JP2000339412A - Recognition information correction system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an efficient recognition system where erroneous recognition which frequently occurs is reduced in recognition of an OCR device and a sound recognition device and correction work by a manual operation is suppressed to be minimum. SOLUTION: An OCR device 101, a word judgement means 102, a word dictionary 103, an example collation processing part 104, an example data base 105, an automatic correction means 106, an input correction part 107, a syntax judgment means 108, a syntax dictionary 109, an example collation processing part 110, an example data base Ill, an automatic correction means 112, an input correction part 113 and an output means 114, are included. Correction work by a manual operation can be suppressed to be minimum even while a recognition device by such conventional technology is used.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to an OCR (Optical Character)
ACTER READER), and relates to a recognition information correction system that corrects erroneous recognition of at least one recognition device selected from speech recognition devices.

[0002]

2. Description of the Related Art In a document processing apparatus, an OCR is used to read a sheet of paper such as a newspaper or a magazine, and the result is stored, and a document process such as machine translation and proofreading support is performed. However, reading by OCR often involves misrecognition and manual correction is indispensable for correction.

[0003] If the contents of conversations such as consultations are documented, it is useful for managing user information and consultation contents. In recent years, speech recognition technology has advanced, and the possibility of performing transcription by speech recognition has appeared (see Reference: Nikkei Electronics 19).
January 11, 1999, pp.43-50). However, the accuracy of transcription by speech recognition is often not sufficient for practical use.

[0004]

As described above, in the prior art, the recognition result of the recognition device must be manually corrected, and the labor cost that is often required becomes enormous, so that practical use is not possible. It was not a target.

An object of the present invention is to provide a practical recognition system (= recognition information correction system) as a total system while minimizing manual correction work while using a conventional recognition device. I have.

[0006]

According to the present invention, in order to solve the above-mentioned problems, an OCR (optical character reader) is provided.
R), at least one recognition device selected from the voice recognition device, a determination unit that analyzes text information as a recognition result of the recognition device and determines necessity of correction, and corrects the determination result of the determination unit. Display means for displaying to the operator the part where the necessity has been found, a case database storing cases in which recognition errors have been corrected, and automatically correcting any cases by referring to the cases; And a learning means for registering the correction result of the operator as a case in the case database as a case.

The determination of the necessity of the correction is performed by referring to a word dictionary prepared in advance, extracting a character string that is not in the word dictionary as an unknown word, and determining whether or not the unknown word exists.

If the word dictionary is made by collecting words extracted by analyzing materials of a business environment using the system, the size of the word dictionary can be minimized. The extraction of unknown words is efficient.

[0009] Further, the determination of the necessity of correction is performed by referring to a syntax dictionary prepared in advance and extracting a syntax that does not match the structure registered in the syntax dictionary from the text information as the recognition result. It is also a good way to do this.

In the learning means, if at least one selected from a word not in the word dictionary and a syntax in the syntax dictionary is extracted by the correction input, if it can be registered in the word dictionary or the syntax dictionary, the learning means Increases the efficiency of the system.

[0011] Further, a process of temporarily storing a signal to be converted into a text, performing a process of recognizing the signal and determining a recognition result at least twice, and performing a comparison process on the plurality of determination process results is performed. By providing a processing unit and selecting and joining portions that do not require correction, signals can be converted to text with the minimum necessary correction processing.

[0012]

Embodiment 1 Hereinafter, a first embodiment of the present invention will be described in detail. FIG. 1 is a block diagram of the system of the present invention. O
CR device 101, word determination means 102, word dictionary 10
3, case collation processing unit 104, case database 105,
An automatic correction unit 106, an input correction unit 107, a syntax determination unit 108, a syntax dictionary 109, a case collation processing unit 110, a case database 111, an automatic correction unit 112, an input correction unit 113, and an output unit 114 are provided.

The OCR device 101 converts the image data read by the scanner into text and extracts text information. This text information is stored in the
02. Here, an OCR device is used as the recognition means, but a voice recognition device may be used.

The word judging unit 102 refers to the word dictionary 103, performs morphological analysis of the recognition data, extracts words not in the dictionary as unknown words, and judges the necessity of correction based on the presence or absence of the unknown words. If no correction is necessary, the text information is sent to the syntax determination means 108 as it is. If the word needs to be corrected, the word case matching unit 104 refers to the case database 105, and if there is a case, it is sent to the automatic correcting unit 106 and corrected. If there is no case, the input correction unit 107 corrects the case according to the input by the operator. The corrected data is sent to the syntax determining means 108.

The input in the input correction unit is registered in the case database 105 as a correction case. At this time, if the correction input is an unregistered word, it is registered in the word dictionary 103.

The syntax determining means 108 refers to the syntax dictionary 109, extracts a syntax that is not in the syntax dictionary from the sent data as an unknown syntax, and determines the necessity of correction based on the presence or absence of the unknown syntax. Output means 1 if no correction is needed
14 is sent. If correction is necessary, the case database 111 is referred to, and if there is a case, it is sent to the automatic correction means 112 and automatically corrected. If there is no case, input correction unit 11
At 3 the correction is made by the input of the operator. The corrected data is sent to the output means and output.

FIG. 2 shows an input correction unit (107, 11 in FIG. 1).
This shows the configuration of 3). Upon receiving the data determined by the case determining means that there is no case, the display means 201 displays a message prompting the operator to make a correction input. The display is desirably a display device such as a monitor, but there is no problem in pointing out a correction portion and input instruction by a synthesized voice. The operator performs a correction input using at least one input means 202 selected from a keyboard, a voice input, and a touch panel. In the registration means 203, the words and the syntax of the pattern not in the dictionary are registered in the dictionary 204, and the case of the correction input is registered in the case database 205, respectively.

[0018]

[Embodiment 2] FIG. 3 is a flowchart showing a second embodiment. The audio signal is recorded by a microphone (not shown), and the audio data is stored by the temporary storage processing 301. The voice recognition processing 304 is performed on the temporarily stored data. In the determination processing 305, a part requiring correction is extracted from the recognition result with reference to a dictionary (not shown) (extraction processing 302A).

Further, extraction processing 302B is performed on the temporarily stored data, and comparison processing 303A is performed.
In the comparison process, the portion requiring correction extracted as a result of the extraction process 302A is replaced with a portion corresponding to 303B determined to have been successfully recognized (portion not determined to require correction).

The comparison processing 303A is performed on the data generated in the comparison processing 303A using the result of the extraction processing 303C.
Perform B. Correction processing 3 for the data thus generated
Send to 06. In the correction process 306, a case database (not shown) is referred to, and if there is a case, automatic correction is performed, and if there is no case, input correction is performed. The output unit 307 outputs text information as a product.

[0021]

According to the apparatus of the first embodiment, which has been learned by reading the literature in the scientific field 1000 times as a material,
When a test sentence of 316 words was input, the correct answer rate of 92.7% on average before correction processing was 9
The accuracy rate was improved to 9.3%.

When the same test sentence was read aloud and its voice data was used by the apparatus of the second embodiment, the correct answer rate was 83.5% in the method without the comparison processing.
The answer rate was improved to 91.9%. As described above, it is possible to minimize the manual correction work while using the recognition device according to the related art.

[0023]

[Brief description of the drawings]

FIG. 1 is a block diagram of a first embodiment;

FIG. 2 is a configuration diagram of an input correction unit

FIG. 3 is a block diagram showing a second embodiment.

Claims (6)

[Claims] 1. A system for converting at least one signal selected from an image and a sound into a text, comprising:
CAL CHARACTER READER), at least one recognition means selected from a voice recognition device, a determination means for analyzing text information as a recognition result of the recognition means and determining the necessity of correction, and a determination result of the determination means Display means for displaying to the operator a portion that has become necessary to be corrected, a case database storing a case in which the recognition error has been corrected,
A correction means for automatically correcting if there is a case with reference to the case, and correcting the recognition result by a correction input of an operator if not, and a learning means for registering the correction input of the operator as a case in the case database. A recognition information correction system provided as a component. 2. The method according to claim 1, wherein the determination of the necessity of the correction is performed by referring to a word dictionary prepared in advance, extracting a character string not in the word dictionary as an unknown word, and determining whether or not the unknown word exists. The recognition information correcting system according to claim 1. 3. The word dictionary according to claim 2, wherein the word dictionary is created by collecting words extracted by analyzing data of a business environment using the system. Recognition information correction system. 4. The determination of the necessity of correction is performed by extracting a sentence that does not match a syntax registered in the syntax dictionary from text information as the recognition result by referring to a syntax dictionary prepared in advance. 2. The recognition information correction system according to claim 1, wherein the recognition information correction is performed based on the presence or absence of a syntax. 5. When at least one selected from a word that is not in the word dictionary and a semantic structure that is not in the syntax dictionary is extracted by the correction input in the learning means, the word is registered in the word dictionary or the syntax dictionary. 2. The recognition information correcting system according to claim 1, wherein: 6. A process for temporarily storing a signal to be converted to text, performing a process of recognizing the signal and determining a recognition result at least twice, and performing a comparison process on the plurality of determination processes. 2. The recognition information correcting system according to claim 1, further comprising a comparison processing unit, and a step of selecting and joining unrequired portions.