Method and apparatus for translation based on a repository of existing translations

• US 7,707,025 B2
• Filed: 06/23/2005
• Issued: 04/27/2010
• Est. Priority Date: 06/24/2004
• Status: Expired due to Fees

• Current Assignee: Sharp Kabushiki Kaisha (Hon Hai Precision Industry Co., Ltd.)
• Original Assignee: Sharp Kabushiki Kaisha (Hon Hai Precision Industry Co., Ltd.)
• Inventors: Whitelock, Peter J.
• Primary Examiner(s): Dorvil; Richemond
• Assistant Examiner(s): He; Jialong
• Application Number:
  US11/159,322
• Publication Number:
  US 20060004560A1
• Time in Patent Office: 1,769 Days
• Field of Search: 704/2, 704/4, 704/5, 704/7
• US Class Current: 704/5
• CPC Class Codes: G06F 40/45 Example-based machine translation; Alignment

Abstract

A method of translating an input sentence in a source language to an output sentence in a target language uses a store comprising a plurality of example sentences. A base example sentence is chosen from the store based on a comparison of the input sentence with a plurality of example sentences, and its paired translation is used as a translation basis. A portion of the input sentence differing from a corresponding portion of the base example sentence is identified. A portion of the translation basis aligned with the base example unmatched is located. The input unmatched portion is used to select a set of subsidiary example sentences from the store. A choice of possible translations corresponding to the input unmatched portion is determined from the set of subsidiary example sentences. A translation is selected from the choice based on a predetermined selection algorithm.

Claims

1. A method performed in a Translation Memory Apparatus for translating an input sequence of data items in a first format to an output sequence of data items in a second format using a store comprising a plurality of example sequences in the first format each paired with its translation in the second format, comprising:

(a) a processor of the apparatus choosing a base example sequence from the store based on a comparison of the input sequence with each of a plurality of example sequences from the store, and using its paired translation as a translation basis;

(b) the processor identifying a portion of the input sequence differing from a corresponding portion of the base example sequence, these portions being designated input and base example unmatched portions respectively and other portions that are not the unmatched portions being designated input and base example matched portions respectively;

(c) the processor locating a portion of the translation basis corresponding to the base example unmatched portion, wherein, when the portion of the translation basis corresponds to the base example unmatched portion and an adjacent base example matched portion, extending the base example unmatched portion to include the adjacent base example matched portion, and extending the corresponding input unmatched portion to include an adjacent input matched portion corresponding to the adjacent base example matched portion;

(d) the processor using the input unmatched portion to select a set of subsidiary example sequences from the store;

(e) the processor determining from the set of subsidiary example sequences a choice of possible translations corresponding to the input unmatched portion;

(f) the processor selecting a translation from the choice based on a predetermined selection algorithm and using the selected translation to replace the portion located in step (c); and

(g) the processor using the result of step (f) as a basis for the output sequence of data items.

2. The method as claimed in claim 1, wherein steps (b) to (f) are repeated for a plurality of input unmatched portions identified in the input sequence.

3. The method as claimed in claim 2, wherein steps (b) to (t) are repeated for every unmatched portion identified in the input sequence.

4. The method as claimed in claim 1, wherein the base example sequence chosen in step (a) is the example sequence from the store which most closely matches the input sequence according to a predetermined measure of closeness.

5. The method as claimed in claim 1, wherein a base example unmatched portion identified in step (b), and its corresponding translation basis portion located in step (c), may be an empty sequence positioned between two adjacent data items, such that the replacement in step (f) is effectively an insertion at a position.

6. The method as claimed in claim 1, wherein in step (f) a morphological variant of the selected translation is used to replace the portion located in step (c) depending on context.

7. The method as claimed in claim 1, wherein further processing is performed on the sequence resulting from step (f) before use as the output sequence of data items.

8. The method as claimed in claim 1, wherein the data items are words.

9. The method as claimed in claim 8, wherein the first and second formats are first and second languages respectively.

10. The method as claimed in claim 9, wherein the input sequence of data items forms a grammatically complete phrase.

11. The method as claimed in claim 10, wherein the input sequence of data items forms a sentence.

12. The method as claimed in claim 1, wherein in step (c) the portion is located using an alignment between parts of the base example sequence and respective parts of its paired translation.

13. The method as claimed in claim 12, wherein the alignment is predetermined.

14. The method as claimed in claim 1, wherein step (d) comprises the processor selecting an example sequence for inclusion in the set if at least one data item in the example sequence matches or corresponds to at least one data item in the input unmatched portion.

16. The method as claimed in claim 1, wherein the predetermined selection algorithm comprises the processor, for each of at least one translation in the choice, allocating to the translation a preference value determined in dependence upon at least one subsidiary example from which the translation is derived.

17. The method as claimed in claim 16, wherein a preference value is determined in dependence upon a comparison between the input sequence and each of the at least one subsidiary example.

18. The method as claimed in claim 17, wherein a preference value is determined by allocating a weighting to the or each such comparison, and combining the allocated weighting(s) in a predetermined manner.

19. The method as claimed in claim 18, wherein a weighting is allocated in dependence upon the number of data items or variants thereof common to the input sequence and the subsidiary example in the comparison.

20. The method as claimed in claim 18, wherein a weighting is allocated in dependence upon one or more of the following properties of a data item common to the input sequence and the subsidiary example:

the position in sequence of the common data item;

the frequency of the common data item in a predetermined corpus;

the concurrence frequency of the common data item with the input unmatched portion; and

the deviation of the concurrence frequency from that predicted by the frequency of the common data item and the input unmatched portion.

21. The method as claimed in claim 20, wherein a weighting is allocated in dependence upon the position in sequence of the common data item relative to the unmatched portion.

22. The method as claimed in claim 18, wherein the allocated weightings are combined by summation.

23. The method as claimed in claim 17, wherein the data item(s) in the input unmatched portion is not included in the comparison.

24. The method as claimed in claim 16, wherein a preference value is determined in dependence upon a comparison between the translation basis and the translation paired to each of the at least one subsidiary example.

25. The method as claimed in claim 16, wherein the predetermined selection algorithm further comprises the processor selecting a translation from the choice in dependence upon allocated preference value(s).

26. The method as claimed in claim 16, wherein a preference value is allocated to the or each translation in the choice.

27. The method as claimed in claim 16, wherein the preference value is determined in dependence upon the or each subsidiary example from which the translation is derived.

28. The method as claimed in claim 1, wherein the predetermined selection algorithm allows manual intervention to influence the translation selected.

29. The method as claimed in claim 1, comprising, where a suitable translation cannot be selected in step (f), the processor subdividing the input unmatched portion into a plurality of input unmatched portions and identifying corresponding respective base example unmatched portions, and performing steps (c) to (f) for each unmatched portion.

30. The method as claimed in claim 1, comprising, where a suitable base example cannot be found in step (a), the processor subdividing the input sequence into a plurality of input sub-sequences, performing steps (a) to (f) for each such sub-sequence, and combining the results from the steps (f) to produce the output sequence in step (g).

31. The method as claimed in claim 1, wherein the store comprises separate stores for use respectively in steps (a) and (d).

15. The method as claimed in claim I, wherein step (e) comprises the processor identifying, for a subsidiary example in the set, a portion of the subsidiary example corresponding to the input unmatched portion, and using a corresponding portion of the translation paired to the subsidiary example to form one of the possible translations in the choice.

32. A Translation Memory system comprising apparatus for translating an input sequence of data items in a first format to an output sequence of data items in a second format using a store comprising a plurality of example sequences in the first format each paired with its translation in the second format, the apparatus comprising:

a unit which chooses a base example sequence from the store based on a comparison of the input sequence with each of a plurality of example sequences from the store, and uses its paired translation as a translation basis;

a unit which identifies a portion of the input sequence differing from a corresponding portion of the base example sequence, these portions being designated input and base example unmatched portions respectively and other portions that are not the unmatched portions being designated input and base example matched portions respectively;

a unit which locates a portion of the translation basis corresponding to the base example unmatched portion, wherein, when the portion of the translation basis corresponds to the base example unmatched portion and an adjacent base example matched portion, extending the one of the base example unmatched portion to include the adjacent base example matched portion, and extending the corresponding input unmatched portion to include an adjacent input matched portion corresponding to the adjacent base example matched portion;

a unit which uses the input unmatched portion to select a set of subsidiary example sequences from the store;

a unit which determines from the set of subsidiary example sequences a choice of possible translations corresponding to the input unmatched portion;

a unit which selects a translation from the choice based on a predetermined selection algorithm and uses the selected translation to replace the portion located by the locating unit; and

a unit which uses the result of the selecting unit as a basis for the output sequence of data items.

33. A computer readable recording medium having stored thereon a computer executable program for translating an input sequence of data items in a first format to an output sequence of data items in a second format using a store comprising a plurality of example sequences in the first format each paired with its translation in the second format, comprising:

(a) choosing a base example sequence from the store based on a comparison of the input sequence with each of a plurality of example sequences from the store, and using its paired translation as a translation basis;

(b) identifying a portion of the input sequence differing from a corresponding portion of the base example sequence, these portions being designated input and base example unmatched portions respectively and other portions that are not the unmatched portions being designated input and base example matched portions respectively;

(c) locating a portion of the translation basis corresponding to the base example unmatched portion, wherein, when the portion of the translation basis corresponds to the base example unmatched portion and an adjacent base example matched portion, extending the base example unmatched portion to include the adjacent base example matched portion, and extending the corresponding input unmatched portion to include an adjacent input matched portion corresponding to the adjacent base example matched portion;

(d) using the input unmatched portion to select a set of subsidiary example sequences from the store;

(e) determining from the set of subsidiary example sequences a choice of possible translations corresponding to the input unmatched portion;

(f) selecting a translation from the choice based on a predetermined selection algorithm and using the selected translation to replace the portion located in step (c); and

(g) using the result of step (f) as a basis for the output sequence of data items.