Research On Information Retrieval Technology Based On Deep Learning

With the exponential growth of various network resources,efficient information retrieval models have become more and more important.The learning to rank model is a key component of the information retrieval model.It has achieved excellent results in the era of information overload.However,the learning to rank model regards ranking as a static process,outputting all documents related to keywords at one time.In the actual information retrieval scenario,there is information transmission between the user and the retrieval model.The user gives feedback,and then the retrieval model returns a new page of content based on this feedback.The existing learning to rank model regards ranking as a one-time process,ignoring user feedback,as a result,the ranking effect of the model is not very well.In response to this type of problem,to improve the sorting effect,this thesis proposes two kinds of interactive deep learning-based information retrieval models.The main tasks include:1)Define the ranking process as a sequence decision process according to the features of reinforcement learning and the use process of retrieval model.Markov modeling is carried out on ranking process,the actions,states and rewards in the process are defined in detail,and an interactive learning to rank model is constructed.2)Long and short-term memory networks is used to model users,the policy gradient method is applied to the interactive learning to rank process.Combining the above two points,this thesis puts forward a learning to rank model based on policy gradient.In the interaction process between the user and the learning to rank model,the retrieval intention is learned according to the user's feedback,the ranking strategy is gradually improved,and the ranking effect is improved.The efficiency of the model is proved.3)Aiming at the interactive learning to rank model,the classic Actor-Critic algorithm is applied to make the ranking strategy better,and the temporal convolutional network is applied to model users.Combining the above two points,this thesis puts forward a learning to rank model based on the Actor-Critic algorithm,and experiments are designed to compare the interactive learning to rank model with some classic models to prove the efficiency of the former.