.76%,相比基线模型BERT提高2.61%.
英文摘要:
      Intent detection is an important task in natural language understanding, which lays a foundation for constructing medical dialogue system. However, medical question features are sparse and confusing, resulting in low accuracy. To solve the above problems, a S-BCN model combining BERT and capsule network is proposed, and the intent detection task is transformed into semantic similarity task. Firstly, the intent detection dataset is constructed into question pair samples and question intent pair samples to train in stages. In the first stage, the question pair samples are input into the BERT layer to generate word vectors matrix. Then, the local semantic features are extracted by the capsule network layer, and the vector representation of the question is obtained. Finally, the semantic similarity of the question pair is obtained by calculating the chord distance of the vector. In the second stage, the question intent is re-entered into the twin network training for the sample, and finally the question intent category is obtained through the scoring module. The method is tested on the intent classification dataset of CBLUE ( Chinese Biomedical Language Understanding Evaluation ), and the accuracy reaches 86.76 %, which is 2.61 % higher than the baseline model BERT.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭