System for text correction adaptive to the text being corrected

• US 5,956,739 A
• Filed: 06/25/1996
• Issued: 09/21/1999
• Est. Priority Date: 06/25/1996
• Status: Expired due to Fees

• Current Assignee: Yeda Research and Development Co. Ltd. (Weizmann Institute of Science)
• Original Assignee: Mitsubishi Electric Information Technology Center America, Inc. (Mitsubishi Electric Corporation)
• Inventors: Roth, Dan, Golding, Andrew R.
• Primary Examiner(s): Zimmerman, Mark K.
• Assistant Examiner(s): GOOD JOHNSON, MOTILEWA
• Application Number:
  US08/669,928
• Time in Patent Office: 1,183 Days
• Field of Search: 707/533, 707/532, 707/534
• US Class Current: 715/209
• CPC Class Codes: G06F 40/232 Orthographic correction, e.g. spell checking or vowelisation

Abstract

A system is provided for correcting users'"'"' mistakes including context-sensitive spelling errors and the like in which an adaptive correction algorithm is utilized which is trained on not only a conventional training corpus, but also on the text which is being corrected, thus to permit the correction of words based on the particular usages of the words in the text being corrected, taking advantage of the fact that the text to be corrected is by and large already mostly correct.

Claims

1. A system for providing context sensitive word correction in which the context of a word in a sentence is utilized to determine which of several alternative or possible correctly-spelled words was intended, comprising:

a conventional training corpus;

a target text having errors in which the word usage is mostly correct, but sometimes incorrect;

means for combining said conventional training corpus and said target text into a combined corpus;

means coupled to said combined corpus for ascertaining the probability that a word in said target text is the correct word based on the occurrence of said word in said combined corpus, such that context sensitive word correction is trained not only on a conventional training corpus, but also on a target text which contains word usage errors whereby the correspondence of said training corpus to said target text need not be strong, and whereby there is no user supplied feedback to supervise the training process.

2. The system of claim 1, and further including means for extracting a feature from said combined corpus and wherein said means for ascertaining the probability that a word in said target text is the correct word includes means coupled to said feature for altering said probability based on the occurrence of said feature in said target text.

3. The system of claim 2, wherein said feature is a context word.

4. The system of claim 2, wherein said feature is a collocation.

5. The system of claim 2, wherein said means for ascertaining the probability that a word in said target text is the correct word includes means for updating the probability that a word could have been intended for the target word via the application of Bayes'"'"' rule.

6. The system of claim 2, wherein said means for ascertaining the probability that a word in said target text is the correct word includes means for assigning an initial probability to a word that could have been intended for the target word in accordance with the ratio of the probability that said word occurred in said combined corpus to the probability that any of the words that could have been intended for the target word occurred in said combined corpus.

7. A method of correcting target text, comprising the steps of:

analyzing uncorrected target text to create a training corpus; and

applying the training corpus to the uncorrected target text to identify a first word as an erroneous word to be replaced to correct the uncorrected target text.

8. The method according to claim 7, further comprising the step of:

replacing the identified first word with a second to correct the uncorrected target text.

9. The method according to claim 8, wherein the first word is a correctly spelled word and the second word is a correctly spelled word.

10. The method according to claim 7, wherein the first word is erroneous based upon a context-sensitive spelling error.

11. The method according to claim 7, wherein the analyzing of the uncorrected target text includes determining correct contextual use of the second word in the uncorrected target text.

12. The method according to claim 7, wherein:

the analyzing includes determining rules of use for the second word; and

the corpus corresponds to the determined rules of use.

13. The method according to claim 7, wherein the training corpus

is a first training corpus and further comprising the steps of:

replacing the identified first word with a second to correct the uncorrected target text;

applying a second training corpus, created by analyzing training text different than the uncorrected target text, to the uncorrected target text to identify a third word as an erroneous word; and

replacing the identified third word with a fourth word to further correct the uncorrected target text.

14. The method according to claim 13, wherein the third word is a misspelled word and the fourth word is a correctly spelled word.

15. The method according to claim 7, wherein:

the analyzing includes identifying inconsistencies in the uncorrected target text; and

the training corpus corresponds to the identified inconsistencies.

16. The method according to claim 7, wherein the analyzing includes determining features that characterize a context in which the first word appears in the uncorrected target text and a context in which the second word appears in the uncorrected target text.

17. The method according to claim 16, wherein the determined features include at least one of:

a particular key word occurring within a distance of the first word and within the same distance of the second word in the uncorrected target text, and

a pattern of part-of-speech and specific words within a context of the first word and within a context of the second word.

18. The method according to claim 16, wherein the analyzing further includes eliminating those of the features which are determined on the basis of less than a threshold amount of data.

19. A system for replacing a word within uncorrected target text with another word to correct the uncorrected target text, comprising:

a training corpus created from uncorrected target text; and

a buffer configured to store the uncorrected target text including a first word which is an erroneous word, and to replace, the first word with a second word, determined on the basis of the training corpus, to correct the uncorrected target text.

20. The system according to claim 19, wherein the first word is a correctly spelled word and the second word is a correctly spelled word.

21. The system according to claim 19, wherein the first word is erroneous based upon a context-sensitive spelling error.

22. The system according to claim 19, wherein the training corpus is a first training corpus and further comprising:

a second training corpus created from other than the uncorrected target text;

the uncorrected target text includes a third word which is an erroneous word; and

the buffer is further configured to replace the third word with a fourth word, determined on the basis of the second training corpus, to correct the uncorrected target text.