A multi-task-assisted extreme multi-label short text classification method using co-occurrence information

Abstract

The invention relates to a multi-task auxiliary extreme multi-label short text classification method using co-occurrence information. , and model the feature information as explicit model input co-occurrence information; construct multi-label text classification tasks and extreme multi-label text classification tasks related to microblog short texts; build multi-task learning task models; Bo short text data pre-trains the multi-task learning task model; fine-tunes the multi-task learning task model; quantifies the neural network output and finally outputs the multi-task prediction result. The present invention utilizes co-occurrence information to design a multi-task learning architecture, and realizes multi-label classification of large-scale short texts. The method can realize stable, accurate and real-time multi-label prediction for large-scale short text data sets under the condition of low industrial deployment cost. .

Description

technical field [0001] The invention belongs to the field of information technology, and relates to natural language processing and text classification methods, in particular to a multi-task auxiliary extreme multi-label short text classification method using co-occurrence information. Background technique [0002] With the increasing speed of text data production, data diversity, and semantic complexity, traditional multi-label text classification methods have been unable to meet daily industrial needs in terms of accuracy and real-time performance. There is an increasing demand for multi-label text classification tasks. [0003] In order to solve the above problems, the existing technologies mostly solve them by methods such as embedding, multi-classifier, tree, and deep learning. Among them, the time complexity of the embedding method is high, and the effect is extremely dependent on the clustering effect in the preprocessing process; the multi-classifier method ignores ...

AI Technical Summary

Problems solved by technology

Among them, the embedding method has high time complexity, and the effect is extremely dependent on the clustering effect in the preprocessing process; the multi-classifier method ignores the information between labels, and regards labels as independent individuals, and because each label requires training A classifier has a huge deployment cost and is difficult to effectively apply to real business scenarios; the tree method cannot solve the long-tail problem in the data set, and is large in scale, high in cost, and poor in accuracy, making it difficult to use stably in industrial scenarios

Existing deep learning methods do not optimize the long-tail problem, but simply increase the number of neurons in the output layer, resulting in their effect is usually not as good as the other three methods

Images