Research On Patent Classifiction Method Intergrating Thematic Expression

As a carrier of innovation,patents hide a large amount of technology and effect knowledge.The realization of effect-based patent classification can enable designers to greatly extend and expand their thinking and provide help for their innovation by reading patents that use the same effect in different fields.At present,the effect-based patent classification is mainly based on concept map matching,which has the problems of poor fault tolerance and low practicability.In recent years,deep learning methods have been widely used in text classification.However,due to the problems of patent covering a wide range of fields and difficult feature extraction,the deep learning model is not effective in patent classification.Therefore,this paper proposes a concept of thematic expression and its acquisition method,and based on it,proposes a deep learning model(T-Bi GRU-ATT model)that integrates thematic expression and attention mechanism.This method can combine patent thematic expression and text features to form high-quality patent features for patent classification.The main work and innovations of this paper are as follows:(1)A concept of thematic expression is proposed for patent classification,and based on the concept of thematic expression,a patent thematic expression acquisition method integrating professional terms is proposed.Firstly,a method of extracting professional terms is proposed,then a topic model combining weighted professional terms is constructed to extract topics from patent data sets,and then a concept of "effect-topic" co-occurrence network is proposed to obtain thematic expressions of patents,which solves the problems of low quality of topic generation and poor practicability of the model.(2)A deep learning model(T-Bi GRU-ATT model)is proposed,By introducing a variety of attention mechanisms,the model integrates patent text features and thematic expressions to classify patents,effectively improving the discrimination of the model,and solving the problems of difficult extraction of patent features and ignoring context information in thematic expressions.In the experimental stage,5000 patents belonging to 25 physical effects are used to verify the effectiveness of the proposed method.The experimental results show that under the same hardware and software environment,the accuracy rate of the patent classification method proposed in this paper reaches more than 77%,the recall rate reaches more than 71%,and the F1 value reaches more than 73%.