MACHINE TRANSLATION IN CONTINUOUS SPACE

• US 20090248394A1
• Filed: 03/25/2008
• Published: 10/01/2009
• Est. Priority Date: 03/25/2008
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Deng, Yonggang, Kingsbury, Brian Edward Doorenbos, Gao, Yuqing, Sarikaya, Ruhi
• Granted Patent: US 8,229,729 B2
• Time in Patent Office: Days
• US Class Current: 704/4
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models

Abstract

A system and method for training a statistical machine translation model and decoding or translating using the same is disclosed. A source word versus target word co-occurrence matrix is created to define word pairs. Dimensionality of the matrix may be reduced. Word pairs are mapped as vectors into continuous space where the word pairs are vectors of continuous real numbers and not discrete entities in the continuous space. A machine translation parametric model is trained using an acoustic model training method based on word pair vectors in the continuous space.

Claims

1. A method for training a statistical machine translation model, comprising:

creating a source word versus target word co-occurrence matrix to define word pairs;

reducing dimensionality of the matrix;

mapping word pairs as vectors into continuous space where the word pairs are vectors of continuous real numbers and not discrete entities in the continuous space; and

training a machine translation parametric model using an acoustic model training method based on word pair vectors in the continuous space.

2. The method as recited in claim 1, further comprising:

clustering the vectors based upon at least one of source words and target words.

3. The method as recited in claim 1, wherein the source word versus target word co-occurrence matrix includes phrases, phrase pairs and word to phrase pairs.

4. The method as recited in claim 1, wherein reducing includes reducing dimensionality of the matrix by employing one of singular-value decomposition (SVD) and non-negative matrix factorization (NMF).

5. The method as recited in claim 1, wherein mapping includes considering word-pairs as phones and phrase-pairs as words.

6. The method as recited in claim 5, wherein pronunciation of a phrase-pair is defined by different word alignments of a source phrase and a target phrase.

7. The method as recited in claim 1, wherein mapping word pairs as vectors includes aligning a source language word to any other target language word for a given sentence pair.

8. The method as recited in claim 1, wherein training includes training the parametric model using at least one of Maximum Likelihood (ML), Maximum-Mutual Information Estimation (MMIE), Minimum Phone Error (MPE) training, Max-margin Training, and Minimum Classification Error (MCE) training.

9. The method as recited in claim 1, wherein training includes adapting the model to at least one of a new domain, language, genre, and speaker.

10. A computer readable medium comprising a computer readable program for training a statistical machine translation model, wherein the computer readable program when executed on a computer causes the computer to perform the steps of:

creating a source word versus target word co-occurrence matrix to define word pairs;

reducing dimensionality of the matrix;

mapping word pairs as vectors into continuous space where the word pairs are vectors of continuous real numbers and not discrete entities in the continuous space; and

training a machine translation parametric model using an acoustic model training method based on word pair vectors in the continuous space.

11. A method for machine translation, comprising:

adapting a parametric model in accordance with a given state, the model including mapped word pairs as vectors in continuous space where the word pairs are vectors of continuous real numbers and not discrete entities in the continuous space, and the model being trained using an acoustic model training method based on word pair vectors in the continuous space;

computing a pair-probability of a word or phrase based on word-pair translation probabilities using the parametric model; and

translating the word or phrase in accordance with a machine translation search that employs the translation probabilities.

12. The method as recited in claim 11, wherein the translation probabilities are combined with at least one of a language model score, and another score to translate the word or phrase.

13. The method as recited in claim 11, wherein adapting includes adapting to at least one of a new domain, genres, speaker, and language

14. The method as recited in claim 11, wherein adapting includes adapting using at least one of maximum likelihood linear regression (MLLR), maximum a posteriori (MAP), feature MLLR (fMLLR), and feature Minimum Phone Error (fMPE).

15. A computer readable medium comprising a computer readable program for machine translation, wherein the computer readable program when executed on a computer causes the computer to perform the steps of:

adapting a parametric model in accordance with a given state, the model including mapped word pairs as vectors in continuous space where the word pairs are vectors of continuous real numbers and not discrete entities in the continuous space, and the model being trained using an acoustic model training method based on word pair vectors in the continuous space;

computing a pair-probability of a word or phrase based on word-pair translation probabilities using the parametric model; and

translating the word or phrase in accordance with a machine translation search that employs the translation probabilities.

16. The computer readable medium as recited in claim 15, wherein the translation probabilities are combined with at least one of a language model score and another score to translate the word or phrase.

17. The computer readable medium as recited in claim 15, wherein adapting includes adapting to at least one of a new domain, genres, speaker, and language

18. The computer readable medium as recited in claim 15, wherein adapting includes adapting using at least one of maximum likelihood linear regression (MLLR), maximum a posteriori (MAP), feature MLLR (fMLLR), and feature Minimum Phone Error (fMPE).