System for parametric text to text language translation

• US 5,805,832 A
• Filed: 06/02/1995
• Issued: 09/08/1998
• Est. Priority Date: 07/25/1991
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Cocke, John, Mercer, Robert Leroy, Della Pietra, Vincent Joseph, Della Pietra, Stephen Andrew, Brown, Peter Fitzhugh, Lai, Jennifer Ceil, Jelinek, Frederick
• Primary Examiner(s): Hayes, Gail O.
• Assistant Examiner(s): HUGHET, WILLIAM N
• Application Number:
  US08/459,454
• Time in Patent Office: 1,194 Days
• Field of Search: 364/419.02, 364/419.1, 364/419.01, 364/419.08, 395/2.49, 395/2.86, 395/751, 395/752, 395/759
• US Class Current: 711/1
• CPC Class Codes: G06F 40/268 Morphological analysis
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models
• CPC Class Codes: G06F 40/49 using very large corpora, e.g. the web
• CPC Class Codes: G06F 40/55 Rule-based translation

Abstract

The present invention is a system for translating text from a first source language into a second target language. The system assigns probabilities or scores to various target-language translations and then displays or makes otherwise available the highest scoring translations. The source text is first transduced into one or more intermediate structural representations. From these intermediate source structures a set of intermediate target-structure hypotheses is generated. These hypotheses are scored by two different models: a language model which assigns a probability or score to an intermediate target structure, and a translation model which assigns a probability or score to the event that an intermediate target structure is translated into an intermediate source structure. Scores from the translation model and language model are combined into a combined score for each intermediate target-structure hypothesis. Finally, a set of target-text hypotheses is produced by transducing the highest scoring target-structure hypotheses into portions of text in the target language. The system can either run in batch mode, in which case it translates source-language text into a target language without human assistance, or it can function as an aid to a human translator. When functioning as an aid to a human translator, the human may simply select from the various translation hypotheses provided by the system, or he may optionally provide hints or constraints on how to perform one or more of the stages of source transduction, hypothesis generation and target transduction.

Claims

1. A text-to-text language translation system, comprising:

a computer processor;

a memory having stored therein a plurality of models, wherein said models are used in text-to-text translation, said plurality of models including;

a parametric translation model for generating a modeled translation probability, wherein said parametric translation model is generated with reference to a translation model source training text and a translation model target training text, said parametric translation model including a first specification of parameters, and

a parametric language model for generating a modeled probability, wherein said parametric language model is generated with reference to a language model training text, said parametric language model including a second specification of parameters; and

means for performing text-to-text language translation using said parametric translation model and said parametric language model.

2. A system according to claim 1, for training text in two parallel corpora, comprising:

means for grouping said source training text into a first plurality of tokens having sentence-like structures; and

means for grouping said target training text into a second plurality of tokens having sentence-like structures; and

means for performing sentence alignment by aligning said first and second pluralities of tokens according to their length.

3. A system according to claim 2, further comprising:

means for determining the probability associated with each of said sentence alignments.

4. A system according to claim 2, further comprising:

means for generating bigram statistics for said tokens; and

means for generating classes of similar elements based on said bigram statistics.

5. A system according to claim 1, further comprising:

means for determining the sum of probabilities for all alignments of units of linguistic structure for a given pair of source and target sentences.

6. A system according to claim 1, further comprising:

means for determining, for a given pair of source and target sentences, the most probable alignment of units of linguistic structure between the sentences.

7. A system according to claim 1, further comprising:

means for determining, for a given pair of source and target sentences and a given partial alignment of units of linguistic structure between the sentences, the most probable completion of said partial alignment.

8. A system according to claim 7, further comprising:

means for adjusting said most probable completion of said partial alignment by changing said alignment; and

means for determining the most probable completion of said new alignment.

9. A system according to claim 8, further comprising:

means for changing said alignment by performing at least one of a swapping and single move operation, said swapping comprising the interchanging of target units of linguistic structure assigned to any two source units of linguistic structure, and said single move comprising the changing of a target unit of linguistic structure assigned to any one source unit of linguistic structure.

10. A system according to claim 1, further comprising:

means for determining, for a given pair of source and target sentences and a completely specified alignment of units of linguistic structure between the sentences, the probability of said alignment.

11. A method for text-to-text language translation, comprising the steps of:

building a parametric translation model to generate a modeled translation probability, comprising the steps of,

storing a translation model source training text,

storing a translation model target training text, and

choosing a first specification of parameters for the translation model so that the modeled translation probability of the source and target training texts is a first unique local maximum value;

building a parametric language model to generate a modeled probability, comprising the steps of,

storing a language model raining text, and

choosing a second specification of parameters for the language model so that the modeled probability of the given training text is a second unique local maximum value; and

performing text-to-text language translation using said parametric translation model and said parametric language model.

12. A method according to claim 11, for training text in two parallel corpora, comprising the steps of:

grouping said translation model source training text into a first plurality of tokens having sentence-like structures;

grouping said translation model target training text into a second plurality of tokens having sentence-like structures;

performing sentence alignment by aligning said first and second pluralities of tokens according to their length.

13. A method according to claim 12, further comprising the step of:

determining a probability associated with each of said sentence alignments.

14. A method according to claim 12, further comprising the steps of:

generating bigram statistics for said tokens; and

generating classes of similar elements based on said bigram statistics.

15. A method according to claim 11, wherein said steps of storing said translation model source training text and said translation model target training text store a source language and an artificial language, respectively.

16. A method according to claim 11, further comprising the steps of:

performing alignment of units of linguistic structure of said source training text with unit of linguistic structure of said target training text.

17. A method according to claim 16, further comprising the step of:

determining a probability associated with each of said unit alignments.

18. A method according to claim 16, further comprising the step of:

determining a sum of probabilities for all alignments of units of linguistic structure for a given pair of source and target sentences.

19. A method according to claim 16, further comprising the step of:

determining, for a given pair of source and target sentences, the most probable alignment of units of linguistic structure between the sentences.

20. A method according to claim 16, further comprising the step of:

determining, for a given pair of source and target sentences and a given partial alignment of units of linguistic structure between the sentences, the most probable completion of said partial alignment.

21. A method according to claim 20, further comprising the step of:

adjusting said most probable completion of said partial alignment by changing said alignment; and

determining the most probable completion of said new alignment.

22. A method according to claim 21, further comprising the step of:

changing said alignment by performing at least one of a swapping and single move operation, said swapping comprising the interchanging of target units of linguistic structure assigned to any two source units of linguistic structure, and said single move comprising the changing of a target unit assigned to any one source unit.

23. A method according to claim 16, further comprising the step of:

determining, for a given pair of source and target sentences and a completely specified alignment of units of linguistic structure between the sentences, the probability of said alignment.

24. A method for translating a first text in a first language into a second text in a second language using a lexical model, comprising the steps of:

inputting the first text into the lexical model, wherein the lexical model comprises a parametric translation model for generating a first probability and a parametric language model for generating a second probability; and

determining, using the lexical model, the second text in the second language that yields a unique local maximum value of a product of the first probability of the parametric translation model and the second probability of the parametric language model.