Intent classification method and system

1. A processing-circuit-implemented method for determining an intent class of a dialogue sentence input in a language containing information-bearing characters, comprising:

obtaining at least one set of word embeddings and character embeddings from the dialogue sentence input to thereby generate at least one set of integrated word-character embeddings, each corresponding to one of the at least one set of word embeddings and character embeddings; and

determining an intent class corresponding to the dialogue sentence input based on the at least one set of integrated word-character embeddings;

wherein in the obtaining at least one set of word embeddings and character embeddings from the dialogue sentence input to thereby generate at least one set of integrated word-character embeddings, one or more of the at least one set of integrated word-character embeddings is updated during backpropagation.

2. The processing-circuit-implemented method of claim 1 , wherein the dialogue sentence input comprises a voice input, and wherein the method further comprises outputting a dialogue based on the determined intent class as at least one of an audio signal and a visual signal.

3. The processing-circuit-implemented method of claim 1 , wherein a number of the at least one set of integrated word-character embeddings is more than one, wherein the determining an intent class corresponding to the dialogue sentence input based on the at least one set of integrated word-character embeddings comprises:

obtaining a prediction result specifying a predicted intent class corresponding to the dialogue sentence input from a model containing each of the more than one set of integrated word-character embeddings; and

ensembling the prediction result from the model containing the each of the more than one set of integrated word-character embeddings to thereby determine an intent class corresponding to the dialogue sentence input.

4. The processing-circuit-implemented method of claim 3 , wherein the obtaining a prediction result specifying a predicted intent class corresponding to the dialogue sentence input from a model containing each of the more than one set of integrated word-character embeddings is realized through a long short-term memory (LSTM) neural network-based approach.

5. The processing-circuit-implemented method of claim 3 , wherein the ensembling the prediction result from the model containing the each of the more than one set of integrated word-character embeddings to thereby determine an intent class corresponding to the dialogue sentence is realized through majority voting.

6. The processing-circuit-implemented method of claim 1 , wherein in the obtaining at least one set of word embeddings and character embeddings from the dialogue sentence input to thereby generate at least one set of integrated word-character embeddings, either or both of the word embeddings and character embeddings in one or more of the at least one set of word embeddings and character embeddings are randomly initialized.

7. The processing-circuit-implemented method of claim 1 , wherein in the obtaining at least one set of word embeddings and character embeddings from the dialogue sentence input to thereby generate at least one set of integrated word-character embeddings, either or both of the word embeddings and character embeddings in one or more of the at least one set of word embeddings and character embeddings are pre-trained.

8. The processing-circuit-implemented method of claim 1 , wherein the language comprises at least one logographic language including Chinese.

9. A processing-circuit-implemented method for determining an intent class of a dialogue sentence input in a language containing information-bearing characters, comprising:

obtaining at least one set of word embeddings and character embeddings from the dialogue sentence input to thereby generate at least one set of integrated word-character embeddings, each corresponding to one of the at least one set of word embeddings and character embeddings; and

determining an intent class corresponding to the dialogue sentence input based on the at least one set of integrated word-character embeddings;

wherein the obtaining at least one set of word embeddings and character embeddings from the dialogue sentence input to thereby generate at least one set of integrated word-character embeddings comprises at least one round of the sub-steps of:

generating a character embedding based input and a word embedding based input from the dialogue sentence input;

generating character embedding vectors and word embedding vectors respectively from the character embedding based input and the word embedding based input;

extracting features from the character embedding vectors to thereby obtain a character feature map;

obtaining a pooled character feature map based on the character feature map; and

integrating the pooled character feature map and the word embedding vectors to thereby obtain integrated word-character embedding vectors;

wherein:

each of the at least round of the sub-steps corresponds to one of the at least one set of integrated word-character embeddings.

10. The processing-circuit-implemented method of claim 9 , wherein the extracting features from the character embedding vectors to thereby obtain a character feature map is realized through a 2-D convolutional neural network (CNN)-based approach.

11. The processing-circuit-implemented method of claim 9 , wherein the obtaining a pooled character feature map based on the character feature map is realized through a max-pooling operation.

12. The processing-circuit-implemented method of claim 9 , wherein in the obtaining a pooled character feature map based on the character feature map, the pooled character feature map has a substantially same dimension as the word embedding vectors.

13. The processing-circuit-implemented method of claim 9 , wherein the integrating the pooled character feature map and the word embedding vectors to thereby obtain integrated word-character embedding vectors is realized through a weighted average-based approach.

14. The processing-circuit-implemented method of claim 9 , wherein the integrating the pooled character feature map and the word embedding vectors to thereby obtain integrated word-character embedding vectors is realized through a concatenation-based approach.

15. The processing-circuit-implemented method of claim 9 , wherein each of the at least one round of the sub-steps further comprises, prior to the generating a character embedding based input and a word embedding based input from the dialogue sentence input:

preprocessing the dialogue sentence input allowing tokenization thereof at a word level.

16. A system, comprising:

at least one processing circuit; and

memory storing instructions that, when executed by the at least one processing circuit, cause the system to:

obtain at least one set of word embeddings and character embeddings from a dialogue sentence input to thereby generate at least one set of integrated word-character embeddings, each corresponding to one of the at least one set of word embeddings and character embeddings; and

determine an intent class corresponding to the dialogue sentence input based on the at least one set of integrated word-character embeddings;

wherein the instructions causing the system to obtain at least one set of word embeddings and character embeddings from a dialogue sentence input to thereby generate at least one set of integrated word-character embeddings, each corresponding to one of the at least one set of word embeddings and character embeddings comprise:

a first instruction causing the system to generate a character embedding based input and a word embedding based input from the dialogue sentence input;

a second instruction causing the system to generate character embedding vectors and word embedding vectors respectively from the character embedding based input and the word embedding based input;

a third instruction causing the system to extract features from the character embedding vectors to thereby obtain a character feature map;

a fourth instruction causing the system to obtain a pooled character feature map based on the character feature map; and

a fifth instruction causing the system to integrate the pooled character feature map and the word embedding vectors to thereby obtain integrated word-character embedding vectors.

17. The system of claim 16 , wherein the instructions further cause the system to preprocess the dialogue sentence input allowing tokenization thereof at a word level.

18. The system of claim 16 , wherein the system is further configured to output a dialogue based on the determined intent class as at least one of an audio signal and a visual signal.

19. The system of claim 16 , wherein the instructions stored in the memory further comprise:

a sixth instruction causing the system to obtain a prediction result specifying a predicted intent class corresponding to the dialogue sentence input from a model containing the integrated word-character embedding vectors.

20. The system of claim 19 , wherein the instructions stored in the memory further comprise:

a seventh instruction causing the system to ensemble the prediction result from more than one model, each containing one set of integrated word-character embedding vectors to thereby determine an intent class corresponding to the dialogue sentence input.