Method, model building method and device for automatically adding punctuation points to real-time text

Abstract

Embodiments of the present invention provide a method for automatically adding punctuation to real-time text, a model construction method, and a device, wherein the model construction method includes: constructing a real-time text automatic punctuation model including a labeling model and a decision model based on reinforcement learning, and converting the non-punctuation The real-time text stream is used as the input stream to input the labeling model. For the characters currently input into the labeling model in the input stream, the labeling model outputs the labeling result of whether the current character is punctuated. The decision model obtains the current character from the input stream, according to the current hidden layer of the labeling model The state evaluates the labeling results, controls whether to write the labeling results to the output stream; trains the labeling model to convergence on the general long-sequence punctuation dataset, and trains the decision model on a dataset containing a preset number of character-punctuation pairs Train until convergence, and obtain the trained real-time text automatic punctuation model. It can automatically and accurately add punctuation to non-punctuation text streams in real-time scenarios.

Description

technical field [0001] The invention relates to the technical field of information processing, in particular to a method for automatically adding punctuation points to real-time text, a model building method and a device. Background technique [0002] With the further development of society, continuous speech recognition (Continuous Speech Recognition, referred to as CSR) technology is more and more widely used, such as intelligent customer service and simultaneous translation and so on. However, the text obtained by the current CSR technology has no punctuation marks, which has brought great limitations to the further use of CSR output results and the development of downstream technology of speech recognition. Therefore, it is necessary to automatically add punctuation marks to the output results of CSR, so as to improve the readability of the results and make CSR technology obtain a broader application prospect. [0003] At present, the existing methods for adding punctua...

AI Technical Summary

Problems solved by technology

[0004] However, the deficiency of the existing method for adding punctuation to the non-punctuation text output by CSR technology mainly lies in that: in real-time scenarios such as speeches and conferences, the text to be punctuated cannot be obtained at one time, but needs to be obtained as the speaker moves. Speech is acquired word by word, that is, real-time text; if the process of outputting text containing correct punctuation recognition results has a large delay compared with the speaker's speech, it will prevent the audience from corresponding the sound with the text in time, and unable to understand the semantic meaning of the speaker in real time , causing greater inconvenience

The reason why the existing methods of adding punctuation to unpunctuated text output by CSR technology cannot respond in time in real-time scenarios is that the training method of the current method determines that the current method is more suitable for adding punctuation to long texts. It takes a certain amount of time to obtain long text, resulting in a delay; at the same time, if you do not use long text but directly add punctuation to short text, the quality of the added result will be significantly reduced

Images