System and method for estimating accuracy of an automatic natural language translation

• US 6,285,978 B1
• Filed: 09/24/1998
• Issued: 09/04/2001
• Est. Priority Date: 09/24/1998
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Medeiros, Sue Ann, Bernth, Arendse, McCord, Michael Campbell, Gdaniec, Claudia Maria
• Primary Examiner(s): Thomas, Joseph
• Application Number:
  US09/159,966
• Time in Patent Office: 1,076 Days
• Field of Search: 704/2, 704/3, 704/4, 704/5, 704/6, 704/7, 704/8, 704/277, 707/536
• US Class Current: 704/7
• CPC Class Codes: G06F 40/51 Translation evaluation

Abstract

A computer system and method for natural language translation uses a translation process to translate a source natural language segment (e.g. English) of one or more source words/elements into a target natural language (e.g. German) segment of one or more target words/elements. An evaluation module determines a confidence measure of the natural language translation. Typically, the confidence measure indicates less confidence as the complexity of the translation increases. Various novel features for determining complexity and confidence measure at different steps in the translation are used. The translation process can be terminated if the confidence measure fails to meet a threshold criterion.

Claims

1. A computer system for natural language translation, comprising:

a parser that parses a source natural language segment into a source parse structure with one or more source elements;

a segment translation process that translates the source parse structure into a target natural language segment through a target language structure; and

an evaluation process that determines a translation complexity of the computer system translating the source natural language segment to the target natural language segment and determines a confidence measure related to a parsing complexity that is a complexity of parsing the source natural language segment, where the confidence measure indicates less confidence as the parsing complexity increases.

2. A computer system for natural language translation, as in claim 1, where the segment translation process terminates if the confidence measure fails to meet a threshold criterion.

3. A computer system for natural language translation, as in claim 1, where the evaluation process determines the translation complexity by measuring a complexity of choices during the translating.

4. A computer system for natural language translation, as in claim 3, where the translation process terminates if the confidence measure fails to meet a threshold criterion.

5. A computer system for a natural language translation, comprising:

a translation process that translates a source natural language segment of one or more source words into a target natural language segment of one or more target words;

a parser that parses the source natural language segment into a source parse structure with one or more source elements including one or more syntactic relationships among the source elements; and

an evaluation process that determines a translation complexity of the computer system translating from the source natural language segment to the target natural language segment by measuring a complexity of choices during the translating and measuring a parsing complexity of the parsing, and further determines a confidence measure of the natural language translation by a complexity combination of the complexity of choices and the parsing complexity, the confidence measure indicating less confidence as the complexity combination increases.

6. A computer system for natural language translation, as in claim 5, where the translation terminates if the confidence measure fails to meet a threshold criterion.

7. A computer system for natural language translation, comprising:

a parser that parses a source natural language segment of one or more source words into a source parse structure with one or more source elements including syntactic relationships among the source elements;

a source evaluation process that generates a source indication of a complexity of choices in producing the source parse structure;

a transfer process that converts the source parse structure into an initial transfer structure in which source elements are replaced by target elements and target relationships are created;

a transfer evaluation process that generates a transfer indication of a complexity of choices in producing the initial transfer structure;

a structural process that converts the initial transfer structure into a target language structure;

a structural transfer evaluation process that generates a structural indication of a complexity of converting the initial transfer structure into the target language structure;

a target generation process that converts the target language structure into a target natural language segment of one or more target words;

a target generation evaluation process that generates a generation indication of a complexity of the converting the target language structure into the target natural language segment; and

a combiner that aggregates any one or more of the source indication, the transfer indication, the structural indication, and the generation indication to produce a single measure of the complexity of the translation by the computer system and determines a confidence measure of a natural language translation that indicates less confidence as the complexity increases.

8. A computer system for natural language translation, as in claim 7, where the translation process terminates if the confidence measure fails to meet a threshold criterion.

9. A system, as in claim 7, where the source indication includes a measure of a complexity of a punctuation.

10. A system, as in claim 7, where the source indication includes a measure of a complexity of an abbreviation.

11. A system, as in claim 7, where the source indication includes a measure of a complexity of a footnote.

12. A system, as in claim 7, where the source indication includes a measure of a complexity of a number of lexical analyses per word.

13. A system, as in claim 7, where the source indication includes a measure of a complexity of a number of different parts of speech.

14. A system, as in claim 7, where the source indication includes a measure of a complexity of an ambiguous combination of parts of speech.

15. A system, as in claim 7, where the source indication includes a measure of an ambiguous construction or ungrammatical construction of the source natural language.

16. A system, as in claim 7, where the source indication includes a measure of a complexity of a for-to construction.

17. A system, as in claim 7, where the source indication includes a measure of a complexity of a preposition without prepositional object.

18. A system, as in claim 7, where the source indication includes a measure of a complexity of a time reference.

19. A system, as in claim 7, where the source indication includes a measure of a complexity of a coordination.

20. A system, as in claim 7, where the source indication includes a measure of a complexity of a parse evaluation score.

21. A system, as in claim 7, where the source indication includes a measure of the amount of a time and a space usage.

22. A system, as in claim 7, where the source indication includes a measure of a complexity of a sentence length.

23. A system, as in claim 7, where the transfer indication includes a measure of a complexity of a lexical transfer.

24. A system, as in claim 7, where the transfer indication includes a measure of the relevance of a lexicon used for target word choice.

25. A system, as in claim 7, where the transfer indication includes a measure of a time used during transfer.

26. A system, as in claim 7, where the structural indication includes a measure of a complexity of a transformation.

27. A system, as in claim 7, where the generation indication includes a measure of a complexity of a target generation.

28. A method executed on a computer system for natural language translation, comprising the steps of:

parsing a source natural language segment into a source parse structure with one or more source elements;

translating the source parse structure into a target natural language segment through a target language structure;

determining a translation complexity of translating the source natural language segment to the target natural language segment; and

determining a confidence measure related to a parsing complexity that is a complexity of parsing the source natural language segment, where the confidence measure indicates less confidence as the parsing complexity increases.

29. A computer system for natural language translation, comprising:

means for parsing a source natural language segment into a source parse structure with one or more source elements;

means for translating the source parse structure into a target natural language segment through a target language structure;

means for determining a translation complexity of translating the source natural language segment to the target natural language segment; and

means for determining a confidence measure related to a parsing complexity that is a complexity of parsing the source natural language segment, where the confidence measure indicates less confidence as the parsing complexity increases.