JPH0218682A - Pattern recognition post-processing system - Google Patents

Abstract

PURPOSE:To execute pattern recognition post-processing at a high speed by automatically registering the information of operation, in which an operator corrects a pattern to be erroneously recognized, as a correcting rule when the pattern recognition post-processing is executed, applying the rule for a following recognized result and automatically correcting an recognition error. CONSTITUTION:The correction rule to be composed of the relation of right solution category to a candidate category set to be automatically extracted from the information, with which the operator corrects the category of the pattern to be erroneously recognized, is applied for the following pattern recognized result. When the candidate category set in the correction rule and the candidate category set to be included in the pattern recognized result are approximated, the right solution category of the correction rule is replaced with the first candidate category of the pattern recognized result. Thus, correction can be easily executed in comparison with correction by the re-registration of pattern feature quantity or grammar processing and the processing can be executed at the high speed.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to pattern recognition of speech, characters, etc. Regarding pattern recognition post-processing method,
.

[Conventional technology]

In recent years, with the development of office automation, character and voice recognition has been widely researched, and some have been put into practical use1.
To recognize a single character or voice, candidate characters and voices (hereinafter referred to as candidate categories) are generated by converting these into patterns, extracting pattern features, and comparing them with standard patterns registered in a recognition dictionary. .

Normally, such pattern recognition improves the recognition rate, so for patterns that are particularly likely to be misrecognized, recognition pattern special! Processing to additionally register three quantities is being carried out. It has also been proposed to improve recognition accuracy by performing grammatical processing such as word matching and morphological analysis on candidate categories.

[Problem to be solved by the invention]

However, in the process of additionally registering the recognized pattern feature 1, since the amount of information of the pattern feature amount is large, there is a problem that a large capacity storage area is required to register a new pattern feature amount. Also, for the above reasons,
There is a problem in that the time required for recognition processing increases, and furthermore, since patterns of new categories (characters and sounds) are registered, recognition accuracy actually decreases. In order to avoid this decrease in recognized viscosity, it is necessary to check whether or not the pattern should be zero-recorded, which poses a problem in that it takes time.

Furthermore, conventional techniques that use grammar processing to improve recognition accuracy have problems such as 1. The time required for recognition processing increases rapidly as the number of candidate categories increases. If there is no correct answer in the candidate category, or if the corresponding word does not exist in the recognition dictionary, the result obtained by grammar processing is correct, but there is a strong tendency to generate incorrect sentences, so the operator must correct f. The problem was that it made the work difficult.

Furthermore, in any of the above conventional pattern recognitions, information about the operation after which the operator corrected the error in the recognition result by W1 is not used.

Therefore, the present invention solves the above-mentioned problems of the prior art,
The purpose of the present invention is to obtain high recognition accuracy with fast and easy operations without increasing the capacity of a recognition dictionary, using information on operations when an operator corrects errors in recognition results.

[Means to solve the problem]

In a pattern recognition method that recognizes an input pattern and generates one or more candidate categories, the present invention enables an operator to correct the category of an erroneously recognized pattern. A correction rule based on the relationship is applied to subsequent pattern recognition results.

If the candidate category set in the correction rule and the candidate category set included in the pattern recognition result are similar to each other, the correct category in the correction rule is replaced with the first candidate category in the pattern recognition result.

Preferably, the automatically extracted correction rule is applied to a set of candidate categories included in the pattern gJ result for which the correct answer has already been corrected, and the misrecognized categories for the total categories corrected by the replacement are correctly corrected. Only those correction rules for which the ratio of the total number of categories exceeds a certain value are applied to subsequent corrections of pattern recognition results.

[Effect] For example, for the correct pattern A, the candidate category set is B
, C. Since A is not included in B or C in this candidate category set, the operator determines that A is genuine. As a result, a correction rule consisting of A, B, and C is obtained.

Through subsequent pattern recognition, a candidate category set that approximates the candidate category set in the correction rule, for example, the same B, C
Suppose that we obtain In this case, the correction rules stipulate that B and C should correct A, so the first candidate category (for example, B) in the candidate category sets B and C is corrected to A. In this way, the correction can be performed more easily than the conventional correction by re-registering the pattern feature amount or by grammar processing, and the processing speed can be increased.

In the above operation, the correction rule for correcting B and C to A is not always correct in subsequent pattern recognition processing. Therefore, the reliability of automatic correction can be improved by applying only the correction rules for which the ratio is equal to or greater than a certain value to the subsequent correction of pattern recognition results.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of the pattern recognition post-processing method of the present invention. In the figure, an automatic correction unit 1 inputs a pattern from a pattern recognition device (not shown),
The input pattern recognition result is automatically corrected according to the correction rules in the correction rule table 2. The automatically corrected result is output to the manual correction control section 3. Note that the correction rules included in the correction rule table 2, which will be described later, are automatically generated and updated sequentially as the operator manually corrects the pattern recognition results, so they are not applied to the initial (first) pattern recognition results. . That is, the first pattern recognition result is directly output to the manual correction control section 3 via the automatic correction section 1. .

The correction rule table 2 defines correction rules when the automatic correction unit 1 automatically corrects the pattern recognition result.

As mentioned above, these correction rules are sequentially generated by the operator manually blocking the pattern recognition results.

The updated 1° manual correction control unit 3 transmits the automatic correction result (or first pattern recognition result) output from the automatic correction unit 1 to the operation unit 6.
Corrections are made in accordance with correction operation instructions given by the operator via the . [Manual l] The primary and secondary control unit creates a correction rule, which will be described later, according to the manual correction operation information for the initial pattern recognition result, and registers the created correction rule in the correction rule table 2 via the correction rule determination unit 5. . After manual correction is completed for pattern recognition results subsequent to the initial recognition results, a new correction rule candidate list and a correction rule update list, which will be described later, are created and sent to the correction rule determining section 5, respectively. In addition, the manual correction control unit 3 creates history information, which will be described later, and
The information is stored in the history information storage section 4.

The correction rule determination unit 5 is the output of the manual correction control unit 3.

The correction rules in the correction rule table 2 are updated using the history information stored in the history information storage section 4, the new correction rule candidate list, and the correction rule update list.

The operating section 6 includes a keyboard, a mouse, a display, etc., and displays information from the manual correction control section 3 on the display, and also gives manual operation instructions to the manual correction control section 3 from the keyboard and mouse.

Next, the operation of this embodiment will be explained in detail with reference to the operation flowcharts shown in FIGS. 2, 3, and 4, and FIGS. 5 to 8. Here, FIG. 2 is an operation 70-chat of the manual correction control unit 3, and FIG.
) is an operation flowchart for subsequent manual correction. 3 is an operation flowchart of the automatic correction control section 3, and FIG. 4 is an operation flowchart of the correction rule determination section 5.

Although this embodiment can be applied to both character and voice pattern recognition processing, a specific example will be described below in which characters are targeted as input patterns.

The results recognized by the character recognition device are first input to the automatic correction section 1. The automatic correction unit 1 automatically corrects misrecognized characters while referring to the correction rule table 2. As described above, the automatic correction section 1 does not operate on the initial recognition result, and the single character recognition result is sent as is to the manual correction control section 3.

FIG. 5(A) shows the procedure for manual correction operations and correction rule generation by the operator for the recognition result a obtained by character recognition of the input sentence "Predict the weather tonight...". Further, FIG. 5(B) shows an example of a display on the display of the operation unit 6.

The manual correction control unit 3 displays the first candidate character as a result of the recognition of each character on the display of the operation unit 6 in a manner similar to the display example on the left side of FIG. 5(B) (as shown in FIG. Step 101). Next, looking at the characters displayed on the display, the operator corrects the misrecognized characters using the keyboard, mouse, etc. of the operation unit 6 (step 102). The correction method is to display and select the next candidates in order if the correct answer is in the candidate character set (in this example, the second candidate characters are used), and if the correct answer is in the candidate character set. If the correct character does not exist, enter the correct character using Kana-Kanji conversion input. The manual n correct result is shown in the display example on the right side of FIG. 5(B).

When the manual correction is completed, the operator sends a manual correction completion signal to the manual correction control section 3 from the keyboard of the operation section 6.

When a manual correction end signal is sent from the operation unit 6, the manual correction control unit 3 creates a relationship (correction rule) between the candidate character set of the recognition result of the character corrected by the operator and the correct character, and calculates the number of corrections. is counted (step 103). All correction rules generated by manual correction of initial character recognition results are sent to the correction rule table 2 via the correction rule determination unit 5.
(Step 104). Part of the contents of the correction rule table 2 in this case is shown in FIG. 5(A). As shown in the figure, for example, for the correct character ``日'', the candidate character sets ``self'' and ``white'' are registered as correction rules. This is expressed as ``Japanese confession.''

However, in this case, the number of corrections is 1 because the correction rule was generated by the first manual correction. .

Next, the manual authenticity control unit 3 processes all the correct characters as a result of the recognition and the corresponding candidate character set into a correction rule judgment (which will be performed later).
It is stored in the history information storage unit 4 as history information (correction results) for use in the operation flow shown in FIG. 4 (step 105). FIG. 6 shows the contents of the history information storage unit 4 after manual correction of the recognition result a shown in FIG. 5(A). Then, the manual correction process ends.

Next, assume that the input sentence ``It will be difficult to predict tomorrow's weather...'' is human-generated. FIG. 7 (Δ) shows the recognition result of this input sentence. The recognition result of this input sentence is first automatically corrected by the automatic correction section 1 shown in FIG. In this automatic correction, first, the similarity (hereinafter referred to as the degree of matching) between the candidate character set in the correction rules included in the automatically created correction rule table 2 and the candidates in the recognition result is calculated. Step 121 in Figure 3). This degree of matching is calculated by the following formula (
Calculate according to 1).

Matching degree - 2XL (x y)/ (L(x)-+-L(
y)) ・ (1) Here, X and y represent the candidate character set of the recognition result and the candidate character set of the correction rule, respectively, and L(x) and L(y) are the number of candidate characters, and L(x y
) represents a candidate character for the logical product set of x and y symbols. next,
It is checked whether the degree of matching is greater than or equal to a certain threshold (step 122). If the degree of matching is above a certain threshold, the correct character of this correction rule is replaced with the first candidate character of the candidate character set of the recognition results (step 123). After the correction rule has been replaced with the correct character in step 123, or when the degree of matching falls below a predetermined threshold in step 122 of FIG. 3, the process proceeds to step 124. Step 1
At step 24, the automatic correction unit 1 determines the recognition result and sends the automatic correction result to the manual correction control unit 3.

In the manual correction control unit 3, step 111 in FIG. 2(B)
, the first candidate character automatically corrected by the automatic correction section 1 is displayed on the display of the operation section 6 (
Example of display in the center of FIG. 7(B)). and,
The operator manually corrects the erroneous characters while looking at the display (step 112). The display example on the right side of FIG. 7(B) does not show the result of the manual correction described above. 21 When the manual correction is completed, the operator sends a manual correction completion signal from the operating section 6 to the manual correction control section 3.

When the operator sends a signal to complete the correction, the manual correction control unit 3 examines the characters manually corrected by the operator and creates the following three sets (step 113).

(1) 'FI1. The first candidate character in the recognition result is incorrect, and the correct automatically corrected character is l' in Figure 8.
[! ■ The first candidate character in the recognition result was incorrect, but the correction rule was not applied, and the character was correctly corrected by manual correction (in the same figure, the first candidate character was incorrect). ...J) C3) The first candidate character in the recognition results was correct, but
Characters that were incorrectly automatically corrected (as shown in the figure [...predicted Q is...]) From now on, these sets will be referred to as set 1. Set 2゜ Set 3
I will call it.

For each character in set 2, the manual correction control unit 3 converts the candidate character set of the recognition result and the correct character of the manually corrected result into a new M
” is sent to the correction rule determining unit 5 as a regular rule candidate list (step 114). Next, the manual correction control unit 3 collects the correction rules applied to the automatic correction of each character for sets 1 and 3, creates a correction rule update list, and sends it to the correction rule determination unit 5 (step 115). At that time, the list for updating correction rules includes information on whether the correction rules are those that were applied to correct corrections (set 1) or those that were applied to incorrect corrections (set 3). Also add information to identify it. In FIG. 8, ◯ indicates correct application, and × indicates incorrect application.

The correction rule determination unit 5 uses the new M correction rule candidate sent from the manual correction control unit 3 to automatically correct the history information (FIG. 6) in the history information storage unit 4 as described above. This is performed by determining the degree of matching (step 131). Here, compare the automatically corrected characters with the correct characters in the history information,
The number of times an incorrect character was correctly corrected (hereinafter referred to as the number of corrections) and the number of times a correct character was incorrectly corrected (
Hereinafter, this will be referred to as the number of false positives), calculate the reliability defined by formula (■), and then calculate the reliability defined by formula
, 5) Only the above new correction rule candidates act as new correction rules and are additionally registered in the correction rule table 2.

Reliability - Number of corrections/(Number of corrections + Number of incorrect corrections)...Next, the corrective rule determination section 5 compares the correction rule update list sent from the manual correction control section 3 with the correction rule table 2. However, regarding the correction rule included in the correction rule update list, if the correction rule was included in the set 1, 1 is added to the number of corrections of the corresponding correction rule in the correction rule table 2, and the correction rule is If the correction rule is included in the set 3, 1 is added to the number of incorrect corrections of the corresponding correction rule in the correction rule table 2 (step 132).

Finally, the correction rule determination unit 5 checks each 5 of the correction rule table 2.
For one regular rule, the reliability is calculated by formula [F] from the number of corrections and the number of incorrect corrections (step 13).
3). Then, correction rules whose reliability is above a certain threshold are removed from the correction rule table 2 (step 134.
135). For example, the correction rule 1 in FIG. Finally, after storing the history information regarding the recognition result in the history information storage section 4, the process ends.

The correction results thus obtained are output to any pig processing system (not shown).

As explained above, the relationship between the correct character and the candidate character set obtained through the operator's operation to correct the incorrectly recognized character is created as a correction rule candidate, and the operator completes the correction of the correction rule candidate. By applying this method to past character recognition results, subsequent character recognition results can be automatically corrected using only correction rules with a high rate of correct correction, thereby increasing recognition accuracy.

The thirty-one embodiments of the present invention described above are not limited to the above-mentioned embodiments, but can also be configured as follows.

M] The number of eye-stopping rules in the correction rule table 2 increases as the number of recognized characters increases, but each correction rule in the correction rule table 2 records the history of the automatic “J positive” application and uses @A The size of the correction rule table 2 can be kept within a certain range by removing from the correction rule table 2 rules with a small degree of accuracy or rules for which a long time has passed since the last application.

Furthermore, the amount of history information used to determine correction rules also increases as the number of recognized characters increases, but by storing only the latest recognition results for several pages, the size can be kept within a certain range.

Further, in the embodiment, learning was performed in units of one character, but the principles of the present invention can also be applied to learning and automatic correction in units of two or more adjacent characters.

Further, in the embodiment described in this text, the candidate characters in the recognition result are up to the second position, but in principle, the same applies no matter how many positions they appear.

In addition, in the embodiment, the correction rule generated from the first input unit (eg, page) is unconditionally adopted without determining whether or not to use it in the correction rule determination unit 5. If history information is built into the device in advance, or if the result of manually correcting the first input unit is used as history information, it is possible to make a judgment based on reliability calculation for the correction rule. .

In addition, in the example, the degree of matching defined by equation (1) as a measure of the similarity between the candidate character set of the correction rule and the candidate character set of the recognition result includes the rank of the candidate character set and the distance value of the recognition result. Although not taken into account, it is also possible to define a measure of similarity between candidate character sets that takes these into account and use it for automatic correction.

The above description has been made using letters as an example of the input cover turn, but when the input cover turn is a voice, the candidate category set of recognition results is output as a set of syllable candidates, as shown in FIG. In the example of Fig. 9, ↓ shows the result of speech recognition of ``Setsuzoku (connection)''. The subsequent processing is performed in the same way as the post-processing of the character recognition results.

FIG. 9 also shows correction rules obtained as a result of manual correction.

〔Effect of the invention〕

Since the present invention is configured as described above, it produces the effects described below.

In the structure of claim 1, when performing correction after pattern recognition, information on an operation by an operator to correct a pattern that has been erroneously recognized is automatically registered as a correction rule, and is applied to subsequent recognition results to correct recognition errors. Automatic correction is performed, and the processing method is simpler than correction methods based on re-registration of pattern feature amounts or grammar processing, so it is possible to speed up processing after pattern recognition.

In the structure of claim 2, since only highly reliable correction M rules are always stored by sequentially determining and updating correction rules, it is possible to easily follow changes in text types and pattern characteristics.
As it is applied, the reliability of automatic correction increases, creating a school deposit effect. Furthermore, word matching, morphological analysis, and the like prevent the generation of grammatically correct but incorrectly corrected sentences, making correction work easier for the operator.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the present invention, FIG. 2 is an operation flowchart of the manual correction control section, FIG. 3 is an operation flowchart of the automatic correction section, and FIG. 4 is an operation flowchart of the correction rule determination section. Fig. 5 is a diagram showing the manual correction operation by the operator for character recognition results and the procedure for generating correction rules, as well as an example of what is displayed on the display of the operation unit, Fig. 6 is a diagram showing the contents of the history information storage unit, and Fig. 7 8 is a diagram showing an example of an automatic correction for a character recognition result, a manual correction operation by an operator, and a display on the display of an operation unit, FIG. 8 is a diagram showing a flow of correction rule determination processing, and FIG. FIG. 3 is a diagram showing an example of a recognition result and a correction rule table when applied to speech recognition. 1... automatic correction section, 2... correction rule table, 3.
...Manual correction control unit, 4...M layer information storage unit, 5...
- Correction rule determination section, 6... operation section. Patent applicant Nippon Telegraph and Telephone Corporation

Claims (2)

[Claims] (1) In a pattern recognition method that recognizes input patterns such as speech and text and generates one or more candidate categories, a set of candidate categories automatically extracted from information that allows the operator to correct categories of patterns that are incorrectly recognized. If a correction rule consisting of the relationship between correct categories is applied to subsequent pattern recognition results, and the candidate category set in the correction rule and the candidate category set included in the pattern recognition result are similar to each other,
A pattern recognition post-processing method characterized by replacing the correct category of the correction rule with the first candidate category of the pattern recognition result. (2) The automatically extracted correction rule is applied to a set of candidate categories included in the pattern recognition results for which corrected correct answers are known, and the misrecognized categories for the total categories corrected by the replacement are correctly corrected. 2. The pattern recognition post-processing method according to claim 1, wherein only correction rules for which a ratio of the total number of categories exceeds a certain value are applied to subsequent correction of pattern recognition results.