TRANSLATION SYSTEM ADAPTED FOR QUERY TRANSLATION VIA A RERANKING FRAMEWORK

• US 20130006954A1
• Filed: 06/30/2011
• Published: 01/03/2013
• Est. Priority Date: 06/30/2011
• Status: Active Grant

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Nikoulina, Vassilina, Lagos, Nikolaos
• Granted Patent: US 8,713,037 B2
• Time in Patent Office: Days
• US Class Current: 707/706
• CPC Class Codes: G06F 16/3337 Translation of the query language, e.g. Chinese to English
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models
• CPC Class Codes: G06F 40/45 Example-based machine translation; Alignment
• CPC Class Codes: G06F 40/51 Translation evaluation

Abstract

A system and method suited to translation of queries are disclosed. The method includes receiving an input query in a source language and outputting a target query, based on a translation of the input query into a target language, different from the source language. The translation is performed with a machine translation system which has been trained on representations of features of translated queries that have been generated by translation of an original query, in the source language, into the target language and a measure of information retrieval performance of each the translated queries, for each of a set of original queries.

Claims

1. A translation method comprising:

receiving an input query in a source language; and

outputting a target query, the target query being based on a translation of the input query into a target language, different from the source language, with a machine translation system which has been trained on representations of features of translated queries, the translated queries having been generated by translation of an original query in the source language into the target language and a measure of information retrieval performance of each the translated queries, for each of a set of original queries.

2. The method of claim 1, wherein the machine translation system is implemented with a computer processor.

3. The method of claim 1, wherein the outputting comprises outputting the target query to a search engine and retrieving information based on the target query.

4. The method of claim 1, wherein the outputting comprises:

translating the input query with a decoder of the machine translation system to generate a set of candidate queries in the target language;

extracting features for each of the candidate queries in the set; and

ranking the candidate queries based on a representation of the extracted features for each candidate query and feature weights for the features learned in the training of the machine translation system.

5. The method of claim 4, wherein the reranking comprises computing a function of the extracted features and learned feature weights for the features.

6. The method of claim 5, wherein the computing of the function comprises computing a translation score for each of the candidate queries as a weighted linear combination of its extracted features.

7. The method of claim 6, wherein the method includes outputting, as the target query, a candidate query which satisfies:

{circumflex over (t)}=argmax<sub>tiε

GEN(q)</sub>Λ

·

F(t_i) 

 

(Eqn.

     1)

where t_iε

(GEN(q)) represents a candidate query generated from a source query q, Λ

represents a set of feature weights learned in training, one weight for each of the features in F(t_i) and ·

represents the dot product.

8. The method of claim 4, wherein at least one of the extracted features is based on parts of speech for the candidate queries.

9. The method of claim 8, wherein at least one of the part of speech features is based on a part of speech for an element of the candidate query and a corresponding part of speech of an element of the input query with which the element of the candidate query is aligned in the translation, wherein each of the elements comprises at least one word.

10. The method of claim 9, wherein the at least one of the part of speech features is also based on a frequency of the element in the candidate query as a translation of element of the input query in a training corpus of bi-sentences, each bi-sentence including a sentence in the source language and a sentence in the target language.

11. The method of claim 4, wherein the outputting of the target query comprises outputting a most highly ranked candidate query as the target query or outputting a combination of more highly ranked candidate queries as the target query.

12. The method of claim 1, wherein the outputting of the target query comprises, for each of a set of candidate queries generated by the machine translation system as a translation of the input query in the target language, generating a features vector based on features extracted from the candidate query.

13. The method of claim 4, wherein the translating of the input query with the machine translation system comprises:

retrieving a set of biphrases, each biphrase comprising at least one word of the input query in the source language and at least one corresponding word in the target language; and

with a scoring function, computing a set of the retrieved biphrases to cover the input query, for each of the set of candidate queries, each candidate query comprising the corresponding words in the target language forming the set of retrieved biphrases.

14. The method of claim 1, wherein the information retrieval performance of each translated query is based on a relevance score for documents in a set of documents retrieved in response to the translated queries.

15. The method of claim 1, wherein the information retrieval performance of each translated query is determined with respect to at least one of the original query and a reference translation thereof in the target language.

16. The method of claim 1, wherein the method includes training the machine translation system, the training including learning feature weights for the features in the set.

17. The method of claim 16, wherein the learning of the feature weights is performed with the margin infused relaxed algorithm.

18. The method of claim 1, wherein at least one of the features is a feature which is not used in generating the translation of the input query.

19. A computer program product comprising a non-transitory computer-readable recording medium which stores instructions for performing the method of claim 1.

20. A translation system comprising non-transitory memory which stores instructions for performing the method of claim 1 and a processor in communication with the memory for executing the instructions.

21. A query translation system comprising:

a decoder which receives a source query in a source language and outputs a set of candidate queries in a target language, each of the candidate queries being a translation of the same source query; and

a reranking module which outputs a target query based on at least one of the candidate queries, the reranking module extracting features of each of the candidate queries and computing a function in which the extracted features are weighted by feature weights, the weights having been learned on features of each of a set of translated queries generated by translation of an original query into the target language and a measure of information retrieval performance of each the translated queries, for each original query in a set of original queries.

22. The query translation system of claim 21, wherein the decoder comprises a phrase-based statistical machine translation system.

23. The query translation system of claim 21, further comprising a processor which implements the reranking module.

25. A computer program product comprising a non-transitory computer-readable recording medium which stores instructions for performing the method of claim 23.

24. A method for training a translation system for translation of queries, comprising:

for each of a set of original queries in a source language;

translating the query to generate a set of translations in a target language;

for each translation in the set of translations, extracting values of features for each of a finite set of features; and

obtaining a measure of retrieval performance for each translation;

learning feature weights for each of the features based on the extracted values of the features and the respective measure of retrieval performance of each translation; and

storing the feature weights for use in translating a new query, different from each of the original queries, from the source language to the target language, whereby candidate translations of the new query are ranked based on their extracted values of features and the stored feature weights.

26. A computer system comprising non-transitory memory which stores instructions for performing the method of claim 24 and a processor in communication with the memory for executing the instructions.