查询意图自动分类的新方法探讨

[Objective] In this paper, we achieve automatic identification of query intent based on Sogou query log data. [Methods] (1) We map the ODP subject class system to Rose's intention class system, using the heuristic and ontology matching method to form annotation rules, and automatically mark the query; (2) We use the LTP tool to extract the natural language level features of the query, including: the characteristics of the word segmentation, the part of speech and the syntactic dependency between the words, and the statistical and user behavior of the query; (3) Based on the marked query log data set and the extracted features, the automatic identification of the query intention is realized by using GBDT. [Results] (1) The scale of the data set marked by the automatic labeling rule proposed in this paper is close to that of the manually labeled data set; (2) The efficiency of the classifier using the feature set training proposed in this paper is superior to the efficiency of the syntactic dependency feature between the unused words; (3) Using the GBDT classification model to obtain the results of the accuracy, precision, recall and F1 values were 0.75,0.76,0.93 and 0.84. [Limitations] This article only uses Sogou query log data, but also need to be further validated on other data sets; the use of natural language processing tools can not fully extract the query syntax and semantic features. [Conclusions] In this paper, we can quickly obtain a large number of trained training data sets; fully extract the characteristics of natural language level can improve the recognition of the intention of the query; GBDT as an assembly type of machine learning model in the efficiency of the identification efficiency is better than the linear classification model (such as Logical Regression and Support Vector Machine).