Research On The Judgment Method Of Semantic Similarity Of Short Sentences Based On Matching-aggregation

As a basic task of natural language processing,semantic similarity discrimination of short sentences plays an extremely important role in downstream tasks such as data mining,information retrieval and machine translation.However,in the current semantic similarity model based on the matching aggregation framework,only the semantic information of a single feature space is considered in the sequence alignment process,and the global information is not fully utilized.In addition,the single-step prediction method that transforms the sequence into a fixed dimension vector will also lead to the loss of important information in the sequence.Therefore,solving the above problems has a positive effect on the improvement of model performance.Aiming at the above two problems,this paper proposes an enhanced sequence alignment method based on mixed global information and a two-way multi-step iterative prediction method based on GRU.The models are designed and experimentally verified for the two parts.The thesis work was supported by the National Key R&D Program Project "Internal and External Connected Trial Execution and Litigation Services Collaborative Support Technology Research"(2018YFC0831300).The main work is as follows:(1)An enhanced sequence alignment method based on mixed global information is proposed and verified by a design model.The method is composed of three parts.The internal alignment method based on Transformer is used to use the global information of the sequence itself,the enhanced interactive alignment based on multi-head attention is used to use the global information of the external sequence,and the global alignment based on Bi-LSTM/Transformer The method is used to fuse the global information inside and outside the sequence.In addition,the factorization machine is used for the second-order interactive enhancement and dimensionality reduction of each dimension in a single component in the sequence.Based on the above content,a semantic similarity model based on mixed global information is designed.The experimental results show that the accuracy rate of the standard model is 89.89%,which surpasses the current mainstream model of the same kind by nearly 0.49%,and the accuracy rate of the non-standard model can reach 89.97%.(2)A two-way multi-step iterative prediction method based on GRU is proposed and verified by designing a model.This method first records the context information of the enhanced alignment feature through Bi-GRU,and then uses GRU to perform iterative prediction from two directions.The initial state of each GRU is a fixed compression vector of a sequence,and the input of each iteration time step is synthesized by calculating the hidden state of the previous step and the complete sequence of another sequence through the attention mechanism.Therefore,the input of each step includes the adjustment of the initial fixed vector and the use of the complete information in another sequence.Integrating the two-way information,the complete information of the two sequences can be used multiple times,and the final prediction result is jointly determined by the prediction results of multiple iteration steps.Based on the above content,a multi-step iterative discrimination model corresponding to semantic similarity based on GRU is designed.Experimental results show that the accuracy rate of the standard model is 89.87%,which surpasses the current mainstream single-step iterative network of the same kind by nearly 0.47%,and the accuracy rate of the non-standard model can reach 89.89%.The above two models are superior to the BERT-Base pre-training model in terms of resource consumption and model speed.At the same time,the model performance is quite competitive and is more suitable for lightweight industrial deployment.