Question and answer matching method based on neural ODE network

Abstract

The invention discloses a question and answer matching method based on neural ODE. The method comprises the following steps: establishing an end-to-end question and answer matching model based on a neural ODE network, which comprises a word vector coding module, a sentence vector coding model block, a feature extraction module, an information interaction module, an information fusion module and a prediction module. The word vector coding module is used for coding each word in a sentence into a word embedding representation; the sentence vector coding module uses cnn to obtain a sentence vector matrix representation, namely shallow semantic information; the feature extraction module uses a neural ODE-based cnn network component to extract a deep semantic information representation from the sentence vector matrix; the information interaction module constructs an interaction semantic feature representation with attention information by using an attention mechanism; the information fusion module fuses the shallow semantic information, the deep semantic information and the interactive semantic feature representation to obtain a global semantic information representation; and the prediction module calculates the global semantic information representation through a softmax function and outputs the relationship of predicted sentences. According to the method, the technical problems that a model is difficult to train along with the deepening of the network structure and the parameter quantity is too large along with the deepening of the structure of a residual neural network are solved.

Description

technical field [0001] The invention belongs to the text matching technology of natural language processing field; In particular, relate to a kind of question-answer matching method based on neural ODE network. Background technique [0002] Question-answer matching is a general term for classification tasks in the field of natural language processing, and has always been the core research direction of researchers. In the text matching task, the model accepts two sentences as input into the model, and outputs a category or a scalar to reflect the relationship between the two sentences. These tasks can be regarded as special forms of question-answer matching. Therefore, the application scenarios of question-answer matching are very extensive, such as public opinion analysis, spam identification, etc. By learning a large amount of data, the model can finally accurately classify text and extract relationships. This is also the standard process for the implementation of natural...

AI Technical Summary

Problems solved by technology

[0022] The technical problem to be solved by the present invention is to overcome the problem of too many residual structure parameters in existing feature extraction and provide a question-answer matching method based on neural ODE network, which uses pre-trained Glove word vectors to encode sentences , to enhance the text representation ability of the neural network, while the feature encoding module extracts higher-dimensional features and improves the performance of the text model of the interactive layer:

Images