Research And Application Of Adaptive Completion Based On Multidimensional Event Knowledge

With the rapid development of digitalization in the field of technology,a large amount of complex scientific and technological datas will be generated.To address this issue,knowledge triplets or event structures can be used to analyze technology big data.Knowledge has timeliness,but the knowledge graph composed of knowledge triplets cannot automatically change over time,resulting in data loss or data failure.The current knowledge completion methods only complete the missing links in the graph,cannot reflect the dynamic changes of the graph,and lacks self-adaptability.The event is a second-order structure composed of several first-order entities.The event structure can directly depict the entity and its behavior,and has a time dimension.It can automatically complete the knowledge graph through innovative elements in technological events,achieving dynamic changes in the knowledge graph.The main research contents of this thesis are as follows:(1)Light multi-source technology event extractionFirstly,a multi-source technology event extraction dataset was constructed based on laboratory technology resources.A lightweight multi-source technology event extraction method is proposed to solve the problem of multi-dimensional technology entities and fewer trigger words in multi-source technology datas,in order to automatically collect event information from multi-source technology datas.This method first discovers possible categories and entities of technological events in event sentences through technological event detection and entity recognition,thereby enhancing the model’s reading and comprehension ability.And use the information structure of the event itself to alleviate the problem of fewer trigger words in multisource scientific and technological datas.Experiments have compared this model with mainstream event extraction models in the same environment.The experimental results show that the accuracy of this model almost reaches the level of mainstream models when the number of parameters is less than other models.(2)Event based adaptive completion method for knowledge graphIn response to the shortcomings of scientific and technological knowledge graph that cannot adaptively change with datas and the error of local information dependence,an event based knowledge graph adaptive completion framework is proposed by comprehensively utilizing scientific and technological event information.Under this framework,real event information to generate completion model training.For completion tasks,technology events are used to connect a single entity with multiple dimensional entities,and technology triplets are converted into soft label prompts to provide more technological information support for the completion model.The combined knowledge driven approach based on multidimensional event information and soft label cues enables pre-training to complement the model’s adaptive technology domain knowledge.Experiments have shown that this method outperforms other methods in three comparison tasks: link prediction,triplet classification,and number of meta paths.Moreover,the completion dataset generated based on real events can improve the performance of the completion model compared to the local random information dataset.