Research On Information Extraction And Question And Answer Method For Integration Of Design,Manufacture,Operation And Maintenance

"Integration" and "synergy" technology oriented to the whole manufacturing process is an important task for the development of Chinese manufacturing industry.Among them,the integration and collaboration of design,manufacturing and operation and maintenance in the whole process of manufacturing is the core and key difficulty.However,the information systems between different departments and different businesses are isolated and scattered,and the data of each system lacks a unified standard,which makes it difficult to support interconnection and integration among businesses,resulting in difficulties in business coordination between design,manufacturing and operation and maintenance.In addition,the design,manufacturing and operation and maintenance business involves not only the process of positive step by step information transmission of each link,but also the process of information acquisition and optimization feedback.In the forward and reverse processes,each link often needs to consult the information of other links as the basis and support for the implementation of the link.Therefore,information retrieval and reference has become a necessary means of business interaction in the whole process of design,manufacturing and operation and maintenance,as well as the most direct and common form of interconnection and collaboration among businesses.Therefore,based on the original business system and data mode,the research on data standardization and information extraction technology among heterogeneous systems,and the formation of unified data description and effective information extraction methods have become the primary task to support data integration and sharing.On this basis,supporting the inter-business information retrieval and knowledge acquisition has become an important means to realize the whole process of inter-business collaboration,which is of great significance.To this end,in view of the existing information extraction methods can not solve some specific problems in the operation and maintenance business of design and manufacturing,as well as the existing problems in problem transformation,user semantic understanding,problem matching and other aspects of the existing retrieval methods,this paper takes the unstructured documents in the operation and maintenance business of design and manufacturing as data support.The research focuses on information extraction and question-and-answer technology supporting the integration of design,manufacturing,operation and maintenance.The main contents are as follows:(1)According to the characteristics of unstructured document data,a multi-modal information extraction method based on projection pursuit is designed.Firstly,based on the idea of projection pursuit method,this method makes use of the isomerism between modes to carry out modal disassembly on the three modal data of text,picture and table in the document,and then carry out information extraction on the three single modes formed by disassembly respectively.It mainly includes text extraction based on multilevel title layering and keywords,table reconstruction and extraction based on secondary traversal and image extraction based on html file format conversion,and finally realizes the acquisition of text,picture and table information in unstructured documents.To ensure the data quality and support the multimodal information set for the follow-up work.(2)According to the requirement of retrieval question answering in this subject,a retrieval question answering method based on multi-modal information set is proposed,which mainly includes the graph retrieval question answering module based on BERT,the text retrieval question answering module based on machine reading comprehension and the text answer control module based on BERT.In order to solve the problem of low accuracy of question answering with graph as the answer,a graph retrieval method based on BERT is proposed.The named entity recognition model based on Bert-Crf is used to obtain the entity in the question,and the attribute similarity retrieval model based on BERT＿SMI is used to output the corresponding picture or table combined with the chart list.Aiming at the problems of low efficiency and coarse granularity of answering with text as the answer,a two-stage text retrieval method based on machine reading comprehension is proposed.In the first stage,BM25 is used to retrieve text paragraphs,which reduces the time spent in the second stage and improves the answering efficiency.In the second stage,Bi LSTM-BERT machine reading comprehension model is used for answer extraction to refine the granularity of text answers.Aiming at the supplementary selection problem of answer modes,a BERT based text answer control module is proposed.After outputting pictures or tables in the Bert-based chart retrieval question answering module,it determines whether text answers need to be supplemented or not by classifying questions to solve the supplementary selection problem of answer modes.(3)Experimental verification and application verification.It mainly includes the experiment and analysis of the graph retrieval module based on BERT,the text retrieval module based on machine reading comprehension and the text question answering control module based on BERT,and verifies the performance and effectiveness of each module of the method in this paper.Then,the feasibility of the proposed method is verified by multi-modal information extraction and question answering tool system.