Text classification algorithm based on bidirectional dynamic routing with label constraint

Abstract

The invention provides a text classification algorithm based on bidirectional dynamic routing with label constraints, which comprises the following steps of: firstly, fusing semantic information and word order information into the same vector through a self-attention mechanism and relative position information coding; then introducing a bidirectional dynamic routing mechanism, and simultaneously considering the conversion amount information from the bottom layer capsule to the high layer capsule and the amount information of the bottom layer capsule required by the high layer capsule; finally, a label embedding constraint method is provided to replace model reconstruction in a capsule network, and the potential unreasonable distribution condition of text capsules is solved; The relative position coding is introduced, the problem that a dynamic routing algorithm neglects the word order of words is solved, and the bidirectional dynamic routing algorithm can give consideration to the problem that the same word has different importance in different types of texts; the provided label embedding constraint method can solve the problem of potential unreasonable distribution of text capsules.

Description

technical field [0001] The invention relates to the technical field of text classification, and mainly relates to a text classification algorithm based on bidirectional dynamic routing with label constraints. Background technique [0002] Because text classification plays an important role in natural language, it has become a relatively active research field in recent years. It has achieved great success in inventions such as document retrieval, web search, and email filtering. Moreover, many natural language processing Tasks can eventually be transformed into text classification tasks, such as sentiment analysis and intent detection. The text classification task is to annotate a given sequence of text with one (or more) class labels. How to construct an effective text feature representation is the most critical step in the text classification task. Existing methods for text classification can be mainly divided into two categories: 1) Traditional methods based on bag-of-wo...

AI Technical Summary

Problems solved by technology

[0004] Purpose of the invention: In order to solve the problems in the above-mentioned background technology, the present invention proposes a text classification algorithm based on two-way dynamic routing with label constraints to solve the problem of potential unreasonable distribution of capsule networks

Images