Multi-source transfer of delexicalized dependency parsers

1. A computer implemented method, comprising:

identifying a parse tree for a source-language sentence, the parse tree being generated based on tagging the source-language sentence and parsing the tagged source-language sentence using a source language lexicalized parser that has been trained using source-language Treebank data;

tagging a target-language sentence with parts of speech tags, where the target-language sentence is a translation of the source-language sentence;

parsing, utilizing one or more processors, the tagged target-language sentence with a delexicalized parser to generate a set of k-best parse trees for the target-language sentence, where the delexicalized parser has been trained using source-language Treebank data;

selecting the best target-language parse tree of the k-best parse trees that most closely aligns with the parse tree of the source-language sentence; and

updating a parameter vector of a target language lexicalized parser based upon the selected best target-language parse tree.

2. A computer implemented method, comprising:

identifying a parse tree for a source-language sentence, the parse tree being generated based on tagging the source-language sentence with non-lexical tags and parsing the tagged source-language sentence using a source language lexicalized parser that has been trained using gold standard source-language data;

tagging a target-language sentence with non-lexical tags, where the target-language sentence is a translation of the source-language sentence;

parsing, utilizing one or more of the processors, the tagged target-language sentence with a delexicalized parser to generate a set of k-best parse trees for the target-language sentence;

selecting the best target-language parse tree of the k-best parse trees that most closely aligns with the parse tree of the source-language sentence; and

updating a parameter vector of a target language lexicalized parser based upon the selected best target-language parse tree.

3. The method of claim 2 , wherein the non-lexical tags are parts of speech tags.

4. The method of claim 2 , wherein the gold standard data is Treebank data.

5. The method of claim 2 , wherein the delexicalized parser has been trained using source-language gold standard data.

6. The method of claim 2 , wherein the delexicalized parser has been trained using gold standard data for a plurality of distinct languages.

7. The method of claim 2 , wherein the delexicalized parser has been trained using concatenating gold standard data for each of a plurality of distinct languages.

8. The method of claim 2 , wherein the delexicalized parser has been trained using weighted gold standard data corresponding to each of a plurality of distinct languages.

9. The method of claim 8 , wherein the gold standard data includes a subset of gold standard data for each of the plurality of languages and a weight is assigned to each of the subsets of gold standard data.

10. The method of claim 2 , wherein the selecting the best target-language parse tree of the k-best parse trees that most closely aligns with the parse tree of the source-language sentence comprises calculating an alignment score that improves with the number of correctly aligned dependencies and worsens with the number of incorrectly aligned dependencies.

11. The method of claim 2 , wherein the source-language sentence is one of a plurality of sentences that includes a plurality of subsets of sentences, at least two of the subsets of sentences are in distinct languages.

12. A system, comprising;

a processor;

memory coupled to said processor, where the memory stores instructions to be executed by the processor to perform steps comprising:

identifying a parse tree for a source-language sentence, the parse tree being generated based on tagging the source-language sentence with non-lexical tags and parsing the tagged source-language sentence using a source language lexicalized parser that has been trained using gold standard source-language data;

tagging a target-language sentence with non-lexical tags, where the target-language sentence is a translation of the source-language sentence;

parsing, utilizing one or more of the processors, the tagged target-language sentence with a delexicalized parser to generate a set of k-best parse trees for the target-language sentence;

selecting the best target-language parse tree of the k-best parse trees that most closely aligns with the parse tree of the source-language sentence; and

updating a parameter vector of a target language lexicalized parser based upon the selected best target-language parse tree.

13. The system of claim 12 , wherein the non-lexical tags are parts of speech tags.

14. The system of claim 12 , wherein the gold standard data is Treebank data.

15. The system of claim 12 , wherein the delexicalized parser has been trained using source-language gold standard data.

16. The system of claim 12 , wherein the delexicalized parser has been trained using gold standard data for a plurality of distinct languages.

17. The system of claim 12 , wherein the delexicalized parser has been trained using concatenating gold standard data for each of a plurality of distinct languages.

18. The system of claim 12 , wherein the delexicalized parser has been trained using weighted gold standard data corresponding to each of a plurality of distinct languages.

19. The system of claim 18 , wherein the gold standard data includes a subset of gold standard data for each of the plurality of languages and a weight is assigned to each of the subsets of gold standard data.

20. The system of claim 12 , wherein the selecting the best target-language parse tree of the k-best parse trees that most closely aligns with the parse tree of the source-language sentence comprises calculating an alignment score that improves with the number of correctly aligned dependencies and worsens with the number of incorrectly aligned dependencies.