Method and apparatus for providing stochastic finite-state machine translation

• US 7,113,903 B1
• Filed: 01/30/2002
• Issued: 09/26/2006
• Est. Priority Date: 01/30/2001
• Status: Expired due to Term

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: AT&T Corporation (AT&T, Inc.)
• Inventors: Riccardi, Giuseppe, Bangalore, Srinivas
• Primary Examiner(s): ABEBE, DANIEL DEMELASH
• Application Number:
  US10/058,995
• Time in Patent Office: 1,700 Days
• Field of Search: 704/2, 704/4
• US Class Current: 704/4
• CPC Class Codes: G06F 40/289 Phrasal analysis, e.g. finite state techniques or chunking
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks

Abstract

A method and apparatus for stochastic finite-state machine translation is provided. The method may include receiving a speech input and translating the speech input in a source language into one or more symbols in a target language based on stochastic language model. Subsequently, all possible sequences of the translated symbols may be generated. One of the generated sequences may be selected based on a monolingual target language model.

Claims

1. A method for providing stochastic finite-state machine translation comprising:

receiving an input in a source language;

translating and decomposing the source language input into one or more symbols in a target language based on stochastic translation model;

generating a plurality of possible sequences of the translated symbols; and

selecting one of the generated sequences based on a monolingual target language model, wherein the selected sequence represents a most likely translation of the source language input

wherein decomposing the source language input into a plurality of symbols further includes decomposing the source language input into phrase level symbols based on joint entropy reduction of the source and target languages.

2. A method for providing stochastic finite-state machine translation comprising:

receiving an input in a source language;

translating and decomposing the source language input into one or more symbols in a target language based on stochastic translation model;

generating a plurality of possible sequences of the translated symbols; and

selecting one of the generated sequences based on a monolingual target language model, wherein the selected sequence represents a most likely translation of the source language input

further including determining alignment information by mapping the symbols in the source language onto the symbols in the target language.

3. The method of claim 2, wherein translating the source language input into one or more symbols in the target language further includes transforming the alignment information into a bi-language string.

4. The method of claim 3, wherein translating the source language input into one or more symbols in the target language further includes training the bi-language string based on a Variable N-gram Stochastic Automaton (VNSA) learning using stochastic transducers.

5. The method of claim 4, wherein translating the source language input into one or more symbols in the target language further includes training the VNSA trained bi-language string based on a bi-phrase learning algorithm using a bi-language corpus.

6. An apparatus for stochastic finite-state machine translation comprising:

a recognizer for receiving an input in a source language;

a lexical selection unit connected to the recognizer for translating the input in the source language into one or more symbols in a target language with reference to a first model database;

a reordering unit connected to the lexical selection unit for generating a plurality of possible sequences of the translated symbols; and

a refining unit connected to the reordering unit for selecting one of the possible sequences of the translated symbols with reference to a second model database,

wherein

the first model database is a stochastic translation model database, and the second model database is a monolingual target language model database,

the recognizer decomposes the input into a plurality of symbols in the source language, and

the recognizer decomposes the input in the source language into phrase level symbols based on joint entropy reduction of the source and target languages.

7. An apparatus for stochastic finite-state machine translation comprising:

a recognizer for receiving an input in a source language;

a lexical selection unit connected to the recognizer for translating the input in the source language into one or more symbols in a target language with reference to a first model database;

a reordering unit connected to the lexical selection unit for generating a plurality of possible sequences of the translated symbols; and

a refining unit connected to the reordering unit for selecting one of the possible sequences of the translated symbols with reference to a second model database,

wherein

the first model database is a stochastic translation model database, and the second model database is a monolingual target language model database,

the recognizer decomposes the input into a plurality of symbols in the source language, and

the lexical selection unit determines an alignment information by mapping the plurality of symbols in the source language into the symbols in the target language.

8. The apparatus of claim 7, wherein the reordering unit generates all possible sequences of the translated symbol based on a Variable Length N-gram Stochastic Automaton model.

9. A method for providing stochastic finite-state machine translation comprising:

receiving an input in a source language;

decomposing the input into a plurality of symbols in the source language;

translating the plurality of symbols into a plurality of symbols in a target language symbol by symbol in accordance with stochastic translation model;

generating a plurality of possible sequences of the symbols in the target language using a Variable N-gram Stochastic Automation algorithm; and

selecting one of the possible sequences based on a monolingual target language model.

10. The method of claim 9, wherein the input is decomposed into the plurality of symbols in the source language based on joint entropy reduction of the source and target languages.

11. The method of claim 10, further including determining an alignment information by mapping each of the plurality of symbols in the source language onto the plurality of symbols in the target language.