BLSTM-PA在食品安全领域语义分析仿真研究

Simulation Research on BLSTM-PA for Semantic Analysis in Food Safety

食品安全领域的智能问答系统旨在对用户通过自然语言进行的食品安全方面的提问做出快速、简洁的反馈,其技术挑战主要在于语义分析和答案句子表示,尤其是在于如何消除问答之间的词汇差距以加强问答匹配能力,以及如何抓取准确的核心单词以增强句子表示能力。尽管基于"短语级别"和众多的注意力模型已经取得了一定的性能提升,但基于注意力的框架都没有很好的重视位置信息。针对上述问题,运用词林和word2vec相结合的方法,提出近义词-主词替换机制(将普通词映射为核心词),实现了语义表示的归一化。同时,受位置上下文提升信息检索性能的启发,假设如果问句中的一个词(称之为问题词)出现在答案句中,问题词的临近词对比偏离词应该被给与更高的权重。基于上述假设,提出了基于双向lstm模型的位置注意力机制(BLSTM-PA)。上述机制给与答案句中问题词的临近文本更高的注意力。以食品安全问答系统为语义分析验证和仿真的平台,通过在食品安全领域数据集(即FS-QA)上进行的对比实验,从MAP和MRR评价指标来看,与基于传统的注意力机制的RNN模型相比,BLSTM-PA实现了5.96%的提升,证明了BLSTM-PA模型的良好性能,同时,集成了提出的问答模型的食品安全问答系统性能也得到了显著的提升。

The intelligent question answering system in the field of food safety aims to provide quick and concise feedback on the questions of food safety in natural language.The technical challenge mainly lies in semantic analysis and answer sentence representation,especially in how to eliminate the vocabulary gap between questions and answers,how to enhance the ability of matching questions and answers,and how to capture accurate core words to enhance the ability of expressing sentences.Although some performance improvements have been achieved based on the"phrase level"and numerous attention models,the attention-based framework does not pay much attention to position information for the above problems.This paper combined Ci-Lin and word2 vec,realized the Synonym-subject mapping mechanism(mapping common words as core words),and achieved the normalization of semantic representation;At the same time,inspired by the position context to improve the information retrieval performance,we assumed that if a word in the question(we call it a question word)appears in the answer sentence,the neighboring word of the question word should be given a higher weight than the away from words.Based on this assumption,we proposed a positional attention mechanism based on the BLSTM model(BLSTM-PA).This mechanism gives higher attention to the neighboring text of the problem word in the answer sentence.We used the food safety question and answer system as a platform for semantic analysis verification and simulation.Based on the comparative experiments conducted on the food safety dataset(FS-QA),BLSTM-PA achieved an improvement of 5.96%compared with the RNN model based on the traditional attention mechanism,demonstrating the good performance of the BLSTM-PA model.At the same time,the performance of the food safety question answering system integrated with the BLSTM-PA model proposed in this paper has also been significantly improved.