Dynamic bi-phrases for statistical machine translation

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:

log

⁢

count

⁡

(

source

i

,

target

i

)

count

⁡

(

source

i

)

where count(source i ) is a number of instances of observing the source phrase source i in a bisentence of the training corpus and count(source i , target i ) is a number of instances of observing both source i and the corresponding target phrase target i in the same bisentence of the training corpus.

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,a expΣ i λ i h 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 .