Cluster storage apparatus for post processing error correction of a character recognition machine

• US 3,969,698 A
• Filed: 10/08/1974
• Issued: 07/13/1976
• Est. Priority Date: 10/08/1974
• Status: Expired due to Term

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Rosenbaum, Walter Steven, Hilliard, John Joseph, Chaires, Anne Marie, Bollinger, Ellen Willis, Ciconte, Jean Marie, Ett, Allen Harold, Kocher, Donald Francis
• Primary Examiner(s): Boudreau, Leo H.
• Application Number:
  US05/513,202
• Time in Patent Office: 644 Days
• Field of Search: 340/146.3 WD, 340/172.5, 340/146.3 ED, 179/1 SA, 179/1 SB, 179/1 SM, 179/1 SD
• US Class Current: 382/225
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/262 using context analysis, e.g. lexical, syntactic or semantic context

Abstract

A cluster storage apparatus is disclosed for outputting groups of valid alpha words as potential candidates for the correct form of an alpha word misrecognized by a character recognition machine. Groups of alpha words are arranged in the cluster storage apparatus such that adjacent locations contain alpha words having similar character recognition misread propensities. Alpha words which have been determined to be misrecognized, are input to the cluster storage apparatus. Numerical values assigned to the characters of which the input word is composed, are used to calculate the address of that group of valid alpha words having similar character recognition misread propensities. The cluster storage apparatus then outputs the accessed groups of alpha words for subsequent processing. The organization of the cluster storage apparatus minimizes the difference in address between alpha words with similar character recognition misread propensities by assigning high numeric values to highly reliable characters, as determined by measuring the character transfer function of the character recognition machine.

Claims

1. A cluster storage apparatus for outputting groups of valid alpha words as potential candidates for the correct form of an alpha word misrecognized by an OCR machine, comprising:

a two-dimenational array of alpha word read only storage locations, each location having a group of alpha words arranged such that adjacent locations contain alpha words having similar OCR misread propensities;

means for assigning numeric values to the characters of the input alpha word based upon the read reliability of the characters;

a first-dimensional accessing means for addressing said locations based upon the values assigned to the characters of which the input alpha word is composed;

a second-dimensional accessing means for accessing said locations based upon the number of characters in said input alpha word;

said first-dimensional accessing means calculating the first-dimensional address as a magnitude ##EQU7## where L_N is the numeric value assigned to each alpha character;

whereby an input alpha word which is potentially in error can be associated with that portion of the read only storage which contains potential candidates for the correct form of the input alpha word.

2. The cluster storage apparatus of claim 1, which further comprises:

said cluster storage being organized so as to minimize the difference in address between alpha words which have similar OCR misread propensities and so as to cluster words of a given character length, as well as words of other lengths that have a significant probability of being malsegmented into said given length;

said propensity being determined by empirical measurement of the OCR character transfer function;

said transfer function being expressed as a series of equations representing each character'"'"'s probability of being confused into a false output character;

said equations being solved for the optimum character value set which assigns higher numeric values to highly reliable characters and higher numeric values to characters which occur more frequently and lower numeric values to less reliable characters and lower numeric values to characters which occur less frequently;

said malsegmentation probability being determined by the OCR character transfer function;

said transfer function being expressed as a series of values representing the probability of a character being malsegmented;

said values being used to calculate the probability of each word being malsegmented;

said word malsegmentation probability being compared with a minimum threshold so that words whose malsegmentation propensity exceeds this threshold are stored with words of adjacent lengths;

whereby said cluster storage is organized in light of a global constraint so that no numeric assignment of two characters which can be misrecognized into one another will differ in location by more than a predetermined error interval.

4. The cluster storage apparatus of claim 1, which further comprises:

said cluster storage being organized so as to minimize the difference in address between phoneme words which have similar speech analyzer misread propensities and so as to cluster words of a given character length, as well as words of other lengths that have a significant probability of being malsegmented into said given length;

said propensity being determined by empirical measurement of the speech analyzer transfer function;

said transfer function being expressed as a series of equations representing each characters probability of being confused into a false output character;

said equations being solved for the optimum character value set which assigns higher numeric values to highly reliable characters and higher numeric values to characters which occur more frequently and lower numeric values to less reliable characters and lower numeric values to characters which occur less frequently;

said malsegmentation probability being determined by the speech analyzer character transfer function;

said transfer function being expressed as a series of values representing the probability of a character being malsegmented;

said values being used to calculate the probability of each phoneme word being malsegmented;

said word malsegmentation probability being compared with a minimum threshold so that words whose malsegmentation propensity exceeds this threshold are stored with words of adjacent lengths;

whereby said cluster storage is organized in light of a global constraint so that no numeric assignment of two characters which can be misrecognized into one another will differ in location by more than a predetermined error interval.

6. The cluster storage apparatus of claim 1, which further comprises:

said cluster storage being organized so as to minimize the difference in address between alpha words which have similar keyboard typographical error propensities and so as to cluster words of a given character probability of being malsegmented into said given length;

said propensity being determined by empirical measurement of the keyboard character transfer function;

said transfer function being expressed as a series of equations representing each character probability of being confused into a false output character;

said equations being solved for the optimum character value set which assigns higher numeric values to highly reliable characters and higher numeric values to characters which occur more frequently and lower numeric values to less reliable characters and lower numeric values to characters which occur less frequently;

said malsegmentation probability being determined by the keyboard character transfer function;

said transfer function being expressed as a series of values representing the probability of a character being malsegmented;

said values being used to calculate the probability of each word being malsegmented;

said word malsegmentation probability being compared with a minimum threshold so that words whose malsegmentation propensity exceeds this threshold are stored with words of adjacent lengths;

whereby said cluster storage is organized in light of a global constraint so that no numeric assignment of two characters which can be mistyped into one another will differ in location by more than a predetermined error interval.

3. A cluster storage apparatus for outputting groups of valid phoneme words as potential candidates for the correct form of a phoneme word misrecognized by speech analyzer machine, comprising:

a two-dimensional array of phoneme word read only storage locations, each location having a group of phoneme words arranged such that adjacent locations contain phoneme words having similar speech analyzer misread propensities;

means for assigning numeric values to the characters of the input phoneme word based upon the read reliability of the characters;

a first-dimensional accessing means for addressing said locations based upon the values assigned to the characters of which the input phoneme word is composed;

a second-dimensional accessing means for accessing said locations based upon the number of characters in said input phoneme word;

said first-dimensional accessing means calculating the first-dimensional address as a magnitude ##EQU8## where L_N is the numeric value assigned to each phoneme character;

whereby an input phoneme word which is potentially in error can be associated with that portion of the read only storage which contains potential candidates for the correct form of the input phoneme word.

5. A cluster storage apparatus for outputting groups of valid alpha words as potential candidates for the correct form of an alpha word mistyped on a keyboard machine, comprising:

a two-dimensional array of alpha word read only storage locations, each location having a group of alpha words arranged such that adjacent locations contain alpha words having similar keyboard typographical error propensities;

means for assigning numeric values to the characters of the input alpha word based upon the typographical error propensity of the characters;

a first-dimensional accessing means for addressing said locations based upon the values assigned to the characters of which the input alpha word is composed;

a second-dimensional accessing means for accessing said locations based upon the number of characters in said input alpha word;

said first-dimensional accessing means calculating the first-dimensional address as a magnitude ##EQU9## where L_N is the numeric value assigned to each alpha character;

whereby an input alpha word which is potentially in error can be associated with that portion of the read only storage which contains potential candidates for the correct form of the input alpha word.

7. The post processing error correction system comprising:

a word generating source having a character transfer function which represents the error propensity of multicharacter words output thereby;

a binary reference matrix having an input line connected to the input of said word generating source, to detect invalid alpha words;

said binary reference matrix having an output control line carrying a binary signal which indicates whether the input alpha word is valid;

a gate means connected to said output from said word generating source and having a control input from said control output of said binary reference matrix, for gating the input alpha word from said word generating source onto a first output line in response to a signal on said control line from said binary reference matrix indicating that said alpha word is valid, and gating said input alpha word onto a second output line in response to a signal from said binary reference matrix control line indicating said alpha word is invalid;

a cluster storage apparatus having an input connected to said second output line from said gating means, to access from an associative memory therein, a group of correct alpha words which have some probability of having been confused with said invalid alpha words input on said second output line from said gate;

the regional context error correction apparatus having an input connected to said output from said gating means and having a second input connected to the output from said cluster storage apparatus for accepting said group of correct alpha words;

said cluster storage apparatus executing a conditional probability analysis to determine which one of the group of correct alpha words most closely corresponds to the invalid alpha word output by said word generating source;

said regional context error correction apparatus outputting the word which most closely corresponds to the invalid alpha word output by said word generating source;

whereby the most probable correct version of a garbled word output from said word generating source, is determined.

8. The post processing error correction system of claim 7 wherein said word generating source is an optical character recognition machine.

9. The post processing error correction system of claim 7 wherein said word generating source is a speech analyzer and said output alpha words are composed of a sequence of phoneme characters.

10. The post processing error correction system of claim 7 wherein said word generating source is a keyboard having a character transfer function representing the propensity for the commission of typographical errors.