MINING BILINGUAL DICTIONARIES FROM MONOLINGUAL WEB PAGES

• US 20090070095A1
• Filed: 09/07/2007
• Published: 03/12/2009
• Est. Priority Date: 09/07/2007
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Gao, Jianfeng
• Granted Patent: US 7,983,903 B2
• Time in Patent Office: Days
• US Class Current: 704/2
• CPC Class Codes: G06F 40/45 Example-based machine translation; Alignment

Abstract

Systems and methods for identifying translation pairs from web pages are provided. One disclosed method includes receiving monolingual web page data of a source language, and processing the web page data by detecting the occurrence of a predefined pattern in the web page data, and extracting a plurality of translation pair candidates. Each of the translation pair candidates may include a source language string and target language string. The method may further include determining whether each translation pair candidate is a valid transliteration. The method may also include, for each translation pair that is determined not to be a valid transliteration, determining whether each translation pair candidate is a valid translation. The method may further include adding each translation pair that is determined to be a valid translation or transliteration to a dictionary.

Claims

1. A method for identifying translation pairs from web pages, the method comprising:

receiving monolingual web page data of a source language;

processing the web page data by;

detecting the occurrence of a predefined pattern in the web page data;

extracting a plurality of translation pair candidates, each of the translation pair candidates including a source language string and target language string;

determining whether each translation pair candidate is a valid transliteration;

for each translation pair that is determined not to be a valid transliteration, determining whether each translation pair candidate is a valid translation; and

adding each translation pair that is determined to be a valid translation or transliteration to a dictionary.

2. The method of claim 1, wherein the source language is Chinese.

3. The method of claim 1, wherein the target language is English.

4. The method of claim 1, wherein the predefined pattern includes a source language string and a target language string contained within parentheses.

5. The method of claim 1, wherein processing further includes:

prior to detecting the occurrence of a predetermined pattern, filtering the web page data to remove non-textual elements of the web page data.

6. The method of claim 5, wherein processing further includes:

segmenting the source language string of each translation pair candidate into one or more words, to thereby produce a segmented translation pair candidate.

7. The method of claim 6, wherein determining whether each translation pair candidate is a valid transliteration further includes:

separating the translation pair candidates into transliteration candidates and semantic translation candidates, based on one or more predefined rules;

converting the target language strings into one or more target language phonemes, each of the target language phonemes having a corresponding source language phoneme in the source language string;

applying an alignment model to determine a probability the source language string matches the target language string in the translation pair candidate, based at least in part on the determined probability each source language phoneme in the source language string is a transliteration of a corresponding target language phoneme in the target language string; and

classifying whether each translation pair candidate is valid transliteration pair, based on an output of the alignment model, and with reference to one or more classifying factors and a classifying threshold.

8. The method of claim 7, wherein the predefined classifying factors are selected from the group consisting of a logarithm of the transliteration probability of transliteration from the source language string to target language string normalized by the number of alignment units;

a normalized logarithm of the transliteration probability from the source language to the target language string;

a ratio of the number of source language alignments and target language alignments in the source language to target language translation;

a ratio of the number of source language alignments and target language alignments in the source language to target language transliteration; and

a perplexity of the source language string with respect to a source language transliteration model.

9. The method of claim 8, further comprising:

ranking each of the translation pair candidates based on a plurality of predefined ranking factors.

10. The method of claim 8, wherein the ranking is accomplished by a multiple nested ranker, and wherein the plurality of ranking factors are selected from the group consisting of a relative frequency of a source language alignment unit;

a length ratio of a target language alignment unit and a source language alignment unit;

a length of a source language alignment unit;

whether translation of a first target language word after reordering is in an alignment unit;

a number of unique terms before the candidate;

whether a word immediately before a suffix is an predetermined indicator word.

11. A system for identifying translation pairs from web pages, the system comprising a computer program configured to be executed on a computing device, the computer program including:

a preprocessing module configured to detect an occurrence of a predefined pattern in monolingual web page data of a source language, and to extract a plurality of translation pair candidates, each of the translation pair candidates including a source language string and target language string; and

a transliteration module configured to process the plurality of translation pair candidates to determine whether each translation pair candidate is a valid transliteration; and

a translation module to process each translation pair candidate that is determined not to be a valid transliteration, to determine whether each translation pair candidate is a valid translation;

wherein the computer program is configured to add each translation pair candidate that is determined to be a valid transliteration or a valid translation to a dictionary.

12. The system of claim 11, wherein the preprocessing module includes a filter configured to filter the web page data to remove non-textual elements of the web page data.

13. The system of claim 12, wherein the preprocessing module includes an extractor configured to extract translation pair candidates from the filtered web page data.

14. The system of claim 13, wherein the preprocessing module includes a segmentor configured to segment the source language string of each translation pair candidate into one or more source language words, to thereby produce a segmented translation pair candidate.

15. The system of claim 14, wherein the transliteration module includes a separator configured to separate the translation pair candidates into transliteration candidates and semantic translation candidates, based on one or more predefined rules.

16. The system of claim 15, wherein the transliteration module includes a phoneme converter configured to convert the target language strings into one or more target language phonemes, each of the target language phonemes having a corresponding source language phoneme in the source language string.

17. The system of claim 16, wherein the transliteration module includes an alignment model configured to determine a probability the source language string matches the target language string in the translation pair candidate, based at least in part on the determined probability each source language phoneme in the source language string is a transliteration of a corresponding target language phoneme in the target language string.

18. The system of claim 17, wherein the transliteration module includes a classifer configured to classify, via a binary classifier, whether each translation pair candidate is valid transliteration pair, based on an output of the alignment model, and with reference to one or more classifying factors and a classifying threshold.

19. The system of claim 18, wherein the translation module includes a ranker configured to rank each of the translation pair candidates based on a plurality of predefined ranking factors.

20. A method for identifying translation pairs from web pages, the method comprising:

receiving monolingual web page data of a source language;

processing the web page data to produce one or more translation pair candidates, each translation pair candidate including a source language string and a target language string;

determining whether one or more of the translation pair candidates is a valid transliteration, at least in part by applying an alignment model to determine a probability the source language string matches the target language string in the translation pair candidate, the alignment model being based on a determined probability that each of one or more source language phonemes in the source language string is a transliteration of a corresponding target language phoneme in the target language string; and

determining whether one or more of the translation pair candidate is a valid semantic translation at least in part by, ranking each of the translation pair candidates based on a plurality of predefined ranking factors, and determining whether a rank of each translation pair candidate is above a predetermined threshold.