Text Retrieval Based On Knowledge Graph Entity Representation And Learning To Rank

In the era of rapid development of the Internet,retrieval has become one of the main sources for people to obtain knowledge,and its fast and convenient use method can help people quickly get the information they want.As one of the most basic retrieval methods,text retrieval has been greatly improved through the traditional Boolean model,bag of words model and some classical sorting algorithms in the past decades,but there are still many problems.On the one hand,in the understanding of natural language,most of the traditional methods are based on the statistical matching of word frequency rather than semantic understanding of the answers to questions,which results in the exclusion of many answers without accurate understanding of search semantics.On the other hand,the traditional sorting model also has the deficiency of learning ability,and the learning effect depends heavily on artificial feature selection and extraction.In the past few years,deep learning models and knowledge graphs have made great progress and made important breakthroughs in various fields.Deep learning neural network,because of its excellent ability to extract features from end to end and the computational benefits brought by the superposition of layers,surmounts traditional machine learning algorithms in many aspects.Word vectors developed on deep learning model and text representation model have been widely applied to various methods of natural language.Knowledge Graph can describe the concept of entity,the relationship between entities,and form a large network diagram.The concept through the manual audit accurate and reliable.Through the knowledge graph,the knowledge about the entity can be stored accurately,and it has reliable performance in the task of question,answer,retrieval,entity connection and so on.With the help of the knowledge graph representation and deep learning model,this paper puts forward a new text retrieval model(Attention-Kernel Entity Similarity Ranking,AKESR),the new model can enhance the word multifaceted semantic understanding,and implement end-to-end feature extraction and sorting,compared with the previous study deep-learning retrieval model can use less data to achieve satisfactory results.The text retrieval model based on graph entity representation and deep learning network proposed in this paper contains the following innovations:1.Introduce multi-relational entity embedding based on knowledge graph into deep learning model.Different from the input based on word vector in traditional network,the input in this paper is based on the multi-relational entity vector trained in knowledge map.According to the existing knowledge map database and the literature in the text database,the knowledge map suitable for this task is built by ourselves.2.Multidimensional similarity matrix is used as the network input to retain relevant information of multidimensional entity matching.The retrieval problem in the data set and the entity in the article are extracted,and the similarity degree matching of corresponding positions is carried out on the dimension of the same relationship to obtain the multidimensional similarity matrix.3.An improved multi-head self-attention mechanism is introduced into the original model network.The self-attention distribution of the word vector in the retrieval problem is combined with the result of feature extraction of the entity network to extract the dependence between words in the retrieval problem.The multidimensional similarity matrix is input into the network,and the feature is extracted by combining gaussian kernel convolution with self-attention mechanism and trained by the full connection layer.Finally,Pairwise is used as the training method for loss to obtain the final training network.Compared with the traditional retrieval method and the existing deep learning retrieval algorithm from multiple perspectives,the experiment proves that this network can effectively reduce the amount of data required for training while introducing the background relationship of entities,and provide a higher accuracy rate.