Text Matching Based On Deep Neural Network

With the rapid development of the digital society,people's requirements in the fields of artificial intelligence such as information retrieval,question answering and dialogue system have begun to appear and intelligent matching algorithms are needed to meet the high demands of users.To solve this problem,natural language processing technology emerges as the times require,providing users with efficient information retrieval services and comfortable human-computer interaction experiences.Text matching is a popular research field in natural language processing technology.The prpblems of dimensionality disaster and data sparse in text representation have become the bottleneck of natural language processing.In recent years,due to the rapid developments of deep learning concerning word embeddings,text matching based on deep neural network has gradually become a new research hotspot.In this paper,a lot of deep text matching models are studied.We propose two text matching algorithms based on multi-semantic document representation and attention mechanism separately.The main work includes:Firstly,this thesis studies word representation methods based on distributed hypothesis,namely the distributed representation(word embedding),including neural network-based word representation,matrix-based word representation and cluster-based word representation.Secondly,this thesis proposes a new deep multiple view sentence representation model(DMVSR)for text matching.DMVSR model can capture long-distance semantic dependencies between texts and obtain multi-granular semantic information.Each sentence representation is generated by a bi-directional long-short-term-memory.The feature extraction and feature selecton are captured by 1D Convolution-neural-network and max-pooling for each sentence representation.The final text matching score is produced by aggregating interactions.The experimental results show that the DMVSR model is better than the traditional text matching algorithm based on single-semantic document representation and multi-semantic document representation.Thirdly,this thesis proposes an attention-hierarchical convolutional neural network model(AHCNbN)for text matching.AHCNN model can feature high-dimensional text semantic information and sort original text semantic unit with selective learning.Hierarchical convolutional neural network(2D-CNN and 3D-CNN)extracts the high dimension features from text interaction matrix and then we perform the pooling operation.Finally,we use attention network to obtain the final text matching score.At the same time,we propose another text matching model named ACNN,which simplifies AHCNN into a single-layer 2D-CNN.The experimental results show both AHCNN model and ACNN model have a better performance than the above deep text matching models.