Sentence Matching Method Research Based On Attention Mechanism Of Fusing Resnet

Nowadays,with the rapid development of digital information,sentence matching has become a research focus in natural language processing tasks.For example,in the tasks such as information retrieval,intelligent question and answer,machine translation,natural language inference,the key problem of research is still the ability to match sentence.In the early days,sentence matching mainly relied on traditional matching techniques,which required manual extraction of features and rules to learn the similarity of matching between sentences.Such features are often limited to specific data sets and were time-consuming,often ignoring the multiplicity of word matching and the deeper connection of sentence matching.With the introduction of deep learning,neural network models can automatically extract features from the original data set to learn sentence features that are useful for contextual matching,while also avoiding the significant overhead of manual feature extraction.Most of the existing deep-level interaction matching models ignore the key information within sentences,and there are certain shortcomings in how to integrate the key information within sentences and the interaction information between sentences.To address the above problems,this paper proposes a sentence matching method based on attention mechanism of fused residual network,which effectively mines the key hierarchical information by capturing the key information within sentences and combining it with the interaction information between sentences to ensure the intra-sentence characteristics and inter-sentence correlations.The main research efforts of this paper are as follows:1.A sentence matching model based on a residual connected attention mechanism is proposed in order to better combine key intra-sentence information and inter-sentence interaction information.Firstly,a BERT Chinese pre-training model is directly used on the word embedding layer to extract word vectors as input,and the word vectors are then used as input through a Bi-LSTM network to obtain local information features.Then the key information within the sentence and the interaction information between the sentences are captured using the intra-sentence self-attention mechanism and the inter-sentence interaction attention mechanism,respectively,and the features of the Bi-LSTM and attention mechanism layers are connected as the output of the current residual layer.The network layers are then stacked through the gated residual mechanism to mine deeper matching information.Finally,the sentence features are obtained through the interaction layer and the final matching score between the two sentences is obtained through a fully connected network.In this way,the matching model can take into account the key information within the sentence and effectively combine it with the interaction information between the sentences during the matching process.The gated residual mechanism can be used to better explore the main information of each network layer,reducing the number of layer of residual blocks and the size of the network structure,and improve the accuracy of sentence matching.2.In order to further capture the key information between sentences and remove redundant information,a Convolutional Neural Networks(CNN)sentence matching based on a densely connected fused self-attentive mechanism feature is proposed.Firstly,the text is represented as a vector matrix at the word embedding layer,and the trainable and untrainable word vectors are used as two channels to capture the local key information features within the sentence using one-dimensional convolutional kernels of different heights.Due to the limitations of convolution in dealing with long text features,the mutual representation information features of the convolution results within different height sentences are then obtained through a self-attentive mechanism to establish a direct link between the whole sentence and highlight the local key information.Then,a matching matrix between different words is created by combining the local key information features,and the deeper information of the matching matrix is obtained through a dense connection network(Dense Net).Finally,a fully connected network is used to obtain the final matching score between the sentences.In this way,the convolutional neural network can be used to avoid the problem of disappearing historical information as the length of the sentence increases,increasing the speed of network training,to better capture the local key information within the sentence directly,to effectively remove redundant information irrelevant to matching,and to improve the matching accuracy of the sentence.