Machine translation using elastic chunks

• US 7,542,893 B2
• Filed: 05/10/2006
• Issued: 06/02/2009
• Est. Priority Date: 05/10/2006
• Status: Expired due to Fees

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Cancedda, Nicola, Dymetman, Marc, Goutte, Cyril, Gaussier, Eric
• Primary Examiner(s): Vo; Huyen X.
• Application Number:
  US11/431,393
• Publication Number:
  US 20070265825A1
• Time in Patent Office: 1,119 Days
• Field of Search: 704 1- 10, 704/277, 704/255, 704/257, 704/251, 704/252
• US Class Current: 704/2
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models

Abstract

A machine translation method includes receiving source text in a first language and retrieving text fragments in a target language from a library of bi-fragments to generate a target hypothesis. Each bi-fragment includes a text fragment from the first language and a corresponding text fragment from the second language. Some of the bi-fragments are modeled as elastic bi-fragments where a gap between words is able to assume a variable size corresponding to a number of other words to occupy the gap. The target hypothesis is evaluated with a translation scoring function which scores the target hypothesis according to a plurality of feature functions, at least one of the feature functions comprising a gap size scoring feature which favors hypotheses with statistically more probable gap sizes over hypotheses with statically less probable gap sizes.

Claims

1. A machine translation method for translating source text from a first language to target text in a second language, comprising:

receiving the source text in the first language;

accessing a library of bi-fragments stored in computer readable memory, each of the bi-fragments including a text fragment from the first language and a text fragment from the second language, at least some of the bi-fragments being modeled as elastic bi-fragments in which words of a fragment are spaced by a gap which is able to assume a variable size corresponding to a number of other words which are to occupy the gap;

retrieving text fragments from the second language from the library corresponding to text fragments in the source text;

generating, by a processor, at least one target hypothesis, each of said target hypotheses comprising text fragments selected from the second language; and

evaluating the target hypothesis with a translation scoring function which scores the target hypothesis according to a plurality of feature functions, at least one of the feature functions comprising a gap size scoring feature which favors hypotheses with statistically more probable gap sizes over hypotheses with statically less probable gap sizes.

2. The method of claim 1, wherein bi-fragments comprising contiguous text fragments are modeled as elastic bi-phrases.

3. The method of claim 1, wherein the generation of a target hypothesis includes generating a partial hypothesis comprising a sequence of words without gaps from the text fragments selected from the second language and storing any unconsumed words of the text fragments used in generating the sequence in a buffer.

4. The method of claim 3, wherein the generation of a hypothesis includes expanding an existing partial hypothesis by at least one of:

adding a word of a new text fragment to the sequence of words; and

retrieving an unconsumed word from the buffer to the sequence of words.

5. The method of claim 1, wherein the gap size scoring feature models gap sizes observed in a training corpus as a distribution which is expressed in terms of a parameter of the distribution.

6. The method of claim 5, wherein the distribution comprises a Poisson distribution and the parameter of the distribution comprises a mean value of the Poisson distribution.

7. The method of claim 5,wherein the distribution ignores gap sizes which are determined to constitute outliers.

8. The method of claim 1, wherein each elastic bi-fragment is associated in the library with at least one parameter of a distribution function which is representative of a distribution of acceptable gap sizes for the elastic bi-phrase, the gap size scoring feature utilizing the parameters of the selected text fragments in the evaluation of the hypothesis.

9. The method of claim 1, wherein the generating of a target hypothesis includes performing a multiple stack beam-search.

10. The method of claim 1, wherein the generating of at least one target hypothesis includes generation of a plurality of partial translation hypotheses, scoring the partial hypotheses, and optionally pruning partial translation hypotheses according to their scores.

11. The method of claim 10, wherein the scoring of the partial hypotheses favors adding source words to a partial translation in which a gap exists in an elastic fragment until a mean gap size of a distribution for the elastic fragment is reached and thereafter favors minimizing further increase to the gap size in the partial translation.

12. The method of claim 1, wherein the evaluation of the at least one hypothesis includes selecting a hypothesis from a plurality of hypotheses which maximizes a log-linear scoring function in which weights are assigned to each of the feature functions.

13. The method of claim 12, wherein the log-linear scoring function further includes a bi-fragment feature h_bf which estimates a log of the probability of observing a bi-fragment (s,t) given the source phrase s from a ratio of the occurrences of the bi-fragment (s,t) and source phrase s in a training corpus, without consideration to the size of any gaps between words in the bi-fragment and wherein the gap size scoring feature h_ELAST estimates a probability for gap sizes assigned to (s,t) in the translation, relative to a gap size distribution associated with the bi-fragment (s,t) in the training corpus.

14. The method of claim 1, wherein the translation scoring function utilizes a log-linear scoring model of the form:

$\begin{array}{cc}\mathrm{Pr}<br><br>\left(t_1^J,d_1^K❘<br><br>s_1^I\right)=\frac{1}{Z_{f_1^I}}<br><br>\exp <br><br>\underset{m=1}{\overset{M}{\sum <br><br>}}<br><br>{\mathrm{\lambda <br><br>}}_m<br><br>h_m<br><br>\left(s_1^I,t_1^J,d_1^K\right)& \left(2\right)\end{array}$

to represent the conditional probability of a target sentence t₁^J and an “

alignment”

d₁^K given a source sentence s₁^I, where each “

feature function”

h_m scores a particular aspect of the relationship between s₁^I and t₁^J, a contribution of each feature function being determined by a weight λ

_m; and

where z_{f 1}^I is a normalization constant and where at least one of the feature functions h_m is selected from;

(1) a “

bi-fragment”

feature h_bf, of the general form;

$\begin{array}{cc}{h}_{\mathrm{bf}}<br><br>\left(s_1^I,t_1^J,d_1^K\right)=\underset{k=1}{\overset{K}{\sum <br><br>}}<br><br>\log <br><br><br><br>\mathrm{Pr}<br><br>\left(\mathrm{target}<br><br>\left(d_k\right)❘<br><br>\mathrm{source}<br><br>\left(d_k\right)\right),& \left(3\right)\end{array}$

where Pr(target(d_k)|source(d_k)) represents a conditional probability of a target fragment given a source fragment, irrespective of their gap sizes; and

(2) a gap size scoring feature (h_ELAST), of the general form;

$h_{\mathrm{ELAST}}<br><br>\left(s_1^I,t_1^J,d_1^K\right)=\underset{k=1}{\overset{K}{\sum <br><br>}}<br><br>\mathrm{EL}<br><br>\left(\mathrm{target}<br><br>\left(d_k\right),\mathrm{source}<br><br>\left(d_k\right)\right)$

 

where EL represents a logarithm of a probability of observing a certain configuration of gap sizes in the target and source fragments of a bi-fragment.

15. The method of claim 14, wherein the translation scoring function outputs a translation for which the log-linear scoring model is maximized.

16. A machine translation system for translating source text from a first language to target text in a second language, comprising:

memory which stores instructions for performing the method of claim 1; and

a processor which executes the instructions.

22. An article of manufacture comprising a program storage medium readable by a computer comprising instructions for executing the method of claim 1.

17. A machine translation system for translating source text from a first language to target text in a second language, comprising:

a library of bi-fragments, each of the bi-fragments including a text fragment from the first language and a text fragment from the second language, at least some of the bi-fragments being modeled as elastic bi-fragments in which words of a fragment are spaced by a gap which is able to assume a variable size corresponding to a number of other words which are to occupy the gap, each of the elastic bi-fragments being associated in memory with a parameter representative of a gap size distribution for the bi-fragment; and

a processor which executes instructions for retrieving text fragments from the second language from the library corresponding to text fragments in the source text, generating at least one target hypothesis, each of said target hypotheses comprising text fragments selected from the second language, and evaluating the hypothesis with a translation scoring function which scores the hypothesis according to a plurality of features, at least one of the features comprising a gap size scoring feature which favors hypotheses comprising text fragments with statistically more probable gap sizes over hypotheses with statically less probable gap sizes.

18. The machine translation system of claim 17, wherein translation scoring function further includes at least one of:

a second feature function which evaluates bi-fragment discontinuities; and

a third feature function which evaluates bi-fragment reorderings.

19. The machine translation system of claim 18, wherein the second feature function counts gaps in the target text of the hypothesis and corresponding source text.

20. The machine translation system of claim 17, wherein the library associates each of a plurality of elastic bi-fragments with a corresponding parameter representative of a probability distribution of gaps in at least one of the source text fragment and the target text fragment.

21. The translation system of claim 17, wherein the processor includes instructions for generation of a hypothesis which includes expanding an existing partial hypothesis by at least one of:

adding a word of a new text fragment to the sequence of words; and

retrieving an unconsumed word from the buffer to the sequence of words.