Extracting treelet translation pairs

1. A computer-implemented method of identifying treelet translation pairs for use in a machine translation system that translates a source language input into a target language output, the method comprising:

accessing a corpus of pairs of aligned, parallel syntactic dependency structures, each pair including a source language dependency structure having nodes that represent lexical items, the nodes being aligned with nodes representing lexical items in a target language dependency structure;

enumerating individual source nodes and combinations of source nodes connected in the source language dependency structure as possible source treelets identifying lexical items, and corresponding dependencies, in the target language dependency structure, that are aligned with the enumerated nodes and combinations of connected nodes, as possible target treelets corresponding to the possible source treelets;

extracting well formed treelet translation pairs from the possible source treelets and possible target treelets using a processor of the computer, by extracting as a well formed treelet translation pair the possible source treelet and corresponding possible target treelet only if the lexical items in the possible source treelet are only aligned with lexical items in the possible target treelet and the lexical items in the possible target treelet are only aligned with lexical items in the possible source;

storing the treelet translation pairs in a data store; and

outputting the treelet translation pairs for use by a component in a machine translation system.

2. The method of claim 1 , wherein each child node of a parent node is considered to be connected with other child nodes of the parent node.

3. The method of claim 1 wherein the source language dependency structures each represent a source language text fragment, and wherein enumerating comprises:

enumerating connected portions of the source language dependency structure regardless of whether the connected portions represent discontiguous sets of words in the source language text fragment.

4. The method of claim 1 wherein the source language dependency structure comprises a source language dependency tree, and wherein enumerating comprises:

enumerating connected sets of nodes that represent a non-linear branch in the source language dependency tree.

5. The method of claim 1 and further comprising:

accessing the stored treelet translation pairs with a decoder to translate an input.

6. The method of claim 5 and further comprising:

generating N-best translations of the input.

7. The method of claim 6 and further comprising:

outputting the N-best translations.

8. The method of claim 1 , wherein extracting comprises extracting a plurality of well formed treelet translation pairs, and wherein extracting comprises:

limiting the step of extracting the well formed treelet translation pairs such that, for each given possible source treelet and its given corresponding possible target treelet, the given possible source treelet and given corresponding possible target treelet are only extracted as a well formed treelet translation pair if the lexical items in the given possible source treelet are only aligned with lexical items in the given corresponding possible target treelet and are not aligned with other lexical items in the possible target treelets that are different than the lexical items in the given corresponding possible target treelet, and the given possible source treelet and given corresponding possible target treelet are only extracted as a well formed treelet translation pair if the lexical items in the given corresponding possible target treelet are only aligned with lexical items in the given possible source treelet and are not aligned with other lexical items in the possible source treelets that are different than the lexical items in the given possible source treelet.

9. A computer-implemented system for identifying treelet translation pairs, from training data, for use in a machine translation system that translates a source language input into a target language output, the system comprising:

a processor;

a treelet pair extractor implemented on the processor to access a corpus of pairs of aligned, parallel syntactic dependency structures stored in a database, each pair including a source language dependency structure having nodes that represent lexical items, the nodes being aligned with nodes representing lexical items in a target language dependency structure; and

the treelet pair extractor being further configured to enumerate sets of source nodes that are connected portions of the source language dependency structure as possible source treelets, wherein the treelet pair extractor is configured to identify lexical items, and corresponding dependencies, in the target language dependency structure, that are aligned with the enumerated connected sets of nodes, as possible target treelets corresponding to the possible source treelets, and to extract as well formed treelet translation pairs the possible source treelets and corresponding possible target treelets only if the lexical items in a given possible source treelet are only aligned with lexical items in a given possible target treelet and the lexical items in the given possible target treelet are only aligned with lexical items in the given possible source treelet, wherein the treelet pair extractor is configured to output the treelet translation pairs for use by a component of a machine translation system.

10. The system of claim 9 wherein each child node of a parent node is considered to be connected with other child nodes of the parent node, and extract well formed treelet translation pairs from the possible source treelets and corresponding aligned portions of the target language dependency structure.

11. The system of claim 9 and further comprising:

a data store storing the extracted treelet translation pairs.

12. The system of claim 9 wherein the source language dependency structures each represent a source language text fragment.

13. The system of claim 12 wherein the treelet translation pair extractor is configured to enumerate connected sets of source nodes in the source language dependency structure regardless of whether they represent discontiguous words in the source language text fragment.

14. The system of claim 9 wherein the source language dependency structure comprises a source language dependency tree, and wherein the treelet pair extractor is configured to enumerate connected sets of nodes that represent a non-linear branch in the source language dependency tree.

15. The system of claim 9 and further comprising:

a decoder configured to access the well formed treelet translation pairs to translate a textual input.

16. The system of claim 15 wherein the decoder is configured to generate and output N-best translations of the textual input.

17. The computer-implemented system of claim 9 , wherein the treelet pair extractor is configured to extract a plurality of well formed treelet translation pairs from the possible source treelets and corresponding possible target treelets, wherein the treelet pair extractor limits the possible source treelets and corresponding possible target treelets extracted as well formed treelet translation pairs such that, for each given possible source treelet and its given corresponding possible target treelet, the given possible source treelet and given corresponding possible target treelet are only extracted as a well formed treelet translation pair if the lexical items in the given possible source treelet are only aligned with lexical items in the given corresponding possible target treelet and are not aligned with other lexical items in the possible target treelets that are different than the lexical items in the given corresponding possible target treelet, and the given possible source treelet and given corresponding possible target treelet are only extracted as a well formed treelet translation pair if the lexical items in the given corresponding possible target treelet are only aligned with lexical items in the given possible source treelet and are not aligned with other lexical items in the possible source treelets that are different than the lexical items in the given possible source treelet.

18. A computer readable storage medium storing computer readable instructions which, when executed by a processor associated with the computer, cause the computer to perform a computer-implemented method for identifying treelet translation pairs for use in a machine translation system that translates a source language input into a target language output, the computer-implemented method comprising:

accessing a corpus of pairs of aligned, parallel syntactic dependency structures, each pair including a source language dependency structure having nodes that represent lexical items, the nodes being aligned with nodes representing lexical items in a target language dependency structure;

enumerating connected sets of source nodes in the source language dependency structure as possible source treelets;

identifying lexical items, and corresponding dependencies, in the target language dependency structure, that are aligned with the enumerated connected sets of nodes, as possible target treelets corresponding to the possible source treelets;

extracting a plurality of well formed treelet translation pairs from the possible source treelets and the possible target treelets using the processor, wherein extracting comprises:

limiting the step of extracting the well formed treelet translation pairs such that, for each given possible source treelet and its given corresponding possible target treelet, the given possible source treelet and given corresponding possible target treelet are only extracted as a well formed treelet translation pair if the lexical items in the given possible source treelet are only aligned with lexical items in the given corresponding possible target treelet and are not aligned with other lexical items in the possible target treelets that are different than the lexical items in the given corresponding possible target treelet, and the given possible source treelet and given corresponding possible target treelet are only extracted as a well formed treelet translation pair if the lexical items in the given corresponding possible target treelet are only aligned with lexical items in the given possible source treelet and are not aligned with other lexical items in the possible source treelets that are different than the lexical items in the given possible source treelet; and

outputting the treelet translation pairs for use by a component in a machine translation system.

19. The computer readable storage medium of claim 18 wherein each child node of a parent node is considered to be connected to other child nodes of the parent node.

20. The computer readable storage medium of claim 18 wherein the method further comprises:

storing the treelet translation pairs in a data store.

21. The computer readable storage medium of claim 18 wherein the source language dependency structures each represent a source language text fragment, and wherein enumerating comprises:

enumerating connected sets of source nodes in the source language dependency structure that may represent discontiguous words in the source language text fragment.

22. The computer readable storage medium of claim 18 wherein the source language dependency structure comprises a source language dependency tree, and wherein enumerating comprises:

enumerating connected sets of nodes that represent a non-linear branch in the source language dependency tree.