Learning translation relationships among words

• US 7,366,654 B2
• Filed: 01/13/2006
• Issued: 04/29/2008
• Est. Priority Date: 06/20/2001
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Moore, Robert C.
• Primary Examiner(s): Dorvil; Richemond
• Assistant Examiner(s): SAINT CYR, LEONARD
• Application Number:
  US11/331,537
• Publication Number:
  US 20060116867A1
• Time in Patent Office: 837 Days
• Field of Search: 704/2, 704/9
• US Class Current: 704/2
• CPC Class Codes: G06F 40/40 Processing or translation of natural language natural language analysis G06F40/20; semantic analysis G06F40/30
• CPC Class Codes: G06F 40/45 Example-based machine translation; Alignment

Abstract

A parallel bilingual training corpus is parsed into its content words. Word association scores for each pair of content words consisting of a word of language L1 that occurs in a sentence aligned in the bilingual corpus to a sentence of language L2 in which the other word occurs. A pair of words is considered “linked” in a pair of aligned sentences if one of the words is the most highly associated, of all the words in its sentence, with the other word. The occurrence of compounds is hypothesized in the training data by identifying maximal, connected sets of linked words in each pair of aligned sentences in the processed and scored training data. Whenever one of these maximal, connected sets contains more than one word in either or both of the languages, the subset of the words in that language is hypothesized as a compound.

Claims

1. A method of calculating translation relationships among words, comprising:

calculating word association scores for word pairs based on co-occurrences of words in each of a plurality of sets of aligned, bilingual units in a corpus;

identifying hypothesized compounds in the units based on the word association scores;

re-calculating the word association scores, given the hypothesized compounds;

ranking the word pairs based on the re-calculated word association scores;

generating transfer mappings that map from words and hypothesized compounds in one language to words and hypothesized compounds in another language, based on the ranking of the re-calculated word association scores; and

obtaining translation relationships based on the transfer mappings.

2. The method of claim 1 wherein the aligned, bilingual units comprise sentences.

3. The method of claim 1 wherein the aligned, bilingual units comprise logical forms.

4. The method of claim 1 wherein obtaining translation relationships comprises:

repeating the step of re-calculating word association scores considering co-occurrences of pairs, including pairs of words, pairs of compounds, and compound/word pairs, in a pair of aligned units only if the pairs are uniquely most strongly associated with one another among all words in the pair of aligned units, to obtain ultimate word association scores.

5. The method of claim 1 wherein recalculating the word association scores, given the hypothesized compounds, comprises:

replacing each hypothesized compound with a token to obtain a re-written corpus; and

re-calculating the word association scores on the aligned units in the re-written corpus.

6. The method of claim 1 wherein identifying hypothesized compounds comprises:

selecting an aligned unit pair having a first unit in a first language and a second unit in a second language; and

identifying hypothesized compounds based on the word association scores failing to show a one-to-one correspondence between words in the first unit and words in the second unit.

7. The method of claim 6 wherein identifying hypothesized compounds based on the word association scores failing to show a one-to-one correspondence comprises:

for each word in the first unit, identifying a most strongly associated word in the second unit; and

for each word in the second unit, identifying a most strongly associated word in the first unit.

8. The method of claim 7 wherein identifying hypothesized compounds based on the word association scores failing to show a one-to-one correspondence further comprises:

identifying maximal connected sets of words in the first and second units based on the identified most strongly associated words in the first and second units.

9. The method of claim 8 wherein identifying hypothesized compounds further comprises:

for each of the first and second units, identifying the words in each multi-word maximal connected set as a hypothesized compound.

10. The method of claim 1 and further comprising:

prior to calculating word association scores, accessing the corpus.

11. The method of claim 10 and further comprising:

parsing the corpus to obtain individual words.

12. The method of claim 1 and further comprising:

after calculating word association scores, pruning word pairs from being subject to further processing based on word association scores.

13. The method of claim 12 wherein pruning comprises:

removing word pairs from further processing if they have a word association score below a predetermined threshold score.

14. The method of claim 1 wherein the words in each of the aligned, bilingual units are converted to lemmas before the step of calculating word association scores.

15. A method of training a machine translation system, comprising:

obtaining a corpus of aligned, bilingual multi-word units;

calculating word association scores for word pairs in the corpus based on co-occurrence of words in the aligned units;

identifying hypothesized compounds based on an absence of one-to-one correspondence between words in the aligned units;

providing a rewritten corpus in which the hypothesized compounds have been replaced by fused tokens;

re-calculating the word association scores using the rewritten corpus;

generating transfer mappings that map from words and fused tokens in one language to words and fused tokens in another language, based on the selected translation relationships;

filtering the transfer mappings based on at least one of;

frequency of appearance;

completeness of parses of the multi-word units;

or completeness of alignment of multi-word units; and

training the machine translation system based on the filtered transfer mappings.

16. The method of claim 15 wherein identifying hypothesized compounds comprises:

selecting an aligned unit pair having a first unit in a first language and a second unit in a second language; and

identifying hypothesized compounds based on the word association scores failing to show a one-to-one correspondence between words in the first unit and words in the second unit.

17. The method of claim 16 wherein identifying hypothesized compounds based on the word association scores failing to show a one-to-one correspondence comprises:

for each word in the first unit, identifying a most strongly associated word in the second unit; and

for each word in the second unit, identifying a most strongly associated word in the first unit.

18. The method of claim 17 wherein identifying hypothesized compounds based on the word association scores failing to show a one-to-one correspondence further comprises:

identifying the maximal connected sets of words in the first and second units based on the identified most strongly associated words in the first and second units.

19. A computer-readable medium comprising computer-executable instructions which, when executed by a computer, configure the computer to:

calculate word association scores for word pairs based on co-occurrences of words in each of a plurality of sets of aligned, bilingual units in a corpus;

identify hypothesized compounds in the units based on word association scores that indicate a lack of one-to-one correspondence between words in the first unit and words in the second unit;

re-calculate the word association scores based on co-occurrences of words and hypothesized compounds;

rank the word pairs based on the re-calculated word association scores; and

generate transfer mappings that map from words and hypothesized compounds in one language to words and hypothesized compounds in another language, based on the ranking of the re-calculated word association scores.