A Pairwise Deep Learning-to-Rank Method

In today's Internet age of information explosion,the importance of information retrieval is self-evident.We are conducting information retrieval almost every moment.For example,query web pages on search engines,watch videos on video websites,and look for papers in the thesis library.A good search engine can directly retrieve the results we want,thereby avoiding people spending a lot of time looking for the information they want.For search engines,the most important part is the ranking model.The ranking model sorts the results in order and puts the most relevant information first.This paper aims to propose a new learning ranking model that can learn the optimal ranking model.We first use the results of traditional unsupervised models to build higher-order features.Then constructed a pairwise learning ranking method.Then we build a Rank Net model based on the previous higher-order features,and then builds a Pair Text CNN model based on the Text CNN method,which can automatically obtain implicit patterns and features from text.Finally,an outer layer deeper neural network with a learnable parameter is used to combine the two to obtain the final model Mix Net.Comparing these three models with other researchers' models on the benchmark public data set,it is found that the effect is better than most models,which shows the advantages of the Pairwise-based model and the combination of explicit pattern matching and implicit pattern matching on the effect.This paper also discusses the effect of different initialization methods and processing methods of the word-embedding method on the results of the model.The size of the dataset used in this paper is relatively small,and the effect of the deep model needs to be tested on a larger dataset with original corpus.Finally,based on the research results,we summarizes the contributions and findings made in this paper.