The Research On Patent Classification Based On Deep Learning

Patent texts are the carrier of science and technology which contain rich knowledge of background,technology,function and effect.Using the function-effect-patent search method,one can learn from the principles and methods of patents in different fields,thus breaking the inertia of thinking,providing product designers with similar patents for reference.At present,the corresponding relationship between patent and effect is mainly dominated by the concept graph matching method,which has the problems of poor matching tolerance and poor practicability.In recent years,using deep learning models to solve the problem of text classification has drawn much attention,and it has been proved to have a great advantage in feature extraction and text representation.There is a problem in the recurrent neural network that the gradient disappears and the semantic information is lost.However the attention mechanism can allocate more attention to the keys of the text,highlighting its importance.This paper proposes a classification algorithm which contains many bi-directional LSTM models based on attention mechanism(Bi_LSTM_ATT).The algorithm respectively uses the abstract,the claim specification,the invention content,the detailed description as the original input to the Bi_LSTM_ATT model,and the four parts features learned are integrated with different weights as the patent text features.The functions are used as the original input of the LSTM model to obtain functional features.Then,the patent text features and functional features are integrated as the overall feature for the patent effect classification.The final output is the type of effect that the patent belongs to.In this paper,a large number of experiments have been conducted on the patent corpora in the field of mechanical physics.Experiments show that the accuracy can reach above 70%,indicating that the algorithm has a certain degree of validity for determining the effect of patents.Validated by engineering application,this algorithm can effectively expand the designer's innovative thinking and promote the emergence of new programs.