Gating cycle acquisition method based on label perception

Abstract

The invention discloses a label perception-based gating cycle acquisition method. The method comprises the following steps of S1, a mixed attention layer: calculating an attention weight between a classification prediction vector of a previous time step and text word embedding of a current time step by the mixed attention layer; s2, a text feature extraction layer: extracting text features by using a convolutional neural network; and S3, a loop acquisition layer: a gating loop unit performs classification prediction according to the classification result of the previous time step and the text features of the current time step. According to the method, the classification result has a higher F1 value (F1score) and a lower Hamming-loss (Hamming-loss). Therefore, the multi-time rereading mechanism of the model can improve the multi-label classification performance. According to the method, for the defects of an existing multi-label classification algorithm, a multi-label classification task is modeled into a cyclic decision process, correlation between labels is considered, and the classification performance of the model is improved.

Description

technical field [0001] The invention belongs to the technical field of natural language processing, and in particular relates to a tag-aware-based gated loop acquisition method. Background technique [0002] Multi-label classification is an important research direction in natural language processing. For the multi-label classification problem, given a document, there may be one or more tags matching it, and there is a correlation between the tags, so it is more challenging than the single-label classification task. Multi-label classification can be used in practical scenarios such as sentiment analysis, information retrieval, and recommendation systems. [0003] There are two main solutions to multi-label classification tasks: problem transformation and algorithm adaptation. Problem transformation methods convert multi-label classification problems into a set of single-label classification problems, which are then processed by single-label classification algorithms. The m...

AI Technical Summary

Problems solved by technology

The binary association algorithm trains a classifier separately for each label, so it cannot take advantage of the correlation between labels

[0005] The above are some related works of multi-label classification, through problem transformation, algorithm adaptation and neural network to solve multi-label classification problems, but these methods do not take into account the semantic information between labels and text

Disadvantages of existing technologies: existing statistical learning multi-label classification algorithms such as binary association algorithm, label power set algorithm, classifier chain algorithm, etc. only consider the first-order or second-order correlation between labels, or consider high-order Calculations are difficult when correlation

Existing deep learning models such as deep convolutional neural networks and CNN-RNN models do not take into account the semantic information between labels and text

Images