Algorithm for context sensitive spelling correction

• US 5,907,839 A
• Filed: 07/03/1996
• Issued: 05/25/1999
• Est. Priority Date: 07/03/1996
• Status: Expired due to Term

• Current Assignee: Dan Roth, Hanger Solutions, LLC (IP Investments Group LLC)
• Original Assignee: Yeda Research and Development Co. Ltd. (Weizmann Institute of Science)
• Inventors: Roth, Don
• Primary Examiner(s): Amsbury, Wayne
• Assistant Examiner(s): Jung, David Yiuk
• Application Number:
  US08/675,618
• Time in Patent Office: 1,056 Days
• Field of Search: 707/1-206, 707/530-534, 704/1-10
• US Class Current: 1/1
• CPC Class Codes: G06F 40/232 Orthographic correction, e.g. spell checking or vowelisation
• CPC Class Codes: Y10S 707/99935 Query augmenting and refining, e.g. inexact access

Abstract

A method determining whether a target word used in a text is a correct word. The method being carried out in a data processor that is being configured according to a given operating system for reading data from, writing data to a storage medium and presenting data on a display screen. The method includes the steps of:

(a) identifying one or more features residing in the vicinity of the target word in a text. The features being associated with said target word in a database stored on said storage medium.

(b) using the features identified in step (a) to retrieve information from the database. The information being indicative as to the likelihood of said target word being in context with the features specified in step (a).

(c) using the information retrieved in step (b) as a criterion for indicating whether the target word is likely to be either the correct word or should it be replaced within said text.

Claims

1. A database which stores information for allowing a computer program to evaluate the usage of a target word within the context of surrounding words in a target text, the database comprising:

a computer-readable medium;

a database structure stored on the computer-readable medium, the database structure including;

a plurality of target words, each one of said target words including at least two members so as to constitute a target word cloud that represents said target word; and

a plurality of features and associated weight values stored with respect to each of said at least two members, the plurality of features reside in the vicinity of said one target word within the target text and are essentially common to all of said members;

wherein the plurality of weight values indicate a contextual relationship at least between the target word and the plurality of features.

2. The database according to claims 1, wherein said plurality of features include word features within ±

k words of the target word in said training text corpus.

3. The database according to claim 1, wherein said features include collocation features, each constituting a pattern of up to l items that reside the vicinity of said target word in said training text corpus.

4. The database according to claim 1, for use in a context related text checking application, wherein said plurality of target words include a plurality of confusion sets each of which including at least two words.

5. The database of claim 4, wherein each confusion set include at least two words having similar pronunciation but different spelling.

6. The database according to claim 1, wherein each one of said target words having either or both of promotion and demotion factor, for altering at least some of the weight values of active features, from among said plurality of features;

the active features being encountered in the vicinity of a candidate word within a text.

7. The database according to claim 6, wherein said candidate word and said target word belong to the same confusion set.

8. The database according to claim 6, wherein all of said target words having the same demotion factor.

9. The database according to claim 6, wherein all of said target words having the same promotion factor.

10. The database according to claim 1, wherein each member, from among said at least two members in a cloud, being associated with a unique promotion factor vis-a-vis the promotion factors of the remaining members from among said at least two instances.

11. The database according to claim 1, wherein each member, from among said at least two members in a cloud, being associated with a unique demotion factor vis-a-vis the demotion factors of the remaining members from among said at least two members.

12. The database according to claim 1, wherein each one of the target word members in a cloud is associated with a second weight value indicating the weight of said target word member in the overall representation of said target word.

17. The method according to claim 13, wherein said database is defined in claim 1.

23. The method according to claim 19, wherein said database is defined in claim 1.

24. A processor being configured according to a given operating system for reading data from, writing data to a storage medium and displaying data on a display screen;

the processor comprising means configured to access a database stored on said storage medium for performing context related text checking, the database being defined in claim 1.

13. A method for determining whether a target word used in a text is a correct word, the method being carried out in a data processor and utilizing a database structure, the data processor being configured according to a given operating system for reading data from, writing data to a storage medium and presenting data on a display screen, the data structure including:

(i) a plurality of target words;

(ii) each one of said target words has at least two members so as to constitute a target word cloud that represents said target word, a plurality of features and associated weight values stored with respect to each of said at least two members, the plurality of features reside in the vicinity of said one target word within the target text, and are essentially common to all of said members, wherein the plurality of weight values indicated a contextual relationship at least between the target word and the plurality of features;

the method comprising;

(a) identifying at least one feature residing in the vicinity of said target word in a text, said target word being in the database stored on said storage medium;

(b) using the at least one feature identified in step (a) to retrieve information from the database by application of the at least two members in the target word cloud, the information being indicative as to the likelihood of said target word being in context with the at least one feature; and

(c) evaluating the information retrieved in step (b) to determine whether the target word is likely to be either the correct word or if it should be replaced within said text.

14. The method according to claim 13, wherein the target word belongs to a confusion set that includes at least one additional candidate word;

and wherein the information stipulated in step (b) being indicative also as to the likelihood of at least one of said candidate words being in context with the at least one feature; and

wherein said step (c) is replaced by the following step (c'"'"')

(c'"'"') using the information retrieved in step (b) as a criterion for determining whether the target word should be either replaced within said text or replaced within said text by one of said candidate words.

15. The method according to claim 13, wherein the information stipulated in step (b) include at least one weight value associated with the at least one feature, respectively;

the plurality of weight values indicating a contextual relationship at least between the target word and the plurality of features.

16. The method according to claim 13, wherein the criterion for indicating whether the target word is the correct word as stipulated in said step (c), being:

##EQU8## w_i being said at least one weight value, x_i being said at least one feature, and θ

being a predetermined threshold.

18. The method of claim 14, wherein step (c'"'"') comprises presenting user with a list of candidate words on said display screen and allowing the user to at least select between the target and the candidate words.

19. A training method for determining whether a target word used in a training corpus text is a correct word, the method being carried out in a data processor and utilizing a database structure on a storage medium, the data processor being configured according to a given operating system for reading data from, writing data to the storage medium and presenting data on a display screen, the data structure including:

(i) a plurality of target words;

(ii) each one of said target words has at least two members so as to constitute a target word cloud that represents said target word, a plurality of features and associated weight values stored with respect to each of said at least two members, the plurality of features reside in the vicinity of said one target word within the target text, and are essentially common to all of said members, wherein the plurality of weight values indicated a contextual relationship at least between the target word and the plurality of features;

the method comprising;

(a) identifying at east one feature residing in the vicinity of said target word in a text, said target word being in the data structure stored on said storage medium;

(b) using the at least one feature identified in step (a) to acquire information from the database by application of the at least two members in the target word cloud, the information being indicative as to the likelihood of said target word being in context with the at least one feature; and

(c) using the information retrieved in step (b) as a criterion for predicting whether the target word is likely to be either the correct word or an incorrect word that should be replaced within said text and in the latter case, altering the information in said data structure.

20. The method according to claim 19, wherein altering said information is effected by utilizing Winnow based technique.

21. The method of claim 20, wherein the Winnow technique is modified to drop from the database representations of features whose weight value drops below a given threshold.

22. The method of claim 20, wherein features which are observed to appear, within various training corpus texts, always together with other features with respect to the same target word, are merged into a common feature having one weight value.