Research And Application Of Question Answering Systems Based On Knowledge Graph For Dangerous Chemical Goods

As one of three most risky industries in the world,chemical industry can lead to serious safety accidents,economic losses and environmental pollution in all stages of chemical dangerous goods processing.In recent years,accidents caused by dangerous chemical goods are very common,during process of production,storage,transportation,use and handling,which may cause injury to human body,animals,plants,and environment,timely and correct handling will reduce chemical pollution.Therefore,the dangerous goods question answering systems in chemical industry can provide professional knowledge and advice.With the development of deep learning technology,algorithms of deep learning are applied to knowledge graph in question answering systems,which can better perform natural language processing,deeper semantic understanding and more efficient automation.In this thesis,focus on the research on question answering systems based on knowledge graph of chemical dangerous goods,which mainly includes data collection and processing,named entity recognition in chemical industry,relationship extraction,question parsing and system construction.The specific research contents are as follows:(1)According to Chemical Knowledge Graph Schema,crawler technology is used to extract text data in chemical field from web pages,and combined with the North Chemical Industry Public Dataset(including chemical dangerous goods cards,accidents caused by chemical dangerous goods in 2000-2019),and is used the expanded database to analyze chemical knowledge graph.For some isolated chemical dangerous goods as some nodes,the relationship between nodes is added manually,the training data set includes 9837 sentences and 226256 words in total.(2)The BERT model is used as language model for named entity recognition tasks in chemical industry,and an improved Transformer is used in encoding stage,and also the relative position encoding with direction and distance perception is used to calculate the attention score,which enhances the model’s expressive ability and captures associative semantics in sequences,which helps to extract nested information in chemical entities.Using BERT-Attention-CNN to achieve relationship extraction,with combining sentence-level information and named entity information in sentences,which can improve the accuracy and robustness of relationship extraction.(3)Design and implement of question answering system for chemical dangerous goods knowledge graph.The main function of the system is question and answer related to chemical dangerous goods,the method based on knowledge graph is selected to realize the analysis of the question sentence,and the parsed question sentence is matched with the pre-defined query template,and the cypher query statement is used to link to the knowledge graph to realize the question-answering function,which can save time for users to obtain information.