Research On Text Retrieval Based On Feature Representations Learning

Text retrieval plays an important role in the Natural Language Processing area,and it is also one of the basic supporting technologies in the era of big data and artificial intelligence.Text retrieval is not only an effective tool for people to obtain needed information from massive internet data,but also a vital prerequisite for the development of other intelligent systems,such as question answer dialogue system.The core problem of text retrieval is searching and determining whether a document is relevant to user query.With the aims of solving this problem,text retrieval models usually map user query and candidate document into a same vector space,and then calculate the relevance between the query and the candi-date document.It involves text feature representation and relevance calculation of text retrieval.There are two types of methods for text feature representation:standard statistics and machine learning.The relevance between queries and documents has two values:the logical value and the real value.When it is the logical value,an extra special sorting algorithm is needed to be used to sort all relevant documents.When it is the real value,the documents can be sorted in accordance with the relevance itself in the model.Under deepgoing and systematic research on the above issues in this paper,the main innovative achievements are as follows:1.This paper proposes a text retrieval model based on salient contex-tual representation.Our model locate the context which has the salient seman-tic matching with the query by using the method of sliding window,and then use the context to obtain the salient contextual representation of the document.Thus,our model can better solve the matching problem between the query and long document.Experimental results show that the model improves the accu-racy of text retrieval and recall precision.2.This paper proposes a text retrieval model based on convolution graph topology representation.Our model transforms the contextual representations of the query and the document into graph topologies and utilizes the Graph Convolution Network to generate new text representations,which contain both contextual information and global structured information.Experimental results show that our model effectively improve the accuracy of text retrieval.3.This paper proposes a text retrieval model based on interactive graph topology representation.Our model joints the bi-directional attention network and Graph Attention Network,and then transforms the contextual representa-tion into a new text representation based on the interactive graph topology rep-resentation.The new text representation contains both interaction and global structure information with different attention weights.Experimental results prove the effectiveness of the model.