Text processing method and device and computing equipment

Abstract

The invention discloses a text processing method and device and computing equipment. The method comprises the steps that a stroke vector matrix is generated based on stroke information of a single character in a word, the single character is an ideographic character, and strokes are minimum continuous stroke units of the single character; inputting the stroke vector matrix into a first neural network for processing to obtain word vectors of words; word vectors of all words in the word sequence are obtained and input into a second neural network to be processed, hidden layer representation of the word sequence is obtained, and the hidden layer representation is used for representing the distribution probability of the next word of the word sequence.

Description

technical field [0001] The present invention relates to the field of natural language processing, in particular to a text processing method, device and computing equipment. Background technique [0002] Complex machine learning language models, especially those based on neural networks, have huge model parameters, which makes the model have higher requirements for storage and operating environments; and a large part of the parameters are occupied in the model In terms of word vector representation, the vocabulary of the model can be as high as more than 200,000 words, and the dimension of each word vector can be hundreds of dimensions. At the same time, although there is a very large vocabulary, this vocabulary is generally fixed during model training. In reality, new words are invented every day. These words that are not in the vocabulary (called out-of-the-table Words or unregistered words) are very difficult to deal with models with fixed vocabularies. [0003] For exam...

AI Technical Summary

Problems solved by technology

[0005] For example, for a general model compression algorithm, one method is simply to eliminate the parameters that are not very important in the model, which is easy to cause loss in the performance of the model; the other is to use a small model to learn a large model method to obtain a small model with fewer parameters. On the one hand, this method may lose performance, and on the other hand, it will increase the training cost of the model.

[0006] For the problem of unregistered words, one method is to represent all low-frequency words with one symbol, and the model will learn a unified representation to represent all low-frequency words. This method processes all low-frequency words together, and inevitably loses every The specific meaning of each word itself makes the system performance decline; another method is to regularly update the system vocabulary, but the language model itself has a relatively high training cost and a long training time, and the cost of regular update is relatively large.

Images