Low-resource task-oriented semantic parsing via intrinsic modeling for assistant systems

1. A method comprising:

receiving one or more training utterances associated with a domain, the domain being associated with

one or more ontology labels that comprise one or more of an intent or a slot;

identifying, for each ontology label, a respective index and a respective span;

obtaining an inventory for the domain, wherein the inventory comprises at least the respective index and the respective span, wherein the respective span comprises a respective descriptive label, and wherein the respective descriptive label comprises a natural-language description of the intent or the slot associated with the one or more ontology labels;

obtaining, based on the one or more training utterances and the inventory by a natural language understanding (NLU) model, a structural representation of a respective training utterance, wherein the structural representation is generated based on a comparison between the respective training utterance and the inventory; and

obtaining an updated NLU model based on the structural representation and the respective training utterance.

2. The method of claim 1 , further comprising:

obtaining, for the one or more training utterances, one or more embeddings, respectively;

obtaining, for the inventory, a linearized string;

obtaining, for each of the one or more embeddings, one or more concatenations, each comprising a respective embedding of the one or more embeddings concatenated with the linearized string; and

inputting the one or more concatenations to the NLU model.

3. The method of claim 1 , wherein the inventory further comprises a respective type for each ontology label.

4. The method of claim 3 , wherein one or more of the types of the inventory are associated with one or more rules, and wherein updating the NLU model is further based on the one or more rules.

5. The method of claim 3 , further comprising:

obtaining, based on the one or more ontology labels, the one or more types and the one or more spans corresponding to the one or more ontology labels, respectively.

6. The method of claim 1 , wherein the inventory is based on a tabular structure comprising one or more tuples, wherein each tuple stores an index, a type, and a span for a corresponding ontology label of the one or more ontology labels.

7. The method of claim 1 , wherein the NLU model is based on a sequence-to-sequence language model.

8. The method of claim 1 , wherein each training utterance comprises one or more utterance tokens, and wherein the inventory comprises a plurality of ontology tokens, wherein each ontology token comprises one or more of an index, a type, or a span corresponding to an ontology label.

9. The method of claim 8 , wherein obtaining the representation comprises selecting one or more of an utterance token from each training utterance or an ontology token from the inventory.

10. The method of claim 8 , wherein obtaining the structural representation comprises selecting one or more ontology tokens from the inventory based on a self-attention mechanism.

11. The method of claim 8 , wherein the structural representation is based on one or more of an utterance token or an ontology token.

12. The method of claim 1 , further comprising:

receiving, from a client system, a user input;

determining, based on the user input by the updated NLU model, one or more intents and one or more slots; and

executing one or more tasks, wherein the one or more tasks are determined based on the one or more intents and the one or more slots.

13. A computer-readable non-transitory medium comprising instructions, which cause operations comprising:

receive one or more training utterances associated with a domain, the domain being associated with

one or more ontology labels for the domain that, comprise one or more of an intent or a slot;

identify, for each ontology label, a respective index and a respective span;

obtain an inventory for the domain, wherein the inventory comprises at least the respective index and the respective span, wherein the respective span comprises a respective descriptive label, and wherein the respective descriptive label comprises a natural-language description of the intent or slot ontology label;

obtain, based on the one or more training utterances and the inventory by a natural language understanding (NLU) model, a structural representation of a respective training utterance, wherein the structural representation is generated based on a comparison between the the respective training utterance and the inventory; and

obtain an updated NLU model based on the structural representation and the respective training utterance.

14. The computer-readable non-transitory medium of claim 13 , wherein the instructions are further configured to cause operations for:

obtain, for the one or more training utterances, one or more embeddings, respectively;

generate, for the inventory, a linearized string;

concatenate each of the one or more embeddings with the linearized string; and

input the one or more concatenations to the NLU model.

15. The computer-readable non-transitory medium of claim 13 , wherein the inventory further comprises a respective type for each ontology label.

16. The computer-readable non-transitory medium of claim 15 , wherein one or more of the types of the inventory are associated with one or more rules, and wherein updating the NLU model is further based on the one or more rules.

17. A method, comprising:

receiving, from a client system, a user input that includes speech data;

determining one or more tasks, based on applying the user input to a trained natural language understanding (NLU) model, wherein the trained NLU model is configured for use based on:

obtaining an inventory for a domain associated with one or more training utterances, wherein the inventory comprises at least a respective index and a respective span for each of one or more of an intent or a slot of one or more ontology labels, wherein the respective span comprises a respective descriptive label associated with the intent or slot each ontology label, and wherein the respective descriptive label comprises a natural-language description of the intent or slot ontology label;

obtaining, based on the one or more training utterances and the inventory by the NLU model, a structural representation of a respective training utterance, wherein the structural representation is generated based on a comparison between the respective training utterance and the inventory; and

executing the one or more tasks.

18. The method of claim 17 , further comprising:

obtaining, in conjunction with the user input, non-speech data;

providing an output from the NLU model and another output based on the non-speech data, to an entity resolution module; and

determining the one or more tasks further in accordance with yet another output of the entity resolution module.

19. The method of claim 18 , wherein the non-speech data includes one or more of:

a location,

a visual,

a gesture and/or touch being performed by a user providing the user input,

a world update,

a social update, and

information related to a context, a person, and/or an activity associated with the user providing the user input.

20. The method of claim 18 , further comprising:

in accordance with providing the output from the NLU model and the non-speech data to the entity resolution module, determining whether to handle the user input in one of (i) a first operational mode that causes the tasks to be performed entirely at a client device, and (ii) a second operational mode that includes performing one or more operations at a remote server separate from the client device.