Research And Implementation Of Representation Learning Algorithm Based On Deep Neural Network

Linked data in physical,biological,social,and information systems can be abstracted into a single-relational graph or multi-relational graph according to the type of relationship.At the same time,due to the graph representation and computing power,the processing of graph-structured data has become one of the current hotspots in academia and industry.Representation Learning technology is the important research directions in Artificial Intelligence(AI),and can effectively represent and calculate graph based on the topology information and semantic information contained in the graph,which plays a pivotal role in related fields such as speech recognition and signal processing,natural language processing and so on.The goal of representation learning is to use machine learning methods and utilize the knowledge structure in the network to represent nodes and edges with low-dimensional and real-valued vectors,thereby realizing the extraction and representation of semantic information of entities and relationships.Based on above,through the depth analysis of the modeling mechanisms of related works,it is determined that this thesis conducts research around representation learning modeling based on semantic mapping,and find that the existing methods have following defects: First,the current mainstream semantic mapping methods underutilizes the information in the graph,and it is difficult to extract richer association features between entities,resulting in lower performance of the representation learning algorithm.Second,the current mainstream algorithm based on semantic mapping mainly focuses on how to design novel vector semantic composition method,ignoring the summary of the impact of different semantic composition methods on the performance of the algorithm,which restricts the development of graph representation learning.This thesis studies the above problems and proposes corresponding solutions.The main contributions are as follows:1.In view of the shortcomings of mainstream semantic composition mathods that are mostly based on linearity and insufficient information utilization,this thesis designed a parallel convolutional semantic composition method,and proposed a novel representation learning based on convolutional neural networks.The algorithm uses the powerful feature extraction capabilities of deep neural networks to fully capture the association of any two entities,alleviating the problem of insufficient use of information in the existing work.The experimental results show that the performance of the proposed algorithm is better than the current mainstream related work.2.Summarize the similarities and differences of the current mainstream algorithms in the design of semantic composition methods,and propose a graph representation learning model based on vector semantic composition and mapping.This model not only maintains the advantages of simplicity and efficiency of the TransE,but also is compatible with the current mainstream semantic mapping model design ideas.It also differs from existing models in design ideas,and provides a research platform for systematic research on different semantic mapping methods.3.The existing vector semantic composition schemes are summarized into three categories,and 12 variants of schemes are proposed.Based on the model proposed by 2,a systematic experimental performance comparison analysis is performed,and some regularities about semantic mapping modeling are obtained: This model can achieve better performance than current mainstream models on a variety of data sets with simple modeling assumptions and low model complexity.The experimental results help to understand the representation learning problem of graphs from a new perspective,and promote further in-depth research and application in the field of learning.