Cross-lingual zero-shot transfer via semantic and synthetic representation learning

1. A computer-implemented method for cross-lingual transfer, comprising:

randomly masking a source corpus and a target corpus to obtain a masked source corpus and a masked target corpus;

tokenizing, by pretrained Natural Language Processing (NLP) models, the masked source corpus and the masked target corpus to obtain source tokens and target tokens, the tokenizing comprising generating features of tree nodes including token representation by the pretrained NLP models, Part-Of-Speech (POS) tagging, and Universal Dependency (UD) relation;

transforming the source tokens and the target tokens into a source dependency parsing tree and a target dependency parsing tree;

inputting the source dependency parsing tree and the target dependency parsing tree into a graph encoder pretrained on a translation language modeling task to extract common language information for transfer, the graph encoder being configured for processing the dependency parsing trees by combining adjacency matrices of the dependency parsing trees with outputs of a preceding graph, and including two graph convolution layers with weights W 1 and W 2 , and biases b 1 and b 2 , generating outputs for the two graph convolution layers; and

fine-tuning the graph encoder and a down-stream network for a specific NLP down-stream task, the fine-tuning including mask language prediction by concatenating the outputs for the two graph convolution layers, the outputs being utilized for the translation language modeling task by training as a filling-in-blank task where mask tokens are predicted by the modeling task based on their surrounding context.

2. The computer-implemented method of claim 1 , wherein the source dependency parsing tree and the target dependency parsing tree comprise token representations by the pretrained NLP models, part of speech tagging, and universal dependency relations.

3. The computer-implemented method of claim 1 , wherein the specific NLP down-stream task is Named Entity Recognition (NER).

4. The computer-implemented method of claim 1 , further comprising training the graph encoder and down-stream network on high-resource languages and evaluated the graph encoder and the down-stream network on low-resource languages.

5. The computer-implemented method of claim 1 , wherein the source corpus and the target corpus are written in respective natural languages.

6. The computer-implemented method of claim 1 , wherein the graph encoder comprises a first graph convolutional layer and a second graph convolutional layer, and wherein outputs of the source corpus and the target corpus on the second graph convolutional layer are concatenated and masked tokens are predicted.

7. The computer-implemented method of claim 1 , wherein each word in the source corpus and the target corpus are tokenized as respective tree nodes of the source dependency parsing tree and the target dependency parsing tree.

8. The computer-implemented method of claim 7 , wherein at least fifteen percent of the source tokens and the target tokens are randomly masked.

9. The computer-implemented method of claim 1 , wherein the cross-lingual transfer is from a high-resource language having a respective label for more than half of the words in the high-resource language to a low-resource language having a respective label for less than half of the words in the low-resource language.

10. A computer program product for cross-lingual transfer, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:

randomly masking a source corpus and a target corpus to obtain a masked source corpus and a masked target corpus;

tokenizing, by pretrained Natural Language Processing (NLP) models, the masked source corpus and the masked target corpus to obtain source tokens and target tokens, the tokenizing comprising generating features of tree nodes including token representation by the pretrained NLP models, Part-Of-Speech (POS) tagging, and Universal Dependency (UD) relation;

transforming the source tokens and the target tokens into a source dependency parsing tree and a target dependency parsing tree;

inputting the source dependency parsing tree and the target dependency parsing tree into a graph encoder pretrained on a translation language modeling task to extract common language information for transfer, the graph encoder being configured for processing the dependency parsing trees by combining adjacency matrices of the dependency parsing trees with outputs of a preceding graph, and including two graph convolution layers with weights W 1 and W 2 , and biases b 1 and b 2 , generating outputs for the two graph convolution layers; and

fine-tuning the graph encoder and a down-stream network for a specific NLP down-stream task, the fine-tuning including mask language prediction by concatenating the outputs for the two graph convolution layers, the outputs being utilized for the translation language modeling task by training as a filling-in-blank task where mask tokens are predicted by the modeling task based on their surrounding context.

11. The computer program product of claim 10 , wherein the source dependency parsing tree and the target dependency parsing tree comprise token representations' by the pretrained NLP models, part of speech tagging, and universal dependency relations.

12. The computer program product of claim 10 , wherein the specific NLP down-stream task is Named Entity Recognition (NER).

13. The computer program product of claim 10 , further comprising training the graph encoder and down-stream network on high-resource languages and evaluated the graph encoder and the down-stream network on low-resource languages.

14. The computer program product of claim 10 , wherein the source corpus and the target corpus are written in respective natural languages.

15. The computer program product of claim 10 , wherein the graph encoder comprises a first graph convolutional layer and a second graph convolutional layer, and wherein outputs of the source corpus and the target corpus on the second graph convolutional layer are concatenated and masked tokens are predicted.

16. The computer program product of claim 10 , wherein each word in the source corpus and the target corpus are tokenized as respective tree nodes of the source dependency parsing tree and the target dependency parsing tree.

17. The computer program product of claim 16 , wherein at least fifteen percent of the source tokens and the target tokens are randomly masked.

18. The computer program product of claim 10 , wherein the cross-lingual transfer is from a high-resource language having a respective label for more than half of words in the high-resource language to a low-resource language having a respective label for less than half of words in the low-resource language.

19. A computer processing system for cross-lingual transfer, comprising:

a memory device for storing program code; and

a processor device operatively coupled to the memory device for running the program code, the processor device being configured for:

randomly masking a source corpus and a target corpus to obtain a masked source corpus and a masked target corpus;

tokenizing, by pretrained Natural Language Processing (NLP) models, the masked source corpus and the masked target corpus to obtain source tokens and target tokens, the tokenizing comprising generating features of tree nodes including token representation by the pretrained NLP models, Part-Of-Speech (POS) tagging, and Universal Dependency (UD) relation;

transforming the source tokens and the target tokens into a source dependency parsing tree and a target dependency parsing tree;

inputting the source dependency parsing tree and the target dependency parsing tree into a graph encoder pretrained on a translation language modeling task to extract common language information for transfer, the graph encoder being configured for processing the dependency parsing trees by combining adjacency matrices of the dependency parsing trees with outputs of a preceding graph, and including two graph convolution layers with weights W 1 and W 2 , and biases b 1 and b 2 , generating outputs for the two graph convolution layers; and

fine-tuning the graph encoder and a down-stream network for a specific NLP down-stream task, the fine-tuning including mask language prediction by concatenating the outputs for the two graph convolution layers, the outputs being utilized for the translation language modeling task by training as a filling-in-blank task where mask tokens are predicted by the modeling task based on their surrounding context.

20. The computer processing system of claim 19 , wherein the source dependency parsing tree and the target dependency parsing tree comprise token representations by the pretrained NLP models, part of speech tagging, and universal dependency relations.