Dynamic bi-phrases for statistical machine translation

• US 9,552,355 B2
• Filed: 05/20/2010
• Issued: 01/24/2017
• Est. Priority Date: 05/20/2010
• Status: Expired due to Fees

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Aziz, Wilker Ferreira, Dymetman, Marc, Cancedda, Nicola, Coursimault, Jean-Marc, Nikoulina, Vassilina, Specia, Lucia
• Primary Examiner(s): WOZNIAK, JAMES S
• Application Number:
  US12/784,040
• Publication Number:
  US 20110288852A1
• Time in Patent Office: 2,441 Days
• Field of Search: 704/2, 704/4, 704/5, 704/9
• US Class Current: 1/1
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models
• CPC Class Codes: G06F 40/45 Example-based machine translation; Alignment

Abstract

A system and a method for phrase-based translation are disclosed. The method includes receiving source language text to be translated into target language text. One or more dynamic bi-phrases are generated, based on the source text and the application of one or more rules, which may be based on user descriptions. A dynamic feature value is associated with each of the dynamic bi-phrases. For a sentence of the source text, static bi-phrases are retrieved from a bi-phrase table, each of the static bi-phrases being associated with one or more values of static features. Any of the dynamic bi-phrases which each cover at least one word of the source text are also retrieved, which together form a set of active bi-phrases. Translation hypotheses are generated using active bi-phrases from the set and scored with a translation scoring model which takes into account the static and dynamic feature values of the bi-phrases used in the respective hypothesis. A translation, based on the hypothesis scores, is then output.

Claims

1. A method for phrase-based translation comprising:

receiving an input of source text in a source language to be translated into target text in a target language;

providing at least one dynamic bi-phrase rule to be used in generation of dynamic bi-phrases for translation of the source text;

for a sentence of the source text;

after receiving the source text, applying the at least one rule to the source text to generate a dynamic bi-phrase;

associating a value of at least one dynamic feature with the at least one dynamic bi-phrase;

retrieving static bi-phrases from a static bi-phrase table stored in memory which each include at least one word of the source text, each of the static bi-phrases being associated with a value of at least one static feature, each static bi-phrase in the bi-phrase table including a pair of phrases, each phrase comprising a sequence of at least one word, one of the phrases in the pair being from the source language and the other phrase being from the target language, the static bi-phrases having been automatically extracted from a training corpus of bisentences;

retrieving any of the dynamic bi-phrases which each cover at least one word of the source text, the retrieved static bi-phrases and dynamic bi-phrases forming a set of active bi-phrases;

generating translation hypotheses for at least a part of the source sentence using active bi-phrases from the set;

scoring the translation hypotheses with a translation scoring model which takes into account the static feature values of static bi-phrases in the hypothesis and dynamic feature values of dynamic-bi-phrases in the hypothesis, wherein in scoring of a hypothesis, the scoring model considers static bi-phrases in the active set of bi-phrases in which the source phrase of the respective static bi-phrase covers at least one source word of the source sentence, wherein the static bi-phrase and a dynamic bi-phrase in the active set of bi-phrases both cover a same source word, such that each source word of the hypothesis is covered by exactly one of the retrieved bi-phrases in the active set of bi-phrases; and

outputting a translation of the source text sentence based on the scoring of the hypotheses;

wherein at least one of the generating at least one dynamic bi-phrase, associating at least one dynamic feature value of a dynamic feature with the at least one dynamic bi-phrase, retrieving static bi-phrases from a static bi-phrase table stored in memory, retrieving any of the dynamic bi-phrases which each cover at least one word of the source text, generating translation hypotheses for at least a part of the source sentence using active bi-phrases from the set, and scoring the translation hypotheses with a translation scoring model is implemented by a computer processor.

2. The method of claim 1, wherein the active static bi-phrases are accorded feature values for the at least one dynamic feature and wherein the at least one active dynamic bi-phrase is accorded a feature value for the at least one static feature.

3. The method of claim 2, wherein the scoring of the translation hypotheses takes into account the active static bi-phrase feature values for the at least one dynamic feature and the at least one active dynamic bi-phrase feature value for the at least one static feature.

4. The method of claim 1, wherein the providing includes providing at least one dynamic bi-phrase generating rule which is configured for generating a dynamic bi-phrase when an instance of a pattern that is able to be instantiated by many possible bi-phrases is observed in a source sentence, the generating of the at least one dynamic bi-phrase including generating a dynamic bi-phrase when one of the instances of the pattern is observed in the source sentence.

5. The method of claim 1, wherein the dynamic bi-phrase generating rule is generated in response to a bi-phrase description input by a user.

6. The method of claim 1, wherein the dynamic bi-phrase is a generic bi-phrase for which instances of the generic bi-phrase are accorded the same dynamic feature values.

7. The method of claim 1, wherein a generic bi-phrase feature value for each of a set of types T_i of dynamic bi-phrase is computed as a function of:

8. The method of claim 7, wherein the count(source_i, target_i) is independent of an alignment between the sentences forming the bisentence.

9. The method of claim 1, wherein when the source text includes an out of vocabulary word, the method includes proposing dynamic bi-phrases to a user which are derived from a training corpus of bisentences using textual entailment techniques.

10. The method of claim 9, wherein the at least one feature value for the dynamic bi-phrase is derived from at least one static bi-phrase comprising an entailed source phrase.

11. The method of claim 1, wherein the at least one dynamic feature is accorded a weight which is taken into account by the scoring model.

12. The method of claim 11, wherein the weight is derived from a development corpus which has been adapted to include examples of bi-sentences in which the dynamic bi-phrase occurs.

13. The method of claim 12, wherein when the source phrase is an out of vocabulary word, the weight is derived from the weight for at least one static bi-phrase.

14. The method of claim 1, wherein the generation of the dynamic bi-phrase includes generating a generic bi-phrase rule comprising a source phrase, target phrase pattern which is fired when an instance of the source phrase of the pattern is observed in the input source sentence, the pattern being selected from:

a) a unit conversion pattern where the source phrase includes a variable amount expressed in a first type of units and the target phrase includes a variable amount expressed in a second type of units with an expression for conversion of the value in the first type of units to the value in the second type of units,

b) a date conversion pattern where a date in one format is converted to the same date in another format, and

c) a numerical expression conversion pattern in which a comma intermediate two digits is replaced with a period between the same two digits.

15. The method of claim 1, wherein each of the static bi-phrases is associated with a value of at least one static feature based on a frequency of the static bi-phrase in the training corpus of bisentences and wherein at least one of the at least one dynamic bi-phrases has no frequency in the training corpus of bisentences on which to base a feature value.

16. The method of claim 1, wherein at least one of the at least one dynamic bi-phrase is based on an output of a rule based machine translation system.

17. The method of claim 1, wherein the at least one dynamic feature value comprises at least two dynamic feature values.

18. The method of claim 1, wherein the translation scoring model is a log linear model of the general form:

(t, a)*=argmax_t,aexpΣ

_iλ

_ih_i(s, a, t) 

 

(Eqn.

     1)

where s is a source text string,

t is a target text string,

a is an alignment of the source text string and target text string,

each h_i is a value of a feature, selected from the static features and the at

least one dynamic feature, and

each λ

_i is a weight applied to the respective feature.

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

20. A non-transitory computer program product comprising a non-transitory recording medium encoding instructions, which when executed on a computer causes the computer to perform the method of claim 1.

21. A system for phrase-based translation comprising:

memory which receives an input of source text in a source language to be translated into target text in a target language;

a static bi-phrase table stored in memory, each of the static bi-phrases in the table being associated with at least one static feature value, each static bi-phrase in the bi-phrase table including a pair of phrases, each phrase comprising a sequence of at least one word, one of the phrases in the pair being from the source language and the other phrase being from the target language, the static bi-phrases having been automatically extracted from a training corpus of bisentences;

a dynamic bi-phrase generator which associates at least one dynamic feature value with at least one dynamic bi-phrase which has been selected for use in translation of the source text into the target language; and

a translation scoring model which is input with hypotheses built from an active set of bi-phrases and scores the hypotheses, the active set including a static bi-phrase covering at least one word of the source text and a dynamic bi-phrase covering the same at least one word, whereby the static bi-phrase and dynamic bi-phrase both cover a same source word of the source text, the model taking into account the static feature values of static bi-phrases that each include at least one of the source words in each of the hypotheses and dynamic feature values of any dynamic bi-phrases in each respective hypothesis.

22. The system of claim 21, wherein the dynamic bi-phrase generator computes weights for the dynamic bi-phrases used in the hypotheses.

23. The system of claim 21, wherein the scoring model does not exclude consideration of static bi-phrases which cover at least one source word covered by the dynamic bi-phrase in scoring of the hypotheses.

24. A method for phrase-based translation comprising:

providing a static bi-phrase table, each of the static bi-phrases being associated with a value of at least one static feature based on a frequency of the static bi-phrase in a training corpus of bisentences, each static bi-phrase in the bi-phrase table including a pair of phrases, each phrase comprising a sequence of at least one word, one of the phrases in the pair being from the source language and the other phrase being from the target language;

receiving an input of source text in a source language to be translated into target text in a target language;

after providing the static bi-phrase table, for a sentence of the source text;

applying at least one rule which, when fired, generates at least one respective dynamic bi-phrase based on the source text in the sentence, the rule comprising a source phrase, target phrase pattern which is able to be instantiated by different bi-phrases and which is fired when an instance of the source phrase of the pattern is observed in the input source sentence;

associating a respective value of at least one dynamic feature with the at least one dynamic bi-phrase;

retrieving static bi-phrases from the static bi-phrase table which each cover at least one word of the source text;

combining the retrieved static bi-phrases and any generated dynamic bi-phrases to form a set of active bi-phrases;

when the active set includes a static bi-phrase covering at least one word of the source text and a dynamic bi-phrase covering the same at least one word whereby the static bi-phrase and dynamic bi-phrase both cover a same source word of the source text, generating translation hypotheses in the target language for at least a part of the source sentence using active bi-phrases drawn from the set of active bi-phrases, where each word in the hypothesis is covered by exactly one of the active bi-phrases;

scoring the translation hypotheses with a translation scoring model which takes into account the static feature values of static bi-phrases in the hypothesis and dynamic feature values of dynamic-bi-phrases in the hypothesis; and

outputting a translation of the source text sentence in the target language based on the scoring of the hypotheses;

wherein at least one of the applying, associating, retrieving, combining, generating, and scoring is implemented by a computer processor.

25. A non-transitory computer program product comprising a non-transitory recording medium encoding instructions, which when executed on a computer causes the computer to perform the method of claim 24.