Research On Text Semantic Matching Method Based On Sequence Model

Along with the coming of Internet times,the information that exists on the Internet has increased at an unprecedented speed,and the ways of acquiring knowledge have also undergone profound changes.Faced with such enormous amount of information resources,under no reasonable management and effective matching functions,massive amounts of information will become useless garbage,which takes space and consumes energy,thus helpless to people's lives.Language is an advanced product of agglomerating wisdom in the course of human development,with its abstractness and complexity,text-matching is very difficult.Therefore,designing an efficient and accurate text-matching system has a wide range of practical significance.Based on the application of recommendation system,this thesis studies the application of semantic representation in text-matching task.For the recommendation algorithm,the quality of the recommendation mainly depends on the quality of the feature extracted from the user's content data.It is not suitable for mining the user's potential interest,so the recommended content is limited.The deep learning method usually used in text representation is more mechanized and less adaptive.For the difficulty of matching due to the inconsistency of the amount of information,as well as semantic loss and information sparsity.In order to solve the above problems,this author study and develop some new method as follows:1.In this thesis,a new text representation method is proposed,in which the TF-IDF algorithm is introduced into the vector representations of texts to weight calculation.The deep semantics represented by the word vector are merged with the shallow semantics of TF-IDF mining to enhance the semantic of keywords and weaken the meaning of irrelevant words in the matching-text.The combination of deep learning method and statistical learning method can make the expression of text semantic more smooth,more flexible and accurate in matching,and facilitate to improve the accuracy of the recommendation algorithm.2.In order to exploit the potential interest of users,a combination of content-based recommendation method and association rules-based recommendation is proposed.The Conditional Random Fields(CRF)algorithm has the ability to take into account transitional probabilities among contextual tags,capture overlapping features and avoid tag offsets,so that user attributes can be recognized and mined for matching text.It is possible to obtain the potential interest and derivative requirements of the user.At the same time,the users are recommended for similarity and correlation,saving the time and cost.Experimental results show that the proposed algorithm obtains higher matching accuracy than the typical feature extraction algorithms and the recommendation is more targeted and more effective.3.In order to improve the text-matching model of neural network based on collaborative filtering recommendation,a method to optimize the weights of Long Short-Term Memory(LSTM)neural network is proposed.The probability distribution of text semantics is calculated through the attention mechanism,and a mean-pooling layer is added to roughly form a global feature perception and enhance the correspondence between the features and its importance.An inner attention-based LSTM model constructed on a neural network can make full use of commodity data and focus on mining key information of less popular products.To some extent,the information loss caused by the unequal quantity informations amount of the matching text is avoided,and the effect of collaborative filtering is improved.Experimental results demonstrate that the proposed algorithm has a good effect on the collaborative filtering recommendation.4.Based on the research of text matching algorithm,a recommendation system is designed and implemented.For the user-entered search statement,the system can automatically analyze and visualize the results.