JPH079654B2 - Katakana word error detection and correction device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial field of application] The present invention automatically creates typographical errors and inconsistencies in katakana contained in an input Japanese character string when creating a Japanese document database or the like. The present invention relates to a katakana word error detection and correction device that automatically detects candidate character strings for correction in the case of typographical error.

[Conventional technology]

With the internationalization of various technologies and information, as well as the transition of Japanese language, words in katakana notation have been frequently used in Japanese documents. For this reason, when converting Japanese documents such as various manuals, scientific papers, manuscripts for publication, and office documents into electronic files, katakana words frequently appear for technical terms, proper nouns, or some common words. Accompanied by,
Misunderstanding of the creator (wrong: interface / correct: interface), input error of document input device represented by Japanese word processor (erroneous: computer / correct: computer), character recognition error represented by Kanji OCR (error:
Various errors such as pitch / correct: pitch) occur. Moreover, not only typographical errors but also fluctuations in the notation according to the document creator frequently occur in katakana words. For example, fax / fax / facsimile are not necessarily erroneous as individual terms, but since they have the same meaning, the notation should be unified.

Conventionally, in such a situation, in the proofreading performed manually, the work of detecting errors in katakana terms and fluctuations in the notation is often due to foreign words in the katakana term, and there is also a tendency to pronounce the katakana as it is. Therefore, compared with kanji terms and hiragana terms, (1) I can't remember all the correct notations of the example. Example: Positive simulation / simulation (mocking) (2) Errors are easily overlooked due to the similarity of the characters. Example: Correct debugging / debugging (to detect program errors) (3) Difficult to extract terms with scattered fluctuations in the document (4) Concerning terms with fluctuations in the notation and unified notation Due to factors such as not being able to remember all
It is difficult to detect errors in katakana characters and unification notation, and to extract correct candidate words when they are detected, and there is a problem that the accuracy of proofreading decreases and the processing time increases.

On the other hand, in order to improve the accuracy of word recognition, a method has been devised in which a reading result of a recognition device is collated at high speed with a recognition dictionary composed of correct terms and a recognition error is eliminated based on the degree of the collation value. (For example, Japanese Patent Application No. 61-248415
issue). FIG. 6 shows the flow of processing of the correction means by the character matching method of this system. Here, 101 is an error character string of the detected katakana, 102 is a typographical error, 103 is a normal character, 104 is a previously created katakana word candidate sequence (a part of the recognition dictionary), and 105 is a character to which a matching value of 1.0 is given. And its position, 106 is the matching value 0.8
Is a character and its position where 106 is moved backward by one character, 107 is a character and a position where a matching value is 0.8 and 105 is moved forward by one character, and 108 is each word candidate As a result of the collation, 109 is an evaluation value, and 110 is a word candidate having the highest evaluation value.

In this example, when the recognition result is a katakana error character string “decoral” (correct answer is “digital”) indicated by 101, the katakana character string is extracted from the recognition result and the position of each character is determined according to the position. Collation values 105, 106, 107 are set respectively. Here, the reason why a collation value is given to a character that is moved forward or backward by one character, such as 106 and 107, is because it is assumed that a missing character or an incorrect character is inserted in addition to the incorrect character. In this case, the collation value is 0 because the occurrence frequency of missing characters and insertion errors is generally lower than that of typographical errors.
It's lowered to 8. After setting the matching value in this way, the matching value set in 105, 106, 107 is given to each candidate of the katakana word candidate sequence 104 which is a part of the recognition dictionary according to the matched character position. Here, 0 is given when there is no corresponding character. After that, an evaluation value (total value of matching values) 109 is calculated for each candidate, and “digital” is extracted as the word candidate 110 having the highest evaluation value.

As described above, the correction value “digital” for the incorrect katakana word “decoral” can be extracted by setting the matching value according to the katakana character and its position and matching with the word candidate. However, with this method, when targeting katakana words that appear in an actual document, the number of word candidates becomes enormous, so that the number of matching targets becomes large and the search time and matching time of word candidates increase. . Also,
When an error occurs in the sound of a katakana word, there is a problem that the wrong word candidate may be extracted because the characters have different shapes.

The 7th case of incorrect correction candidate extraction due to the generation of goose
Shown in the figure. Here, 111 of "x" mark is a word candidate extracted by mistake. This example is for the case of an error in the sound of chorus, which occurs frequently in Katakana-to-Kanji conversion of Roman alphabet input in a Japanese word processor. In other words, “SO” and “SYO” are caused by one key operation error during input,
Moreover, because the character shape is completely different, the Katakana word candidate containing "SHO" is given a higher evaluation value than the Katakana word candidate containing "SO", and the correct word "Solid" is not extracted at the top and is incorrect. The word candidates that have been extracted will be extracted.

In addition, the same katakana word is often used in a document many times, but in this case, even if the katakana word with the same meaning has fluctuation in its notation, it may be properly used. is there. For example, fax / fax / facsimile etc. have the same meaning, so the notation should be unified. However, any katakana word notation is a correct notation, and since it is not detected as an error, conventionally it is not possible to detect the error of the katakana word due to the fluctuation of the notation, and it is possible to extract the correction candidate accordingly. could not.

[Problems to be Solved by the Invention]

As described above, conventionally, with the increase in the number of katakana word candidates,
There has been a problem that the accuracy of manual error detection and the accuracy of extraction of correction candidates are reduced and the processing time is increased, which is a burden on the calibration work. In addition, in the method of extracting correction candidates by collating the katakana character string with a candidate dictionary of correct katakana words at high speed, when the katakana words that appear in the actual document are targeted, the number of word candidates becomes enormous and the word candidates are searched. There has been a problem that the processing time such as time and matching time increases, or that the accuracy of extraction of correction candidates decreases when an error occurs in jumble. Furthermore, there is a problem that correction candidates cannot be extracted for katakana words that have fluctuations in the notation.

It is an object of the present invention to provide a Katakana word error detection and correction device which solves the above-mentioned conventional problems and improves processing speed and correction accuracy.

[Means for Solving the Problems]

In order to achieve the above-mentioned object, in a katakana word error detection and correction device of the present invention, a morphological analysis for performing word division and part-of-speech recognition on an input Japanese sentence character string using a Japanese word dictionary and a grammar dictionary. A means, a katakana word extracting means for extracting a katakana word consisting of a katakana character string recognized by the word division and part-of-speech recognition together with the division information and the certification information, and when there is no corresponding katakana word in the Japanese word dictionary, Katakana word error detection means for recognizing an error in a katakana word by recognizing the case where there is no grammatical connection, and two consecutive characters forming the character string of the katakana word as a correction candidate and its position are used as keys. All combinations of a Katakana word candidate dictionary created in advance, a Katakana character string in which an error is detected, and a character string obtained by shifting this character string by one character before and after Katakana word correction candidate group extracting means for extracting the primary correction candidate group including the correct katakana word by using the two katakana characters and their positions as keys, and the katakana word in which the error is detected. For the character string and the extracted Katakana word correction candidate group, the Katakana word consonant vowel separating means for converting the Katakana character string into Roman letters and further separating it into the consonant string and the vowel string symbol string, and the Katakana character in which the error is detected. With respect to the column and the extracted katakana word correction candidate group, a katakana word collating means for collating a symbol string separated into a consonant sequence and a vowel sequence separately from the front and the back to obtain a collation value, and a collation value set in advance. And a katakana word candidate selecting unit that selects the extracted katakana word correction candidate as a correct answer candidate when the value is higher.

Furthermore, from the Japanese sentence character string input by the document input device to a katakana word with fluctuations in the document,
Katakana character string extraction means for extracting the katakana character string, katakana character string consonant vowel separation means for converting the extracted katakana character string into Roman letters, and further separating the consonant string and the vowel string symbol string, and the extracted mutual For the Katakana character string of No., if the concatenation string and the vowel string of the Katakana character string are separately collated from the front and the back to obtain the collation value, and the collated character strings are not the same In addition, when the matching value is higher than a preset value, there is provided means for detecting that the katakana word notation is inconsistent between the extracted katakana character strings.

[Action]

In the present invention, in order to correct an input error and a character recognition error mixed in a katakana word, a katakana word candidate dictionary having two consecutive characters and its position as a key for correcting a character string of a katakana word that is a correction candidate is used. Created in advance, the katakana character string detected by the fact that the recognition result is an unknown word and the character string obtained by shifting this character string by one character before and after are all two combinations of katakana characters and their positions as keys,
The Katakana word candidate dictionary is searched to extract a first-order correction candidate group including the correct correct Katakana word, and the detected Katakana character string and the extracted Katakana word candidate group are each a Katakana character string. Is converted to Roman letters and is further separated into consonant and vowel symbol strings, and the separated symbol strings are collated separately from the front and the back to calculate the collation value, and the collation value is higher than the preset value. In this case, the extracted katakana word correction candidate is selected as the correct answer candidate.

Furthermore, for katakana words that have fluctuations in the document, extract the katakana character strings in the document, convert each of the extracted katakana character strings to romaji, and then convert them into consonant and vowel string symbols. Separate the symbol strings separated into consonant strings and vowel strings separately from the front and back to calculate the matching value, and if the matching character strings are not the same, and the matching value is When the value is higher than a preset value, it is detected that the katakana word notation is inconsistent between the extracted katakana character strings.

As a result, it detects errors in Katakana words that are difficult to proofread manually, including errors, among input errors in document input devices such as Japanese word processors, character recognition errors, and mistakes made by the creator. Even in this case, it is possible to extract correction candidates with high correction accuracy, and moreover, correction candidates in which at least one or more two character parts match are narrowed down as a primary correction candidate group to be collated, so that the processing speed is greatly improved. be able to. Furthermore, it detects the inconsistency in the notation of katakana words contained in the input document and extracts correction candidates, so it is possible to reduce the load of the proofreading work of the inconsistency in the notation and improve the correction accuracy. it can.

〔Example〕

 An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of a katakana word error detection / correction apparatus showing an embodiment of the present invention. In FIG. 1, 10 is a document input device, 20 is a processing device composed of a CPU and a memory, and 3
0 is a Japanese word dictionary, 40 is a grammar dictionary, 50 is a Katakana word candidate dictionary with two katakana characters and their positions as keys, and 60 is an error-corrected document. CPU / memory processor 20
In terms of processing function, is a morphological analysis processing unit 22 that performs word division and part-of-speech recognition using the input processing unit 21 and dictionaries 30 and 40.
And a Katakana word extraction processing unit 23 using word division and part-of-speech recognition of the morpheme analysis processing unit 22, a Katakana word error detection processing unit 24 for detecting an error in a Katakana word, and a Katakana word candidate dictionary 50. A Katakana word correction candidate group extraction processing unit 25 for extracting the first correction candidate group, and a Katakana word consonant vowel separation processing unit for converting a Katakana character string into Roman characters and further separating it into a consonant string and a vowel string symbol string. 26, a katakana word matching processing unit 27 that matches a symbol string separated into a consonant string and a vowel string to extract a matching value, and a Katakana word candidate selection processing unit 28 that selects a correct answer candidate.

In the processing device 20, in the input processing unit 21, the document input device
After converting the input character code into a processable Japanese character string from 10, the morphological analysis unit 22 uses the Japanese word dictionary 30 and the grammar dictionary 40 to perform word division and part-of-speech recognition. With respect to this word division and part-of-speech recognition result, in the katakana word extraction processing unit 23, a Katakana word consisting of a Katakana character string is extracted together with the division information and the certification information, and in the Katakana word error detection processing unit 24, the Japanese word dictionary 30 An error in a katakana word is detected using the unknown word recognition result that occurs when there is no corresponding katakana word or when there is no grammatical connection relationship.

On the other hand, a Katakana word candidate dictionary 50 is prepared in advance using two consecutive characters that form a character string and the position of the Katakana word that is a correction candidate as a key.

In the Katakana word correction candidate group extraction processing unit 25, the Katakana character string in which an error is detected by the Katakana word error detection processing unit 24 and the character string obtained by shifting the character string by one character before and after the character string are all adjacent to each other. The katakana word candidate dictionary 50 is searched using the characters and their positions as keys to extract a first-order correction candidate group including correct katakana words. Next, regarding the katakana character string detected by the katakana word error detection processing section 24 and the correction candidate group of the primary katakana word extracted by the katakana word correction candidate group extraction processing section 25, the katakana word consonant vowel separation processing section 26. In, each Katakana character string is converted to Roman letters, further separated into consonant strings and vowel string symbol string, further, in the Katakana word matching processing unit 27, each separated consonant and vowel symbol string separately forward and. The collation value is calculated by collating from the back. And a katakana word candidate selection processing unit
In 28, if the matching value is higher than the preset value,
The extracted katakana word correction candidate is selected as the correct answer candidate, and the error-corrected document 60 is created.

FIG. 2 is an explanatory diagram showing a specific example of the processing of the katakana word candidate group extraction processing unit 25 for extracting the primary katakana word correction candidate group in FIG.

In FIG. 2, 101 is an error character string of the detected katakana, 102 is a typographical error, and 103 is an orthographic character. 121 is a character string for extraction of the primary correction candidate group, 121-1 is a character string directly extracted from the error character string 101, 121-2 is a character string obtained by moving 121-1 forward by one character, 121-3 Is a character string obtained by moving 121-1 backward by one character. 122 is 121-1, 121-2 and 121-3
It is correction candidate search data which is a key for searching the Katakana word candidate dictionary 50, which is composed of two Katakana characters extracted from a combination of all two adjacent characters and their positions. 123 is the two-character part of the katakana of 122, 124 is the character position of the first character of 123, 123-1 to 123-3 and 124-1 to 124-
Reference numeral 3 is a part of each correction candidate search data. 13
0 is the key part of the katakana word candidate dictionary 50, 131 is the heading part of the katakana word, 131-1 is the primary correction candidate group extracted in this example, 131-2 is the katakana word candidate not extracted, 132
Is a two-character katakana character portion of the key portion 130 of the katakana word candidate dictionary, 133 is a character position of the key portion 130, and 134 is a pointer to the heading portion 131 of the katakana word indicated by the key portion 130. Reference numeral 135 denotes the extracted primary katakana word correction candidate group.

Here, a process of extracting from the katakana word candidate dictionary 50 a small number of first-order katakana word correction candidate groups that are targets of the collation process for the detected katakana character string will be described.

First, the correction candidate search data 122 is created as a key for detecting the katakana word candidate dictionary 50 from the “modanibunu” of the detected katakana character string 101. For this reason, "modanibunu" of the katakana character string 121-1 and the character string 121-1 were moved forward by one character, and "danibnu" of the character string 121-2 and the character string 121-1 were moved backward by one character. “Δmodanibunu” (Δ is a blank character) of the character string 121-3 is created as the correction candidate extraction data 121, and the correction candidate search data 122 consisting of two adjacent katakana characters and the position of the leading character 122 of all combinations. To extract.

For example, the first character "Da" in 121-2 has "Dani", "Dada", and "Damo" as the combination of the two characters.
The character position of the leading character of 123-1 "Dani" is the first character, so 124-1 becomes "1". Similarly, the first character “MO” of 121-1 is “MONI”, “MODA”, and “MOMO” as the combination of the two characters, and the “MODA” of 123-2 in this is the first character. Since the character position of is the first character, 124-2 becomes "1". Similarly, the second character “ni” of 121-2 has “nib”, “nii”, and “nida” as the combination of the two characters, and the “nib” of 123-3 in this is the beginning of that. Since the character position of the character is the second character, 124-3 is “2”.

In FIG. 2, correction candidate search data 122 is displayed with 123 and 124 shifted according to the difference in character position.
Here, in order to create the correction candidate extraction character string 121,
Shifting the detected character string 101 by 1 character before and after is because the detected character string may not only contain typographical errors but may also include missing characters, erroneous insertions, and character replacements. This is because it is necessary to comprehensively search the candidate group.

Next, the katakana word candidate dictionary 50 is searched using the extracted correction candidate search data 122. For example, 123-2
In the case of “moda” and “1” of 124-2, since the katakana word candidate dictionary 50 is searched using the two characters “moda” from the first character as a key, in this example, “modernist”, “modernism”, The four correction candidates 131-1 of "modality" and "modern madam" can be extracted. Similarly, by searching the katakana word candidate dictionary 50 using all two characters of the correction candidate search data 122 and the character position information thereof as keys,
The primary correction candidate group 135 can be extracted. Where 131-2
The “moment” is not related to the correction candidate search data and is not extracted.

In this way, since it is possible to search for and narrow down at least one or more two-letter parts of the katakana word candidates that are the target of the matching process, it is possible to perform conventional manual processing and computer-based all katakana The search time can be significantly reduced, and the processing accuracy can be improved, as compared with the collation based on the search of word candidates.

FIG. 3 is a specific example of the processing of the Katakana word consonant vowel separation processing unit 26, the Katakana word matching processing unit 27, and the Katakana candidate selection processing unit 28 for matching the Katakana word correction candidates and selecting the correct answer candidate in FIG. FIG.

In FIG. 3, 101 is an error character string of the detected katakana, 102 is a typographical error, and 103 is an orthographic character. 141 is a conversion table for separating consonant and vowel symbol strings from Katakana characters. 142 is a character to be separated of the detected katakana character string 01, 143 is a consonant symbol string, and 144 is a vowel symbol string. 145 is a katakana word candidate. Reference numeral 146 indicates matching with each candidate from the front side, 147 indicates matching with each candidate from the rear, 148 indicates matching of consonant symbol strings, and 149 indicates matching of vowel symbol strings. 150 is the result of matching the symbol strings, 150-1 is the matching part, 150
-2 is a portion that does not match at all from the beginning, 151 is a matching value by the matched character string of the consonant symbol string, 152 is the total matching value in the matching from the front, 153 is the number of matching target symbol strings, 154
Is the total matching value. Reference numeral 155 is a correction candidate selected by the total matching value.

Here, "Shorid" of the detected katakana character string 101
With respect to the extracted first katakana word candidate group (in this example, “shopping” and “solid”), the katakana character strings are converted into roman letters, and further separated into consonant strings and vowel string symbols, The consonant symbol sequence and the vowel symbol sequence that have been separated are collated separately from the front and the rear, and a collation value is calculated to obtain a total collation value 154 for each candidate. For example, as a consonant sequence and a vowel sequence, “Solid” to “SYRD” “OIO”, and correction candidates “Shopping” to “S”, respectively.
"SYD" and "OIO" are converted and separated from "YPNG", "OIU", and "solid". Furthermore, the collation from the front is performed in 146, and the collation from the rear is performed in 147, and the number of coincidences is calculated as a collation value. If there is no match, it will be 0.0. Next, a comprehensive evaluation value 154 indicating the degree of matching is obtained by the following formula.

As a result, when the total matching value is higher than the preset value (1.0 in this example), it is selected as a katakana word correction candidate. In this example, since “solid” has a total collation value of 1.5, it is selected as 155.

In this way, even for errors in katakana words that include a syllabary that could not be dealt with by the conventional word matching method using character shapes, correct candidates can be selected by finely matching the correction candidates, and the correction accuracy can be selected. Can be improved.

FIG. 4 is a block diagram showing another embodiment of the present invention, showing a katakana word error detection / correction device having a function of detecting inconsistency of notation of katakana words. In FIG.
Reference numeral 10 is a document input device, and 20 is a processing device including a CPU and a memory. Here, in terms of processing functions, the processing device 20 includes an input processing unit 21, a Katakana character string extraction processing unit 23 ′ for extracting a Katakana character string from a character string of an input document, and a Roman character for the Katakana character string. Katakana word concatenation processing that calculates the matching value by matching the katakana word consonant vowel separation processing unit 26 that is converted to and further separated into the consonant string and the vowel string symbol string and the symbol string separated into the consonant string and the vowel string. Section 27 and katakana word notation inconsistency detection processing section 29 for detecting inconsistency in notation of katakana words
Can be divided into

In the processing device 20, in the input processing unit 21, the document input device
After converting the character code input from 10 into a Japanese character string that can be processed, the Katakana character string extraction unit
Extract at 23 '. For all of the extracted Katakana character strings, first, the Katakana word consonant vowel separation processing unit
In 26, each of the katakana character strings is converted into Roman letters, and further separated into consonant and vowel character symbol strings, and then in the katakana word matching processing section 27, the separated consonant and vowel symbol strings are separated. The collation value is calculated by collating from the front and the rear. Then, in the katakana word notation inconsistency detection processing unit 29, when the mutual character strings to be collated are not the same and when the collation value is higher than a preset value, the notation is not uniform between the extracted katakana character strings. To detect the possibility of

FIG. 5 is an explanatory diagram showing a specific example of the processing for detecting the inconsistency of the notation of the katakana word in FIG.

In FIG. 5, 161 is a katakana character string extracted from the input Japanese sentence character string, 162 is an orthography for unification of notations, 163 is a consonant symbol string converted from each katakana character string, and 164 is converted A symbol string 165 of the separated vowels is a processing example.

In this example, katakana words such as "photograph" and "cut glass" are written as correct katakana characters in the input document, and fluctuations "photo" and "hot" (correct answer: hot) Or the typographical error "Grak" (correct answer:
This is the case when a word including (graph) exists. Even in this case, by separating the input katakana word into consonant and vowel symbol strings and collating them, the “photograph” is “photograph” and the total collation value is 1.0 (in the embodiment,
Since the set value of the collation value that is regarded as the fluctuation of the notation is 1.0), it is understood that the notation is inconsistent due to the fluctuation of the notation of the katakana word. However, since the total matching values of "cut glass" and "photograph" or "photograph" are 0.4, respectively, the degree of similarity is low and it can be seen that they are completely different words. In this way, the inconsistency in the notation of scattered Katakana words contained in the input document is extracted from the document without performing the process of searching the Katakana word candidate dictionary even if it is not an error. Since it can be detected by collating the Katakana character strings with each other, it is possible to improve the detection accuracy and processing efficiency of the unification notation of the Katakana word.

As described above, in the katakana word error detection / correction device of the present invention, a katakana word candidate dictionary in which two consecutive characters that form a character string and a character position thereof are used as keys for a correction candidate katakana word is created in advance, Convert the character code input from the document input device into a processable Japanese character string,
If an error in a Katakana word is detected due to an unknown result of the morphological analysis of the converted Japanese sentence character string, two Katakana characters and the character position of the detected Katakana character string are used as keys in advance. The created corresponding Katakana word candidate dictionary is searched to extract the primary correction candidate group, and the detected Katakana word and the extracted primary correction candidate group are converted into Katakana character strings, respectively. , Further, it is separated into consonant strings and vowel character strings, and the separated symbol strings are collated separately from the front and the back to extract the matching value. When the matching value is higher than a preset value, the extracted katakana A word correction candidate is selected as a correct answer candidate.

Furthermore, for katakana words that have fluctuations in the document, extract the katakana character strings in the document, convert each of the extracted katakana character strings to romaji, and then convert them into consonant and vowel string symbols. Separate the symbol strings separated into the consonant sequence and the vowel sequence separately from the front and back to calculate the matching value, and if the matching character strings are not the same, and the matching value is When the value is higher than a preset value, it is detected that the katakana word notation is inconsistent between the extracted katakana character strings.

As a result, it is possible to detect an error in a katakana word that is difficult to proofread manually, and to extract a correction candidate with high correction accuracy even if it contains a syllabary, and moreover, a correction candidate in which at least one two-character part matches Is narrowed down as the primary correction candidate group to be collated, so that the processing speed can be greatly improved. Furthermore, it detects the inconsistency in the notation of katakana words contained in the input document and extracts correction candidates, so it is possible to reduce the load of the proofreading work of the inconsistency in the notation and improve the correction accuracy. it can.

〔The invention's effect〕

According to the invention of claim (1), an error in a katakana word which is difficult to proofread manually by an input error or a character recognition error in a document input device such as a Japanese word processor or a misunderstanding of a creator is detected, It is possible to extract correction candidates with high correction accuracy even when they include syllabary, and moreover, the correction candidates in which at least one two-character portion matches are narrowed down as the primary correction candidate group to be collated,
The processing speed can be greatly improved.

Further, according to the invention of claim (2), since the inconsistency of the notation of the katakana word included in the input document is detected and the correction candidate is extracted, the load of the proofreading work of the inconsistency of the notation manually is performed. And the correction accuracy can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram of a Katakana word error detection / correction device showing an embodiment of the present invention, and FIG. 2 is an explanatory view showing a concrete example of a process of extracting a primary katakana word correction candidate group in the present invention. FIG. 3 is an explanatory diagram showing a specific example of a process of collating katakana word correction candidates and selecting a correct answer candidate according to the present invention;
FIG. 5 is a block diagram of a katakana word error detection and correction apparatus having a function of detecting inconsistency of katakana word notation according to another embodiment of the present invention. FIG. 5 detects inconsistency of notation of katakana words in the present invention. FIG. 6 is an explanatory diagram showing a concrete example of the process, FIG. 6 is an explanatory diagram showing a flow of a process of extracting a correction candidate in the conventional method, and FIG. 7 is an explanatory diagram showing an example of erroneous correction candidate extraction in the conventional method. . 10 ... Document input device, 20 ... Processing device, 21 ... Input processing unit, 22
... Morphological analysis processing unit, 23 ... Katakana word extraction processing unit, 23 '
... Katakana character string extraction processing unit, 24 ... Katakana word error detection processing unit, 25 ... Katakana word correction candidate group extraction processing unit, 26 ... Katakana word consonant vowel separation processing unit, 27 ... Katakana word matching processing unit, 28 ... Katakana word Candidate selection processing unit, 29 ... Katakana word notation inconsistency detection processing unit, 30 ... Japanese word dictionary, 40 ... Grammar dictionary, 50 ... Katakana word candidate dictionary, 60 ... Error corrected document.

Claims (2)

[Claims] 1. A Katakana word error detection and correction device for detecting an input error or character recognition error mixed in a Katakana word and extracting a correction candidate for the error, a Japanese word dictionary and a grammar for an input Japanese sentence character string. Morphological analysis means for word division and part-of-speech recognition using a dictionary, katakana words consisting of katakana character strings recognized by the word division and part-of-speech recognition, and katakana word extraction means for extracting together with the division information and the certification information, A Katakana word error detection unit that detects an error in a Katakana word by recognizing that there is no corresponding Katakana word in the Japanese word dictionary or that there is no grammatical connection, and a character string for the Katakana word that is a correction candidate. A means for creating in advance a katakana word candidate dictionary having two consecutive characters and its position as keys, and the above-mentioned error-detected katakana character string. The first correction candidate that includes the correct katakana word by searching the katakana word candidate dictionary using two adjacent katakana characters of the combination and their position as the key Katakana word correction candidate group extraction means for extracting the group, the error detected Katakana character string and the extracted Katakana word correction candidate group, respectively, to convert the Katakana character string to Roman letters, further consonant strings and vowel strings Katakana word consonant vowel separating means for separating into a symbol string, the error-detected Katakana character string and the extracted Katakana word correction candidate group, the symbol string separated into consonant strings and vowel strings separately forward and backward Katakana word matching means for finding a matching value by matching from, and Katakana selecting the extracted Katakana word correction candidate as a correct answer candidate when the matching value is higher than a preset value. A katakana word error detection / correction device comprising: a word candidate selection means. 2. A Katakana character string extracting means for extracting a Katakana character string from an input Japanese sentence character string, converting the extracted Katakana character string into Roman characters, and further converting it into a consonant string and a vowel string. Katakana character string consonant vowel separation means to be separated, and for each of the above extracted mutual katakana character strings, the concatenation string of the katakana character string and the symbol string separated into the vowel string are collated separately from the front and back, and the collation value If the mutual character strings to be collated are not the same and the collation value is higher than a preset value, there is an inconsistency in the katakana word notation between the extracted katakana character strings. The katakana word error detection and correction device according to claim 1, further comprising: a detecting unit.