基于意图—槽位注意机制的医疗咨询意图理解与实体抽取算法

自然语言理解作为医疗对话中的关键组成部分，包含意图识别和槽位填充两个重要的子任务。为建立意图和槽位的相互促进关系，实现语义层次上的建模，提出了基于意图—槽位注意机制的医疗咨询意图理解与实体抽取算法。首先，收集医疗信息网站上用户的医疗健康提问文本，基于医学知识归纳总结了24类医疗意图和5种槽位，构建了中文医疗健康咨询数据集（CMISD-UQS）；然后，引入槽位选通机制来建模意图和槽位向量之间的显式关系，设计了意图—槽位注意机制层，构建了意图上下文信息以意图标签向量方式嵌入到槽位的方式。在公共数据集ATIS和SNIPS上与八种代表性算法的对比实验结果表明，所提算法优于所比较的八种算法；在CMISD-UQS数据集上的测试结果表明，所提算法的医疗意图识别准确率、语义槽填充F1值、句子级语义框架准确率分别为78.1%、94.9%和73.2%，均优于其他对比算法。

Intent understanding and entity extraction algorithm for medical consultation based on intent-slot attention mechanism

Natural language understanding is a key component in medical dialogue, contains two important subtasks of intent classification and slot filling. In order to establish the mutual facilitation relationship between intents and slots and to achieve modeling at the semantic level, this paper proposed an intent understanding and entity extraction algorithm for medical consultation based on the intent-slot attention mechanism. Firstly, this paper collected users'' health question texts on medical information websites, summarized 24 types of medical intents and 5 types of slots based on medical knowledge, and constructed the Chinese medical and health consultation dataset(CMISD-UQS). Then, this paper introduced a slot selection pass-through mechanism to model the explicit relationship between intent and slot vectors, and designed an intent-slot attention mechanism layer to construct a method to embed intent context information into slots as intent label vectors. Finally, the results of comparative experiments with eight representative algorithms on public datasets ATIS and SNIPS show that the proposed algorithm is better than other algorithms. The test results on CMISD-UQS dataset show that the accuracy of medical intent classification, semantic slot filling F1 value, and sentence-level semantic frame accuracy of the proposed algorithm reaching 78.1%, 94.9%, and 73.2%, respectively, which are better than other comparison algorithms.