Balancing out-of-dictionary and in-dictionary recognition scores

• US 7,899,251 B2
• Filed: 06/05/2006
• Issued: 03/01/2011
• Est. Priority Date: 06/05/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Slavik, Petr
• Primary Examiner(s): Hung; Yubin
• Assistant Examiner(s): Yentrapati; Avinash
• Application Number:
  US11/446,872
• Publication Number:
  US 20070280537A1
• Time in Patent Office: 1,730 Days
• Field of Search: None
• US Class Current: 382/181
• CPC Class Codes: G06V 30/2272 with lexical matching
• CPC Class Codes: G06V 30/36 Matching; Classification

Abstract

Various technologies and techniques are disclosed that improve handwriting recognition operations. A balancing factor is calculated that can be used in recognition mode to compare out-of-dictionary recognition scores with in-dictionary recognition scores. Correct ink samples of words are provided, with some in-dictionary and some out-of-dictionary words. One or more tuning sets are generated from the samples, the tuning set(s) having a percentage of out-of-dictionary words based upon what is expected for a typical user. A handwriting recognizer is run against the tuning set(s) to determine statistics. The statistics are based upon the recognized answer for each word compared to the correct answer. The statistics are used to calculate the balancing factor. During recognition mode, a user inputs handwriting to be recognized. The balancing factor is used to compare and combine an in-dictionary alternate list with an out-of-dictionary alternate list to reach a recognition decision.

Claims

1. A computer implemented method for generating a second out-of-dictionary recognition score for an out-of-dictionary recognized word for handwriting recognition, comprising:

analyzing, with a microprocessor, a tuning word to find an in-dictionary tuning word and an out-of-dictionary tuning word that substantially match the tuning word;

selecting a balancing factor, based on a probability the in-dictionary tuning word and the out-of-dictionary tuning word correctly match the tuning word, comprising one of;

selecting a value for the balancing factor that is less than a resulting quotient of the probability of the in-dictionary tuning word being correct divided by the probability of the out-of-dictionary tuning word being correct, if the in-dictionary tuning word is correct and the out-of-dictionary tuning word is not correct; and

selecting a value for the balancing factor that is greater than a resulting quotient of the probability of the in-dictionary tuning word being correct divided by the probability of the out-of-dictionary tuning word being correct, if the in-dictionary tuning word is not correct and the out-of-dictionary tuning word is correct; and

combining the balancing factor with a first out-of-dictionary recognition score for a word determined, during a recognition process, to be an out-of-dictionary word to generate a second out-of-dictionary recognition score for the out-of-dictionary recognized word.

2. The method of claim 1, comprising comparing the second out-of-dictionary recognition score for the out-of-dictionary recognized word to an in-dictionary recognition score for an in-dictionary recognized word.

3. The method of claim 2, comprising combining a list of recognized in-dictionary words with a list of recognized out-of dictionary words using the second out-of-dictionary recognition score and the in-dictionary recognition score.

4. The method of claim 3, comprising using the combined list to make a handwriting recognition decision.

5. The method of claim 1, the second out-of-dictionary recognition score comprising a product of the balancing factor and the first out-of-dictionary recognition score.

6. The method of claim 1, the first out-of-dictionary recognition score comprising a combination of a matching probability of the out-of dictionary recognized word and a weak language model probability for the out-of dictionary recognized word.

7. The method of claim 1, comprising providing the tuning word from a set of tuning words.

8. The method of claim 1, comprising receiving user input to adjust the balancing factor.

9. The method of claim 7, comprising randomly generating the set of tuning words.

10. A computer-readable storage media having computer-executable instructions that when executed on a computer via a microprocessor perform a method comprising:

generating a tuning set from a set of provided correct ink samples of words, the tuning set comprising at least some in-dictionary words and a number of out-of-dictionary words, the number of out-of-dictionary words determined by a percentage of out-of-dictionary words expected for a typical user;

running a handwriting recognizer against the tuning set to determine a set of statistics for a plurality of words in the tuning set, the set of statistics based at least in part upon a probability that a word from the tuning set matches a correct answer; and

calculating a balancing factor based on the set of statistics for use in comparing an in-dictionary recognition score with an out-of-dictionary recognition score in a recognition operation, comprising one of;

selecting a value for the balancing factor that is less than a resulting quotient of a probability of an in-dictionary tuning word being correct divided by a probability of an out-of-dictionary turning word being correct, if the in-dictionary tuning word is correct and the out-of-dictionary tuning word is not correct; and

selecting a value for the balancing factor that is greater than a resulting quotient of the probability of the in-dictionary tuning word being correct divided by the probability of the out-of-dictionary tuning word being correct, if the in-dictionary tuning word is not correct and the out-of-dictionary tuning word is correct.

11. The computer-readable media of claim 10, the method comprising generating the tuning set randomly.

12. The computer-readable media of claim 10, the method comprising determining the percentage of out-of-dictionary words expected for the typical user.

13. A computing device for handwriting recognition, comprising:

one or more processors; and

a computer readable storage medium comprising computer executable instructions that when executed by the one or more processors perform operations comprising;

segmenting handwritten input from a user into a plurality of segments;

determining a set of matching probabilities for at least a portion of the segments;

generating an in-dictionary alternate list comprising a plurality of in-dictionary recognition scores;

generating an out-of-dictionary alternate list comprising a plurality of out-of-dictionary recognition scores;

calculating a balancing factor using a tuning set comprising in-dictionary tuning words and out-of-dictionary tuning words, the calculating comprising one of;

selecting a first value for the balancing factor that is less than a resulting quotient of a first probability of an in-dictionary tuning word being correct divided by a second probability of an out-of-dictionary tuning word being correct, if the in-dictionary tuning word is correct and the out-of-dictionary tuning word is not correct; and

selecting a second value for the balancing factor that is greater than a resulting quotient of a third probability of the in-dictionary tuning word being correct divided by a fourth probability of the out-of-dictionary tuning word being correct, if the in-dictionary tuning word is not correct and the out-of-dictionary tuning word is correct;

comparing the in-dictionary alternate list with the out-of-dictionary alternate list using the balancing factor;

combining the in-dictionary alternate list and the out-of-dictionary alternate list, based on the comparison of the in-dictionary alternate list with the out-of-dictionary alternate list, to create a combined alternate list; and

making a final recognition decision regarding the handwritten input using the combined alternate list.

14. The method of claim 13, comprising displaying one or more characters representing the final recognition decision to the user.

15. The method of claim 13, comprising multiplying the balancing factor by the respective out-of-dictionary recognition scores to generate corresponding revised out-of-dictionary recognition scores.