Sequence classification for machine translation

• US 7,783,473 B2
• Filed: 12/28/2006
• Issued: 08/24/2010
• Est. Priority Date: 12/28/2006
• Status: Expired due to Fees

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: AT&T Intellectual Property II LP (AT&T, Inc.)
• Inventors: Kanthak, Stephan, Haffner, Patrick, Bangalore, Srinivas
• Primary Examiner(s): ALBERTALLI, BRIAN LOUIS
• Application Number:
  US11/647,080
• Publication Number:
  US 20080162111A1
• Time in Patent Office: 1,335 Days
• Field of Search: None
• US Class Current: 704/2
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models

Abstract

Classification of sequences, such as the translation of natural language sentences, is carried out using an independence assumption. The independence assumption is an assumption that the probability of a correct translation of a source sentence word into a particular target sentence word is independent of the translation of other words in the sentence. Although this assumption is not a correct one, a high level of word translation accuracy is nonetheless achieved. In particular, discriminative training is used to develop models for each target vocabulary word based on a set of features of the corresponding source word in training sentences, with at least one of those features relating to the context of the source word. Each model comprises a weight vector for the corresponding target vocabulary word. The weights comprising the vectors are associated with respective ones of the features; each weight is a measure of the extent to which the presence of that feature for the source word makes it more probable that the target word in question is the correct one.

Claims

1. A machine translation method of classifying source symbol sequences into target symbol sequences, the method comprising

classifying, by the machine, each symbol of the source sequence independently of the other symbols of the source sequence, the classifying being based on symbol models, each of at least ones of the symbol models being a function of training sequence context information, and

classifying, by the machine, the target symbol sequence based on the independently classified source sequence symbols.

2. The method of claim 1 wherein each symbol model is associated with a respective symbol of a target vocabulary and was generated as a function of contextual information about symbols in a plurality of training sequences.

3. The method of claim 2 wherein each symbol model is a respective set of weights each associated with respective ones of a plurality of feature definitions, at least one of the feature definitions defining a relationship between a given symbol in a given source sequence and one or more of the other symbols in the source sequence.

4. The method of claim 3 wherein said classifying each symbol of the source sequence independently of the other symbols of the source sequence comprises generating for at least ones of the target vocabulary symbols a respective probability, said probability for a given target vocabulary symbol being a function of the associated symbol model'"'"'s weights and being a further function of which of the feature definitions are met by the source symbol being classified.

5. A machine translation method of translating words in a source natural language sentence into corresponding words in a target natural language sentence, the method comprising

for a particular source sentence word, determining, by the machine, a probability for each one of a plurality of target vocabulary words, the probability being the probability that said each one target vocabulary word is the correct translation of said particular source sentence word, said probability being a function of a set of feature values and being a further function of a set of weights associated with said each one target vocabulary word, the feature values indicating which of a plurality of feature definitions are met by said particular source sentence word, at least one of the features being contextual information about said particular source sentence word, the weights each being associated with a respective one of the features, and

selecting, by the machine, a particular one of the target vocabulary words as being the correct translation of the source sentence as a function of the probabilities thus determined.

6. The method of claim 5 wherein the selected target vocabulary word is the target vocabulary word having the highest of said probabilities.

7. The method of claim 5 wherein each of said weights is a measure of the probability that a word in the source sentence translates to said each one of said target vocabulary words when the source sentence word has the feature in question.

8. The method of claim 5 wherein said probability is a function of the sum of the weights associated with feature definitions that are met by said particular source sentence word.

9. The method of claim 5 wherein the weights associated with said each one of said target vocabulary words are the result of discriminative training based on a) training sentences in the source language, b) the corresponding sentences in the target language, and c) alignment information indicating which words in each source language training sentence correspond to which words in the corresponding target language sentence.