Cross-domain label-adaptive stance detection

1 . A method, comprising:

receiving a natural language input;

tokenizing an entirety of the natural language input by a shared tokenizer to identify tokens in the natural language input;

parsing an entirety of the tokens by a plurality of domain expert encoder blocks to produce a corresponding plurality of domain encodings for the natural language input;

parsing the entirety of the tokens by a global encoder block to produce a global encoding for the natural language input separately and in parallel to the parsing of the entirety of the tokens by the plurality of domain expert encoder blocks;

processing the plurality of domain encodings and the global encoding by a shared label embedding layer to produce a probability distribution for a stance of the natural language input, wherein the shared label embedding layer receives the plurality of domain encodings and the global encoding separately and in parallel, and wherein the shared label embedding layer masks unrelated labels and keeps visible only labels from target datasets for a sample in the received natural language input; and

outputting the stance for the natural language input,

wherein the shared tokenizer, the plurality of domain expert encoder blocks, the global encoder block, and the shared label embedding layer comprise an architecture of a natural language analysis model.

2 . The method of claim 1 , wherein each domain encoding of the plurality of domain encodings includes a set of probabilities for whether the natural language input exhibits a candidate stance from a plurality of stances used by a corresponding domain expert encoder block of the plurality of domain expert encoder blocks.

3 . The method of claim 1 , wherein the label embedding layer is a domain-adversarial neural network (DANN) trained via a domain-adversarial loss function to minimize a task objective and maximize a confusion level in a domain classifier.

4 . The method of claim 1 , wherein the stance is associated with a first name in a first domain expert encoder block of the plurality of domain expert encoder blocks and with a second name in a second domain expert encoder block of the plurality of domain expert encoder blocks.

5 . The method of claim 1 , wherein the stance is associated with a category outside of a classification scheme used by at least one domain expert encoder block of the plurality of domain expert encoder blocks.

6 . The method of claim 5 , wherein the stance is associated with a category outside of any classification scheme used by any domain expert encoder block of the plurality of domain expert encoder blocks.

7 . The method of claim 1 , wherein the stance is associated with a first name in a first domain expert encoder block of the plurality of domain expert encoder blocks and not with the first name in a second domain expert encoder block of the plurality of domain expert encoder blocks.

8 . A system, comprising:

a processor; and

a memory including instructions that when executed by the processor, the system performs operations including:

receiving a natural language input;

tokenizing an entirety of the natural language input by a shared tokenizer to identify tokens in the natural language input;

parsing an entirety of the tokens by a plurality of domain expert encoder blocks to produce a corresponding plurality of domain encodings for the natural language input;

parsing the entirety of the tokens by a global encoder block to produce a global encoding for the natural language input separately and in parallel to the parsing of the entirety of the tokens by the plurality of domain expert encoder blocks;

processing the plurality of domain encodings and the global encoding by a shared label embedding layer to produce a probability distribution for a stance of the natural language input, wherein the shared label embedding layer receives the plurality of domain encodings and the global encoding separately and in parallel, and wherein the shared label embedding layer masks unrelated labels and keeps visible only labels from target datasets for a sample in the received natural language input; and

outputting the stance for the natural language input,

wherein the shared tokenizer, the plurality of domain expert encoder blocks, the global encoder block, and the shared label embedding layer comprise an architecture of a natural language analysis model.

9 . The system of claim 8 , wherein each domain encoding of the plurality of domain encodings includes a set of probabilities for whether the natural language input exhibits a candidate stance from a plurality of stances used by a corresponding domain expert encoder block of the plurality of domain expert encoder blocks.

10 . The system of claim 8 , wherein the label embedding layer is a domain- adversarial neural network (DANN) trained via a domain-adversarial loss function to minimize a task objective and maximize a confusion level in a domain classifier.

11 . The system of claim 8 , wherein the stance is associated with a first name in a first domain expert encoder block of the plurality of domain expert encoder blocks and with a second name in a second domain expert encoder block of the plurality of domain expert encoder blocks.

12 . The system of claim 8 , wherein the stance is associated with a category outside of a classification scheme used by at least one domain expert encoder block of the plurality of domain expert encoder blocks.

13 . The system of claim 12 , wherein the stance is associated with a category outside of any classification scheme used by any domain expert encoder block of the plurality of domain expert encoder blocks.

14 . The system of claim 8 , wherein the stance is associated with a first name in a first domain expert encoder block of the plurality of domain expert encoder blocks and not with the first name in a second domain expert encoder block of the plurality of domain expert encoder blocks.

15 . A non-transitory computer-readable storage device including instructions that when executed by a processor performs operations comprising:

receiving a natural language input;

tokenizing an entirety of the natural language input by a shared tokenizer to identify tokens in the natural language input;

parsing an entirety of the tokens by a plurality of domain expert encoder blocks to produce a corresponding plurality of domain encodings for the natural language input;

parsing the entirety of the tokens by a global encoder block to produce a global encoding for the natural language input separately and in parallel to the parsing of the entirety of the tokens by the plurality of domain expert encoder blocks;

processing the plurality of domain encodings and the global encoding by a shared label embedding layer to produce a probability distribution for a stance of the natural language input, wherein the shared label embedding layer receives the plurality of domain encodings and the global encoding separately and in parallel, and wherein the shared label embedding layer masks unrelated labels and keeps visible only labels from target datasets for a sample in the received natural language input; and

outputting the stance for the natural language input,

wherein the shared tokenizer, the plurality of domain expert encoder blocks, the global encoder block, and the shared label embedding layer comprise an architecture of a natural language analysis model.

16 . The non-transitory computer-readable storage device of claim 15 , wherein each domain encoding of the plurality of domain encodings includes a set of probabilities for whether the natural language input exhibits a candidate stance from a plurality of stances used by a corresponding domain expert encoder block of the plurality of domain expert encoder blocks.

17 . The non-transitory computer-readable storage device of claim 15 , wherein the label embedding layer is a domain-adversarial neural network (DANN) trained via a domain-adversarial loss function to minimize a task objective and maximize a confusion level in a domain classifier.

18 . The non-transitory computer-readable storage device of claim 15 , wherein the stance is associated with a first name in a first domain expert encoder block of the plurality of domain expert encoder blocks and with a second name in a second domain expert encoder block of the plurality of domain expert encoder blocks.

19 . The non-transitory computer-readable storage device of claim 15 , wherein the stance is associated with a category outside of a classification scheme used by at least one domain expert encoder block of the plurality of domain expert encoder blocks.

20 . The non-transitory computer-readable storage device of claim 15 , wherein the stance is associated with a first name in a first domain expert encoder block of the plurality of domain expert encoder blocks and not with the first name in a second domain expert encoder block of the plurality of domain expert encoder blocks.