Granted Patent B2

US 12,547,834 · App. 18/469,461 · Granted Feb 10, 2026

Inventors: Yuxiang Zou (Beijing, CN); Zejun Ma (Beijing, CN)

Assignee: BEIJING YOUZHUJU NETWORK TECHNOLOGY CO., LTD.

G06F40/284G06F40/263G10L13/10

Abstract

Claims (73)

1 . A method for text analysis, comprising:

acquiring a text comprising words originating from different languages;

performing token conversion on every word in the text in a language-independent manner, to obtain a token sequence; and

performing feature extraction on the token sequence, and processing a target task based on an extracted feature, to determine an analysis result for the text.

2 . The method according to claim 1 , wherein the performing token conversion on every word in the text in the language-independent manner to obtain the token sequence comprises:

acquiring a multilingual vocabulary, wherein the multilingual vocabulary is used to represent a mapping relationship between words in a plurality of languages and tokens;

querying, in the multilingual vocabulary, for tokens corresponding to the words in the text, to determine the token sequence.

3 . The method according to claim 2 , wherein the querying, in the multilingual vocabulary, for tokens corresponding to the words in the text, to determine the token sequence comprises:

querying, in the multilingual vocabulary, for tokens corresponding to the words in the text; and

in a case that there is no query result for a word, dividing the word into sub-words, and querying, in the multilingual vocabulary, for tokens corresponding to the sub-words, to determine the token sequence.

4 . The method according to claim 1 , wherein the performing feature extraction on the token sequence, and processing the target task based on the extracted feature, to determine an analysis result for the text comprises:

acquiring a target analysis model, wherein the target analysis model is obtained by analyzing a sample token sequence in at least one preset language, and the target analysis model comprises a feature extraction unit and at least one type of target task processing unit;

inputting the token sequence into the feature extraction unit for feature extraction, to obtain the extracted feature; and

inputting the extracted feature into the at least one type of target task processing unit respectively, to determine the analysis result for the text.

5 . The method according to claim 4 , wherein the acquiring the target analysis model comprises:

acquiring a preset analysis model, wherein the preset analysis model comprises a pre-trained feature extraction unit and at least one preset task processing unit;

acquiring the sample token sequence in the at least one preset language and a task label corresponding to the sample token sequence;

inputting the sample token sequence into the preset analysis model to determine a prediction result corresponding to each of preset task processing units; and

calculating a loss function based on the prediction result and the task label to update parameters of the pre-trained feature extraction unit and parameters of the at least one preset task processing unit, to determine the target analysis model.

6 . The method according to claim 5 , wherein the calculating the loss function based on the prediction result and the task label to update parameters of the pre-trained feature extraction unit and parameters of the at least one preset task processing unit, to determine the target analysis model comprises:

for each of target tasks, acquiring a prediction category corresponding to the target task in the prediction result;

calculating a first loss function based on the prediction category and a target category corresponding to a preset task, to determine a first loss;

acquiring candidate prediction categories in the target task, wherein the prediction category is one of the candidate prediction categories;

calculating a second loss function based on the candidate prediction categories to determine a second loss; and

updating the parameters of the pre-trained feature extraction unit and the parameters of the at least one preset task processing unit based on the first loss and the second loss, to determine the target analysis model.

7 . The method according to claim 6 , wherein the calculating the second loss function based on the candidate prediction categories to determine the second loss comprises:

counting a quantity of each of candidate prediction categories;

determining a loss value of each candidate prediction category based on the quantity, wherein the loss value is negatively correlated with the quantity; and

determining the second loss based on loss values of all the candidate prediction categories.

8 . The method according to claim 4 , wherein the text is in a language other than the preset language, a multilingual vocabulary comprises the language of the text, and the multilingual vocabulary is used to represent a mapping relationship between words in a plurality of languages and tokens.

9 . The method according to claim 1 , wherein the target task comprises prosodic boundary prediction, and the method further comprises:

acquiring a phoneme sequence corresponding to the text; and

performing speech synthesis based on the phoneme sequence and predicted prosodic boundaries, to determine a target speech.

10 . The method according to claim 9 , wherein the target task further comprises part-of-speech prediction, and the acquiring a phoneme sequence corresponding to the text comprises:

in a case that the text comprises a polyphonic word, determining a phoneme of the polyphonic word based on part-of-speech of the polyphonic word, to determine the phoneme sequence corresponding to the text.

11 . An electronic device, comprising:

a memory and a processor, wherein the memory and the processor are in communication with each other, the memory stores computer instructions, and the processor is configured to, when executing the computer instructions, implement:

acquiring a text comprising words originating from different languages;

performing token conversion on every word in the text in a language-independent manner, to obtain a token sequence; and

performing feature extraction on the token sequence, and processing a target task based on extracted feature, to determine an analysis result for the text.

12 . The electronic device according to claim 11 , wherein the processor is further configured to implement:

acquiring a multilingual vocabulary, wherein the multilingual vocabulary is used to represent a mapping relationship between words in a plurality of languages and tokens;

querying, in the multilingual vocabulary, for tokens corresponding to the words in the text, to determine the token sequence.

13 . The electronic device according to claim 12 , wherein the processor is further configured to implement:

querying, in the multilingual vocabulary, for tokens corresponding to the words in the text; and

in a case that there is no query result for a word, dividing the word into sub-words, and querying, in the multilingual vocabulary, for tokens corresponding to the sub-words, to determine the token sequence.

14 . The electronic device according to claim 11 , wherein the processor is further configured to implement:

acquiring a target analysis model, wherein the target analysis model is obtained by analyzing a sample token sequence in at least one preset language, and the target analysis model comprises a feature extraction unit and at least one type of target task processing unit;

inputting the token sequence into the feature extraction unit for feature extraction, to obtain the extracted feature; and

inputting the extracted feature into the at least one type of target task processing unit respectively, to determine the analysis result for the text.

15 . The electronic device according to claim 14 , wherein the processor is further configured to implement:

acquiring a preset analysis model, wherein the preset analysis model comprises a pre-trained feature extraction unit and at least one preset task processing unit;

acquiring the sample token sequence in the at least one preset language and a task label corresponding to the sample token sequence;

inputting the sample token sequence into the preset analysis model to determine a prediction result corresponding to each of preset task processing units; and

calculating a loss function based on the prediction result and the task label to update parameters of the pre-trained feature extraction unit and parameters of the at least one preset task processing unit, to determine the target analysis model.

16 . The electronic device according to claim 15 , wherein the processor is further configured to implement:

for each of target tasks, acquiring a prediction category corresponding to the target task in the prediction result;

calculating a first loss function based on the prediction category and a target category corresponding to a preset task, to determine a first loss;

acquiring candidate prediction categories in the target task, wherein the prediction category is one of the candidate prediction categories;

calculating a second loss function based on the candidate prediction categories to determine a second loss; and

updating the parameters of the pre-trained feature extraction unit and the parameters of the at least one preset task processing unit based on the first loss and the second loss, to determine the target analysis model.

17 . The electronic device according to claim 16 , wherein the processor is further configured to implement:

counting a quantity of each of candidate prediction categories;

determining a loss value of each candidate prediction category based on the quantity, wherein the loss value is negatively correlated with the quantity; and

determining the second loss based on loss values of all the candidate prediction categories.

18 . The electronic device according to claim 14 , wherein the text is in a language other than the preset language, a multilingual vocabulary comprises the language of the text, and the multilingual vocabulary is used to represent a mapping relationship between words in a plurality of languages and tokens.

19 . The electronic device according to claim 11 , wherein the target task comprises prosodic boundary prediction, wherein the processor is further configured to implement:

acquiring a phoneme sequence corresponding to the text; and

performing speech synthesis based on the phoneme sequence and predicted prosodic boundaries, to determine a target speech.

20 . A non-transitory computer readable storage medium having computer instructions stored thereon, wherein the computer instructions are configured to cause a computer to implement:

acquiring a text comprising words originating from different languages;

performing token conversion on every word in the text in a language-independent manner, to obtain a token sequence; and

performing feature extraction on the token sequence, and processing a target task based on extracted feature, to determine an analysis result for the text.

Assignments (3)

Priority Claims (1)

Continuity (1)

References Cited (10)