Text Matching Technology Based On Deep Learning

Text matching is a key task in natural language processing technology,which is used to determine the similarity or matching degree between two texts.It has a wide range of application scenarios,such as semantic reasoning,text classification,information retrieval,question answering systems,etc.At present,there are the following problems in text matching:firstly,the pre trained model lacks the ability to represent text semantics,and the word vectors generated by the pre trained model have anisotropy issues,making it difficult to directly apply them to text semantic matching.Second,the ability to extract text Semantic information is insufficient.At present,the common practice is to directly use the pre training model coding,without considering the context between the words that constitute the current sentence,and there is a problem of insufficient ability to extract Semantic information.Third,it ignores the text syntactic structure information,meaning,syntax and structure,which constitute the three elements of the text.At present,the common practice is to match based on the text’s Semantic information,ignoring the text’s syntactic dependency structure information.The thesis proposes an improved text matching model to address the shortcomings of existing research on text matching tasks.Firstly,in response to the phenomenon of "collapse" in the semantic representation of sentences in text pre training models,a text representation transfer model based on contrastive learning is proposed.Drawing inspiration from the Sim CSE model,a dropout mechanism is adopted to sample positive and negative samples.Considering that amplification within a batch can lead to false negative samples and the "collapse" of the representation vector.A new contrastive learning framework was proposed to alleviate the impact of sampling bias in BSim CSE.Firstly,sample matching weights were used to eliminate the impact of false negative samples;Secondly,display data augmentation(EDA,back translation,text paraphrasing)methods are used to maximize the expansion of the negative sample dataset.By using Fine tune on unsupervised corpora in the target domain,the text representation generated by the model is more suitable for the data distribution of downstream tasks,improving the semantic representation ability of the text.Through experimental results,it has been proven that BSim CSE has significant performance improvement compared to other model models under the same settings in sentence semantic matching(STS)tasks.Secondly,aiming at the problem that text matching model only considers text semantic matching and ignores syntax and structural characteristics,a hybrid text matching model based on context Semantic information and graph convolution neural network is proposed.A method based on multi head self attention mechanism for text semantic encoding is proposed.In text syntactic dependency encoding,graph neural networks are used to obtain information about text syntactic dependency structures.In terms of feature fusion,in order to comprehensively and effectively apply semantic features and syntactic dependency features,the thesis designs a forward network that combines residual connections and layer normalization to efficiently fuse semantic and syntactic dependency features.The model proposed in this article achieved the best results on all three datasets.At the same time,we also conducted ablation experiments on the model and evaluated the impact of different modules of the DSSGM model on the final results.Finally,performance analysis experiments were conducted on the model,and the DSSGM model also achieved a relatively ideal effect.