JPH07110844A - Japanese document processor - Google Patents

Abstract

(57) [Abstract] [Purpose] The present invention provides a Japanese recognition input device for characters and voices, a Japanese input device from a keyboard, and the like.
A method of pointing out the possibility of an input error by checking the connection frequency of three consecutive target characters in a sentence input from a recognition device or a keyboard using the "three-character connection table for target characters" Is. Furthermore, especially in the recognition device, there are recognition candidate characters in the prior art. [Structure] The candidate character of the first character of the recognition result is replaced with the target character for inspection, and the candidate character is more frequently connected than the character of the first character of the recognition result, and a certain condition is satisfied. In this case, it is automatically corrected to a candidate character with a high frequency, assuming that the first character of the original recognition result was wrong. As the number of target characters increases, harmful effects will appear. Therefore, in the present invention, symbols frequently used in hiragana, katakana, and Japanese are set as the target characters.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a Japanese voice, a character recognition device, and a processing device at the time of input of a word processor.

[0002]

2. Description of the Related Art Conventionally, as a post-processing method for a Japanese input device, particularly a recognition device, a method based on a collation check with a word dictionary and a grammatical connection check between words has been used (Umeda: "Word Dictionary". Determining Power of Characters in Character Recognition Using "The Institute of Electronics, Information and Communication Engineers, Vol. J72-D-II N
o. 1 pp. 22-31 January 1989, and Ikehara et al .: "Automatic detection of Japanese typographical errors by word analysis program and extraction of correction candidates by secondary Markov model" IPSJ Journal Vol. 25 No. 2 pp. 298
-305 March 1984).

There is also a method of utilizing a two-character concatenation probability (Sugimura et al .: "Unreadable character determination processing using character concatenation information", IEICE Transactions Vol. J68-D.
No. 1 pp. 64-71 January 1985).

These methods can cover the processing of Japanese sentences including all character types such as kanji, hiragana and katakana, and the post-processing correction effect is high.

[0005]

In the above method, a huge word dictionary used for analysis is required, processing is complicated, processing takes time, and much time is spent for processing Kanji that is not easily mistaken. Therefore, there is a problem of post-processing when using morphological analysis, such as a lot of waste of processing.

[0006]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and is a Japanese document processing apparatus for pointing out and correcting an input error in a Japanese document. A three-character connection frequency table of target characters created by statistical processing, a means for extracting a sequence of Japanese target characters including the input Kanji, and an incorrect input of the sequence of extracted characters by the three-character connection frequency table. From the result of the check, and a means for pointing out to the user the spelling of a character sequence with an extremely low frequency from the result of the check (EN) A Japanese document processing device comprising a means for correcting an arrangement of characters having an extremely low frequency and a sequence of characters having a high frequency.

[0007]

[Operation] Kanji errors are often seen at first glance, which is an unavoidable problem when inputting. Kana input errors are difficult to find in the proofreading stage. Since the sequence of kana has a small amount of features, it is more likely to be misrecognized than the sequence of kanji, so it is more reliable to use the probability of three-letter sequences than to use the probability of two letters. Then, it is possible to reduce the number of input errors as a whole by pointing out and correcting.

[0008]

EXAMPLES Examples of the present invention when realized by OCR will be described below with reference to the tables, but the present invention is not limited to these.

(Creation of target character three-character connection frequency table) First, a target character three-character connection frequency table is created. However, here, a table is created with hiragana characters (83 characters), katakana characters (86 characters), and symbols as target characters. Enter Japanese text data,
Hiragana characters (83 characters) Katakana characters (86 characters),
For and symbols, the character strings are extracted and the frequency of existence of the three character strings is obtained. Here, the character for which the frequency is to be obtained (the hiragana character, katakana character, and symbol) is called the target character. The three-character string is a character string in which the target character is continuously three characters. Each character is ordered by the target character code table (Table 1).

[0010]

[Table 1]

For example, when the target character code table is as shown in Table 1, the position number of "a" is "0", the position number of "a" is "1", and the position number of "i" is "2". The position numbers of "" and "I" are "3", and the numbers are determined according to the order of arrangement. With the target code table decided, a large amount of Japanese text data is input and processed. FIG. 1 shows a flowchart of an example of the processing for creating the connection frequency table for the three target characters. The process proceeds by reading Japanese text data character by character. For example, assume that the data shown in Table 2 below is input.

[0012]

[Table 2]

It is checked whether the three-letter spelling is established. (Step 1) If the three-character spelling is established, 1 is added to the table at the position described in the three-character buffer, and if the three-character spelling is not established, the process proceeds to step 2. (Step 2) Read one character. If it is the end of the file, the process ends. (Step 3) It is checked whether the read character is the target character. (Step 4) In step 4, if the result of the check is the target character, it is shifted to the left position of the three-character spelling buffer and the character position is entered at the end of the three-character spelling buffer. (Step 5) If the check result in step 4 is not the target character, the three-character buffer is cleared and the process returns to step 1. (Step 6) For example, in this process, six types of character strings shown in the following Table 3 are extracted with respect to the first data input “same as,” in Table 2.

[0014]

[Table 3]

From Table 1, the position of "ji" is 23.
Since the position of "mo" is 65 and the position of "no" is 45,
Every time the character string “one time” appears, 1 is added to the tables 123, 65, 45).

When all the Japanese text data is processed, the number of times the target character string exists in the processed Japanese text data is recorded in the table. A value obtained by dividing the number of times of existence of these by the total number of times is defined as a “target character three-character connection frequency table” (FIG. 2).

However, this table has a size of the number of target characters × the number of target characters × the number of target characters, and the size is too large to be practical as it is. Therefore, only the target character strings existing by the table reduction program described below are stored. This makes the table size 3
It is reduced to less than K bytes.

(Method for reducing target character three-character connection table)
Basically, this reduction method is to reduce the created “target character three-character connection frequency table” with a frequency of 0 (flowcharts FIGS. 3 and 4).

A table (search table 1 tabcount [MOJISUU]) having a size corresponding to the target character is prepared. (Step 1) The first character of the three-character connection frequency table is fixed as the grouping key, and the first character of the three-character connection frequency table is fixed to check how many elements exist in the same key. When the frequency of the connection frequency table is not 0, 1 is added to the search table. (Step 2) The search table is written in the compressed table file (step 3). Check the 3 character connection frequency table for the number of times of the elements of the table, and when the content of the table is not 0, the position of the second character of the table, the position of the third character, and the numerical value of the frequency table corresponding to that position are displayed. Write to compressed table. (Step 4) When reading the compressed table file, the reverse operation may be performed.

(Input error indication / correction processing) The flow of input error indication / correction processing performed by using the "target character three-character connection frequency table compression table" created by the above method is shown below (FIG. 5). 6 and 7 are flowcharts and FIG. 8 is a block diagram).

An input device such as an OCR system performs input in a text data format. This process refers to a step of extracting a three-character spelling of a target character string such as kana (target character three-character spelling extraction step) from the entire input Japanese character string and the frequency of occurrence of the three-character spelling from the frequency table. Step (table reference step) and step of checking existence frequency of three-letter spelling (check step)
When the existence frequency is below a certain threshold, the three-letter spelling is registered as a character string that may cause an input error (input error candidate extraction step), and the position of the input error candidate is notified to the user (input error correction). Result output step).

The target character three-character spelling extraction step is O
The Japanese character spelling output from the CR recognition unit is read character by character. For example, in the target character spelling extraction step,

[0023]

[Table 4]

When the character string "is output from the recognition unit of the OCR, first," the same "is read into the buffer and it is checked whether it is the target character. The check is performed using the target character code table. If it is not the target character, nothing is done and the next character is read. Next, "ji"
Is read and similarly it is checked whether it is the target character. If it is the target character, it is copied to the three-character spelling buffer and the next character is read. Since "mo" is also the target character, it is copied to the place next to "ji" in the three-character spelling buffer. Similarly, "no" is also copied to the three-character spelling buffer. At this point, the three-character spelling of the target character is completed, and the target character spelling extraction step ends.

The table reference step obtains the frequency of three-letter spelling by referring to the "target character three-letter connection frequency table compression table" from the three-letter spelling sent from the target letter spelling extraction step. In the check step, the frequency obtained in the table reference step is low, and when the value exceeds a certain threshold value, a flag indicating a possibility of an error is added to the first "jimono" part as the 3-character spell check information. send.

Returning again from the target character spelling extraction step, "to" is read. Here, the three-letter spelling "monoto" is created and sent to the table reference step. Similar processing is performed until the read character is not the target character. If there is a flag with an error possibility, the position of the flag with an error possibility is checked in the input error candidate extraction step, and when the flag with an error possibility continues beyond the threshold, the three characters Characters in the overlapped part of the spelling convey information as a possible error. At the same time, the user is pointed out that there is a possibility of error by reversing the output display. Next, the candidate character is referenced, and the erroneous character that may be an error is replaced with the candidate character. Next, the frequency of the three-character spelling replaced with the candidate of the “target character three-character connection frequency table compression table” is referred to. At this time, the original frequency is compared with the frequency when the candidate is replaced, and if it exceeds the threshold, the process proceeds to the input error correction result output step.

In the case of notifying the user of the position of the input error candidate (input error correction result output step), the three-letter spells, which are frequently extracted in the input candidate extraction step, are output as the correction result.

An example of input and output of this processing will be shown below.

<Example of input error input / correction processing>

[0030]

[Table 5]

The first-order recognition result of this character string is the table target character, and a character string having the number of characters of 3 or more is sent to the input error indication / correction processing frequency table reference section. That is, in the upper row, it is a single character string "not a scale."

<Example of output result of correction processing indicating correction of input error>

[0033]

[Table 6]

“Ha” is modified to “ba”.

In the above embodiment, the Japanese OCR is used for explanation, but the present invention is not limited to this and may be used for a mistake check in the post-processing of an input device such as a voice recognition device or a word processor. .

In addition, it has a "target three-syllable connection frequency table" created by statistical processing from a sentence in which only the hiragana characters in which the kanji part of a general sentence is corrected into a kana, and the arrangement of the inputted kana characters is incorrect by the above table Equipped with a means for checking whether there is a possibility of inputting and a means for pointing out to the user a sequence of characters that has an extremely low frequency as a spelling that may cause an erroneous input. It may be used as a processing device for performing a mischeck on a character string of kana just before it is converted into the kanji.

[0037]

As described above, according to the present invention, there is a problem in the post-processing using the conventional technique of morphological analysis, that is, a huge word dictionary used for the analysis is required, the processing is complicated, and the processing time is long. It is possible to solve the problem that a large amount of time is spent on the processing of kanji that are not erroneously made, and the processing is wasteful.

[Brief description of drawings]

FIG. 1 is a flowchart showing a target character three-character connection frequency table creation process according to an embodiment of the present invention.

FIG. 2 is a configuration diagram of a connection frequency table of three target characters according to an embodiment of the present invention.

FIG. 3 is a flowchart illustrating a method of reducing a target character three character connection frequency table according to an embodiment of the present invention.

FIG. 4 is a flowchart showing a method of reducing a target character three character connection frequency table according to an embodiment of the present invention.

FIG. 5 is a flowchart showing an input error indication and correction process according to an embodiment of the present invention.

FIG. 6 is a flowchart showing input error indication and correction processing according to an embodiment of the present invention.

FIG. 7 is a flowchart showing an input error indication and correction process according to an embodiment of the present invention.

FIG. 8 is a block diagram showing a configuration according to an exemplary embodiment of the present invention.

Claims (1)

[Claims] 1. A Japanese document processing apparatus for pointing out and correcting an input error in a Japanese document, comprising a three-character connection frequency table of target characters created by statistical processing from general sentences, and input Kanji characters. Means for extracting a sequence of Japanese target characters including "," means for checking whether or not there is a possibility of erroneous input of the sequence of extracted characters by the three-character connection frequency table, and a frequency from the result of the check. Means for pointing out to the user the spelling of the extremely low sequence of characters as a spelling that may be erroneous input, and means for correcting the sequence of extremely low frequency characters from the result of the check as described above, A Japanese document processing device characterized by comprising.