Research On Short Text Similarity Calculation Method Based On Siamese Structural Model

Short text similarity calculation method based on deep learning has always been an important research in the field of natural language processing.However,the existing deep learning models in short text similarity calculation tasks have some shortcomings,such as insufficient text semantic extraction,poor context combining ability,and low ability to understand polysemy.Nowadays,the mainstream method is to improve the accuracy of short text similarity calculation through more complex or parameterized models.However,this type of method consumes a lot of computer power,increases model training time,and reduces the universality of the model.In response to the existing problems in short text similarity calculation methods based on deep learning,this article has done the following work:(1)A Siamese short text similarity calculation model based on BERT was designed to address the problems of traditional short text similarity calculation models.In this method,the overall model is designed as a Siamese structure.In the word vector embedding part,the Siamese-BERT model is used for vector embedding,and the Bi-LSTM network is used for feature extraction.At the same time,the self attention mechanism is added to design the Siamese-BERT-Bi-LSTM-att model.After model optimization,the comparison and ablation experiments are conducted on the standard data set.The experiment proves that this method enhances the model’s semantic understanding ability of short text,improves the accuracy of short text similarity calculation and F1 value.(2)Although the accuracy of the Siamese-BERT-Bi-LSTM-att model has been improved to some extent,it was found in the experimental process that this method sacrifices too much computational power compared to traditional models,resulting in a significant increase in training time.In response to this issue,this article designs a Siamese-ALBERT-Bi-GRU-att model,which effectively reduces the processing time of the word embedding layer and feature extraction layer.And comparative experiments and ablation experiments were conducted on the standard dataset MSRP and the standard dataset QQP.The experiments showed that the Siamese-ALBERT-Bi-GRU-att model not only improved the model’s ability to understand short text semantics,but also shortened the training time of this type of improved method model,increasing its universality.This article designs two short text similarity calculation models,which combine the advantages of dynamic word vector training models and feature extraction networks to increase the overall similarity calculation accuracy of the model.At the same time,a new model is designed to reduce training time and enhance the universality of the model without affecting the accuracy of short text similarity calculation.