Statistical machine translation method using dependency forest

1. A method of generating a translation rule where the method is performed by a translation apparatus comprising a dependency parser, a translation rule extractor, a language model trainer, a memory and a decoder, the method comprising:

performing, with the dependency parser, a dependency analysis for a bilingual corpus;

generating, with the dependency parser, a dependency tree based on the dependency analysis;

generating, with the dependency parser, a dependency forest for the bilingual corpus by combining a plurality of dependency trees, wherein the dependency forest is aligned with a source sentence string;

generating, with the dependency parser, the dependency forest by connecting nodes forming the plurality of dependency trees by a hyperedge, wherein respective nodes of the dependency forest are connected by the hyperedge and the hyperedge is configured to pack all dependents having a common head;

extracting, with the translation rule extractor, a translation rule based on the dependency forest;

collecting, with the language model trainer, all heads and dependents by listing all hyperedges of the dependency forest;

generating, with the language model trainer, a dependency language model based on the collected heads and dependents;

storing the translation rule into the memory;

reading, with the decoder, the stored translation rule from the memory;

receiving, with the decoder, a source language text; and

generating, with the decoder, a target language text from the source language text by applying the translation rule.

2. The method of claim 1 , wherein the method further comprises:

searching for a plurality of well-formed structures for each node within the dependency forest; and

extracting the translation rule when dependency structures within the plurality of well-formed structures correspond to a word alignment.

3. The method of claim 2 , wherein the method comprises:

obtaining the plurality of well-formed structures by manipulating a fixed structure of dependents of the node, wherein the plurality of well-formed structures have k-best fixed and floating structures.

4. The method of claim 1 , wherein the nodes are distinguished by a span.

5. The method of claim 1 , wherein extracting the translation rule is based on string-to-forest aligned corpus.

6. The method of claim 1 , wherein the method comprises:

maintaining a plurality of best well-formed structures for each node by searching for a well-formed structure for the node.

7. The method of claim 6 , wherein the method comprises:

obtaining the plurality of best well-formed structures by connecting fixed structures of dependents of the node.

8. The method of claim 1 , wherein the method further comprises:

translating a source language by using the translation rule and the dependency language model.

9. An apparatus for generating a translation rule comprising:

a non-transitory computer-readable medium; and

instructions stored on the computer-readable medium, the instructions configured to, when executed by a processor, cause the processor to:

perform a dependency analysis for a corpus of a pair of languages;

generate a dependency tree based on the dependency analysis;

generate a dependency forest for the corpus of the pair of languages by combining a plurality of dependency trees, wherein the dependency forest is aligned with a source sentence string;

generate the dependency forest by connecting nodes forming the plurality of dependency trees by a hyperedge, wherein respective nodes of the dependency forest are connected by the hyperedge and the hyperedge is configured to pack all dependents having a common head;

collect all heads and dependents by listing all hyperedges of the dependency forest;

generate a dependency language model based on the collected heads and dependents;

search for a plurality of well-formed structures for each node within the dependency forest;

extract a translation rule when dependency structures within the plurality of well-formed structures correspond to a word alignment;

store the translation rule;

read the stored translation rule;

receive a source language text; and

generate a target language text from the source language text by applying the translation rule.

10. The apparatus of claim 9 , wherein:

the plurality of well-formed structures is obtained by controlling a fixed structure of dependents of the node, wherein the plurality of well-formed structures have k-best fixed and floating structures.

11. A statistical machine translation apparatus comprising:

a dependency parser configured to:

generate a dependency tree by performing a dependency analysis for a source sentence and a target sentence of a corpus of a pair of languages;

generate a dependency forest for the source sentence and the target sentence by combining a plurality of dependency trees; and

generate the dependency forest by connecting nodes forming a plurality of dependency trees by a hyperedge, wherein the hyperedge packs all dependents having a common head;

a translation rule extractor configured to extract a translation rule by using the dependency forest;

a language model trainer configured to:

generate a dependency language model by using the dependency forest of the target sentence;

collect all heads and dependents by listing all hyperedges of the dependency forest; and

generate the dependency language model based on the collected heads and dependents; and

a decoder configured to convert a source sentence text to a target sentence text based on the extracted translation rule and the generated dependency language model.

12. The apparatus of claim 11 , wherein the translation rule extractor is configured to search for a plurality of well-formed structures for each node within the dependency forest and to extract the translation rule when dependency structures within the plurality of well-formed structures correspond to a word alignment.