| Predicate mapping,which is a key module in knowledge base question answering systems,determines the search space of knowledge base and accuracy of question answering.Many predicate mapping modules in knowledge base question answering systems do not pay attention to the problems of nested relations and reverse questions appearing generally in real scenarios,which make the systems not friendly and reliable in the production environment.Besides,building a semantic-aware and expression-sensitive predicate mapping module is another huge challenge.Based on the above problems,the syntax enhanced predicate mapping model and the question-template reverse predicate mapping model are designed for the predicate mapping module of the knowledge base question answering systems.The syntax enhanced predicate mapping model uses long short-term memory cells and the designed Multi-Input long shortterm memory units to dynamically construct the dependent syntactic structure network of questions,extract the semantic features of questions and map them to the predicate paths of the knowledge base.The question-template based reverse predicate mapping model is proposed to create question templates from triples in the knowledge base by the TransE with the seq2 seq model.It maps questions to the question templates to complete the predicate mapping in reverse.The question-template reverse predicate mapping model solves the nesting problem and reverse questions by constructing question templates.At the same time,in order to verify the validity of the syntax enhanced predicate mapping model and the question-template reverse predicate mapping model,PM2KBQA,a knowledge base question answering system for the open domain,takes the two models as the predicate mapping module.Besides,PM2KBQA is applied to the manufacturing field to create a manufacturing knowledge base question answering system MIKBQA.The syntax enhanced predicate mapping model,question-template based reversed predicate mapping model,and their PM2KBQA are tested on the SimpleQuestion and NLPCC ICCPOL 2016 KBQA datasets.Experiments prove the syntax enhanced predicate mapping model and question-template based reversed predicate mapping model achieve higher accuracy than the state-of-the-art model in several cases.More importantly,the two models help our system beat the state-of-the-art knowledge base question answering system in most cases.Besides,MIKBQA is applied to the manufacturing datasets and gets good test results. |
