Intent detection via multi-hop unified syntactic graph

1. A method for detecting business intent from a business intent corpus by employing an Intent Detection via Multi-hop Unified Syntactic Graph (IDMG), the method comprising:

parsing each text sample representing a business need description to extract syntactic information including at least tokens and words;

tokenizing the words of the syntactic information to generate sub-words for each of the words by employing a multi-lingual pre-trained language model;

aligning the generated sub-words to the tokens of the syntactic information to match ground-truth intent actions and objects to the tokenized sub-words;

generating a unified syntactic graph;

encoding, via a multi-hop unified syntactic graph encoder, the unified syntactic graph to generate an output; and

predicting an intent action and object from the output.

2. The method of claim 1 , wherein the syntactic information further includes at least part-of-speech for each of the words, dependency relation, and abstract meaning relation.

3. The method of claim 1 , wherein generation of the unified syntactic graph includes token node construction, node relation construction, and inter-sentence relation construction.

4. The method of claim 1 , wherein the multi-hop unified syntactic graph encoder includes a syntactic structure attention component and a multi-head semantic attention component.

5. The method of claim 4 , wherein the syntactic structure attention component receives edge relations and node attributes as input from the unified syntactic graph and the multi-head semantic attention component receives sequential embeddings as input.

6. The method of claim 5 , wherein, if a first node and a second node are directly connected, an edge embedding is initialized as a type of dependency relation and fine-tuned during a training process.

7. The method of claim 5 , wherein if a first node and a second node are not directly connected, an edge embedding is a sum of all edge segments.

8. A non-transitory computer-readable storage medium comprising a computer-readable program for detecting business intent from a business intent corpus by employing an Intent Detection via Multi-hop Unified Syntactic Graph (IDMG), wherein the computer-readable program when executed on a computer causes the computer to perform the steps of:

parsing each text sample representing a business need description to extract syntactic information including at least tokens and words;

tokenizing the words of the syntactic information to generate sub-words for each of the words by employing a multi-lingual pre-trained language model;

aligning the generated sub-words to the tokens of the syntactic information to match ground-truth intent actions and objects to the tokenized sub-words;

generating a unified syntactic graph;

encoding, via a multi-hop unified syntactic graph encoder, the unified syntactic graph to generate an output; and

predicting an intent action and object from the output.

9. The non-transitory computer-readable storage medium of claim 8 , wherein the syntactic information further includes at least part-of-speech for each of the words, dependency relation, and abstract meaning relation.

10. The non-transitory computer-readable storage medium of claim 8 , wherein generation of the unified syntactic graph includes token node construction, node relation construction, and inter-sentence relation construction.

11. The non-transitory computer-readable storage medium of claim 8 , wherein the multi-hop unified syntactic graph encoder includes a syntactic structure attention component and a multi-head semantic attention component.

12. The non-transitory computer-readable storage medium of claim 11 , wherein the syntactic structure attention component receives edge relations and node attributes as input from the unified syntactic graph and the multi-head semantic attention component receives sequential embeddings as input.

13. The non-transitory computer-readable storage medium of claim 12 , wherein, if a first node and a second node are directly connected, an edge embedding is initialized as a type of dependency relation and fine-tuned during a training process.

14. The non-transitory computer-readable storage medium of claim 12 , wherein if a first node and a second node are not directly connected, an edge embedding is a sum of all edge segments.

15. A system for detecting business intent from a business intent corpus by employing an Intent Detection via Multi-hop Unified Syntactic Graph (IDMG), the system comprising:

a memory; and

one or more processors in communication with the memory configured to:

parse each text sample representing a business need description to extract syntactic information including at least tokens and words;

tokenize the words of the syntactic information to generate sub-words for each of the words by employing a multi-lingual pre-trained language model;

align the generated sub-words to the tokens of the syntactic information to match ground-truth intent actions and objects to the tokenized sub-words;

generate a unified syntactic graph;

encode, via a multi-hop unified syntactic graph encoder, the unified syntactic graph to generate an output; and

predict an intent action and object from the output.

16. The system of claim 15 , wherein the syntactic information further includes at least part-of-speech for each of the words, dependency relation, and abstract meaning relation.

17. The system of claim 15 , wherein generation of the unified syntactic graph includes token node construction, node relation construction, and inter-sentence relation construction.

18. The system of claim 15 , wherein the multi-hop unified syntactic graph encoder includes a syntactic structure attention component and a multi-head semantic attention component.

19. The system of claim 18 , wherein the syntactic structure attention component receives edge relations and node attributes as input from the unified syntactic graph and the multi-head semantic attention component receives sequential embeddings as input.

20. The system of claim 19 ,

wherein, if a first node and a second node are directly connected, an edge embedding is initialized as a type of dependency relation and fine-tuned during a training process; and

wherein if the first node and the second node are not directly connected, the edge embedding is a sum of all edge segments.