General purpose correction of grammatical and word usage errors

• US 9,262,397 B2
• Filed: 12/07/2010
• Issued: 02/16/2016
• Est. Priority Date: 10/08/2010
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Inventors: Gamon, Michael, Knig, Christian
• Primary Examiner(s): Desir, Pierre-Louis
• Assistant Examiner(s): KOVACEK, DAVID M
• Application Number:
  US12/961,516
• Publication Number:
  US 20120089387A1
• Time in Patent Office: 1,897 Days
• Field of Search: 704/1, 704 4- 8, 704/9, 704/10, 704251-257, 704/270, 704/276, 704E15019-E15024, 704E11001-E11007, 704E13001-E13014, 707758-759, 707/765
• US Class Current: 1/1
• CPC Class Codes: G06F 40/00 Handling natural language data speech analysis or synthesis, speech recognition G10L
• CPC Class Codes: G06F 40/20 Natural language analysis semantic analysis of natural language G06F40/30
• CPC Class Codes: G06F 40/253 Grammatical analysis; Style critique
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/19 Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules

Abstract

Architecture that detects and corrects writing errors in a human language based on the utilization of three different stages: error detection, correction candidate generation, and correction candidate ranking. The architecture is a generic framework for generating fluent alternatives to non-grammatical word sequences in a written sample. Error detection is addressed by a suite of language model related scores and other scores such as parse scores that can identify a particularly unlikely sequence of words. Correction candidate generation is addressed by a lookup in a very large corpus of “correct” English that looks for alternative arrangements of the same or similar words or subsequences of these words in the same context. Correction candidate ranking is addressed by a language model ranker.

Claims

1. A computer-implemented error correction system, comprising:

an error detection component that collects scores based on word usage in a sequence of multiple words intended to form a fluent language expression, generates a signal for presence of an error in a subsequence of two or more words of the sequence based on the scores, and detects an erroneous subsequence of two or more words in the sequence of multiple words;

wherein the error detection component operates as a sliding window covering subsequences of two or more words of the sequence of multiple words;

a potential candidate generation component that generates potential candidate strings of words related to the erroneous subsequence of words from a corpus, each of the candidate strings having a variable number of words compared a number of words of the erroneous subsequence;

a candidate selection component that selects and presents one or more of the potential candidate strings as candidate strings for correction of language fluency introduced by the erroneous subsequence of words; and

a processor that executes computer-executable instructions associated with at least one of the error detection component, potential candidate generation component, or candidate selection component.

2. The system of claim 1, wherein the error detection component detects the erroneous subsequence of words based on scores from at least one of multiple language models or phrase count tables that detect an unlikely sequence of words.

3. The system of claim 1, wherein the candidate selection component selects a candidate string based on a score of the potential candidate strings from an associated language model.

4. The system of claim 1, wherein the scores for error detection are used as features in a classification or sequence model that learns how scores of the language models indicate presence of the erroneous subsequence of words.

5. The system of claim 1, wherein the corpus used by the potential candidate generation component is represented as a suffix array.

6. The system of claim 1, wherein the potential candidate generation component generates the potential candidate strings from a corpus based on context.

7. The system of claim 1, wherein the potential candidate generation component generates the potential candidate strings based on comparison of number of words in the potential candidate strings to number of words in the erroneous subsequence.

8. The system of claim 1, wherein the potential candidate generation component generates the potential candidate strings based on number of words in the potential candidate strings that match words in the erroneous subsequence.

9. A computer-implemented error correction method, comprising acts of:

receiving a sequence of multiple words intended to form a fluent language expression;

detecting an erroneous subsequence of two or more words in the sequence of multiple words defined by a sliding window covering subsequences of two or more words of the sequence of multiple words;

generating potential candidate strings of words related to the erroneous subsequence of words from a corpus of alternative arrangements of same or similar words, or subsequences of words in a same context, each of the candidate strings having a variable number of words compared a number of words of the erroneous subsequence;

selecting candidate strings from one or more of the potential candidate strings for correction of dysfluency of the language expression by the erroneous subsequence of words and presenting potential corrections to a user; and

utilizing a processor that executes instructions stored in memory to perform at least one of the acts of detecting, generating, or selecting.

10. The method of claim 9, further comprising collecting scores from one or more language models and processing the scores to detect the erroneous subsequence of words.

11. The method of claim 9, further comprising utilizing language model scores in a sequence model to detect the erroneous subsequence of words in the intended fluent language expression and location of the subsequence in the expression.

12. The method of claim 9, further comprising finding a most likely sequence of error and non-error states based on language model scores.

13. The method of claim 9, further comprising presenting only candidate strings that rank higher than the sequence of words under detection.

14. The method of claim 9, further comprising selecting and presenting only a top number of candidate strings based on a ranking component that ranks the candidate strings.

15. A computer-implemented error correction method, comprising acts of:

receiving a sequence of multiple words intended to form a fluent language expression;

collecting scores from one or more language models based on subsequences of two or more words in the sequence of multiple words defined by a sliding window covering two or more words of the sequence of multiple words;

detecting an erroneous subsequence of two or more words based on the scores collected for each subsequence of words in the sequence;

generating potential candidate strings of words related to the erroneous subsequence of words from a corpus of alternative arrangements of same or similar words or subsequences of words, based on context-related criteria, each of the candidate strings having a variable number of words compared a number of words of the erroneous subsequence;

selecting candidate strings from the one or more of the potential candidate strings for correction of dysfluency of the language expression by the erroneous subsequence of words and presenting the candidate strings to a user as corrections to the dysfluency; and

utilizing a processor that executes instructions stored in memory to perform at least one of the acts of collecting, detecting, generating, or selecting.

16. The method of claim 15, further comprising detecting the erroneous sequence of words based on unlikelihood of an entirety of the sequence of words in the sliding window being correct.

17. The method of claim 15, further comprising selecting the candidate strings based on context-related criteria that is at least one of n-word distance from an original corresponding input sequence, semantically similar words to the original intended expression, or morphologically similar words in the original intended expression.

18. The method of claim 15, further comprising selecting the candidate strings based on context-related criteria that includes comparison of a number of words in the candidates strings to number of words in the erroneous subsequence or a number of words in the candidate strings that match words in the erroneous subsequence.

19. The method of claim 15, further comprising presenting a subset of the candidate strings that rank higher than an original corresponding input sequence of words defined by the sliding window.