JPH064716A - Character recognizing device - Google Patents

Abstract

PURPOSE:To improve the character recognition rate by extracting keywords and correcting a character based on an evaluation reference matched with the contents of a document. CONSTITUTION:An inputted document picture is processed once by a character recognizing part 1, a word retrieving part 3, a caluse retrieving part 6, a clause evaluation value calculating part 9, and a clause selecting part 10 to output a corrected character string 11, and keywords are extracted from the corrected character string 11 by a keyword extracting part 12 and used to restore characters other than candidates by a keyword partial coincidence retrieval part 14 and a candidate word adding part 15, and a weight is added to the evaluation value by a keyword weight adding part 16, and the corrected character string 11 is updated to recognize inputted characters.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device for reading and recognizing characters in sentences such as documents.

[0002]

2. Description of the Related Art In recent years, with the development of databases, there has been an increasing demand for a character recognition device having a high speed and a high recognition rate.

A conventional character recognition device is shown, for example, in IPSJ Journal Vol.30 No.11 pp.1394-1401. FIG. 4 shows a conventional character recognition device.
The character correction unit of the character recognition device includes a word search unit 3, a word dictionary 4, a phrase search unit 6, a grammar dictionary 7, a phrase evaluation value calculation unit 9, and the like, and the character correction unit includes the character recognition units 1 to 1. Receives n candidate characters per character as input.
The word search unit 3 of the character correction unit obtains a set of candidate words from the set of candidate characters by using the word dictionary 4. The phrase searching unit 6 uses the grammar dictionary 7 from the set of candidate words to select candidate phrases and obtain a set of candidate phrases. The phrase evaluation value calculation unit 9 evaluates the evaluation value of the character recognition unit 1, the frequency of words, the length of characters, and the like for each candidate phrase, and derives a phrase evaluation value indicating the likelihood of the phrase. The phrase selecting unit 10 selects the most likely phrase based on the phrase evaluation value and obtains the modified character string 11.

As described above, by using the word dictionary and the grammar dictionary, it is possible to correct a character that is difficult to determine by the character recognition unit 1 only, based on the knowledge of the word and the grammar.

[0005]

However, in the above-mentioned character recognition device, the knowledge utilized for correcting the recognized characters output from the character recognition unit 1 is the knowledge of words and grammar for general sentences. is there. However, many documents have their own characteristics. For example, in a patent document, many words unique to the patent are described. In this way, the style of the sentence and the frequency of the words used are usually different depending on the content of the document, but such unique information held by the document was not used to correct the recognition character. .

Further, in the conventional method, the correct character is selected from the n candidate characters output from the character recognition unit 1 for correction, but if the correct character is not in the n candidates, the correction is impossible. there were.

Therefore, there is a problem that the character recognition rate becomes low because the unique information of the document is not used as described above.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and an object of the present invention is to provide a character recognition apparatus capable of increasing the character recognition rate.

[0009]

According to the present invention of claim 1, a candidate character recognition means for reading a character string of a recognition target sentence to obtain a candidate character group, and a candidate character string created from a word dictionary and a candidate character group. A word phrase search means for obtaining a candidate word group by using the candidate word group and a grammar dictionary, and an evaluation value of at least the lexical and grammatical correctness of at least the candidate phrase group for each phrase. An evaluation value calculating means for calculating, a clause selecting means for selecting a clause from a candidate clause group according to the calculation result, and outputting a selection character string created by the selected clause, and the outputted selection A keyword extracting means for extracting a keyword from a character string based on a predetermined criterion, and a keyword weighting means for weighting an evaluation value with respect to a clause including the keyword are provided. Which recognizes character recognition apparatus a character string read by using the weighting.

According to the present invention of claim 2, a candidate character recognizing means for reading a character string of a recognition target sentence to obtain a candidate character group,
A word phrase search means for obtaining a candidate word group by a candidate character string created from a word dictionary and a candidate character group, and obtaining a candidate phrase group by the candidate word group and a grammar dictionary, and at least lexical and at least the candidate phrase group. Evaluation value calculation means for calculating an evaluation value for each phrase in consideration of grammatical correctness, and a selection character string created by selecting a phrase from a candidate phrase group according to the calculation result and selecting the phrase And a keyword extracting means for extracting a keyword from the output selected character string based on a predetermined criterion, and a partial match searching means for performing a partial match search between the candidate character group and the keyword. A candidate word adding means for adding the partially matched keyword as a candidate word to the candidate word group, and using the candidate word group to recognize the read character string. A character recognition device that.

[0011]

According to the present invention of claim 1, the candidate character recognition means reads the character string of the recognition target sentence to obtain a candidate character group, and the word / phrase search means is created from the word dictionary and the candidate character group. A candidate word group is obtained by using the candidate word group and the candidate phrase group is obtained by the grammar dictionary, and the evaluation value calculating means obtains the evaluation value of at least the lexical and grammatical correctness of the candidate phrase group. Each time, the bunsetsu selecting means selects a bunsetsu from the candidate bunsetsu group according to the operation result, outputs a selection character string created by the selected bunsetsu, and the keyword extracting means outputs the selected selection. The keyword is extracted from the character string based on a predetermined criterion, and the keyword weighting means weights the evaluation value of the phrase containing the keyword, and reads it using the weighting. Performs recognition of string.

According to the present invention of claim 2, the candidate character recognition means reads the character string of the recognition target sentence to obtain a candidate character group,
The word phrase search means obtains a candidate word group from the candidate character string created from the word dictionary and the candidate character group, obtains the candidate phrase group from the candidate word group and the grammar dictionary, and the evaluation value calculation means at least the candidate phrase group. Of at least the lexical and grammatical correctness, the evaluation value is calculated for each bunsetsu, and the bunsetsu selecting means selects a bunsetsu from the candidate bunsetsu group according to the operation result, and creates the selected bunsetsu. The selected character string is output, the keyword extraction means extracts a keyword from the output selected character string based on a predetermined criterion, and the partial match search means performs a partial match search between the candidate character group and the keyword. Then, the candidate word adding means adds the partially matched keyword as a candidate word to the candidate word group, and recognizes the read character string using the candidate word group.

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings showing its embodiments.

FIG. 1 is a block diagram of a character recognition apparatus according to an embodiment of the present invention. That is, the character recognition unit 1 performs character recognition from a character image, and a candidate character set 2 (n) having n candidate characters from the first candidate character to the nth candidate character for each character in a character string of length m ( Output the candidate character group).

The word search unit 3 searches the word dictionary 4 to select a candidate word that matches the word existing in the word dictionary 4 from the combination (candidate character string) of the candidate character set 2 to select the candidate word set. 5 (candidate word group) is created. The phrase searching unit 6 refers to the grammar dictionary 7 to select a candidate phrase that is a combination of words that can be a phrase, and a candidate phrase set 8
Create (candidate phrase group). The bunsetsu evaluation value calculation unit 9 calculates an evaluation value (Equation 1) based on the lexical and grammatical correctness of the bunsetsu searched by the bunsetsu searching unit 6 based on the length and frequency of the words in the bunsetsu.

[Equation 1] E = aE _a + bE _b + cE _c + dE _d where E _a : output of the character recognition device 1 E _b : length of phrase E _c : frequency of occurrence of independent word contained in phrase E _d : of word For the lengths a, b, c and d, the constant phrase selection unit 10 selects a phrase having a large evaluation value among the phrase candidates and generates a modified character string 11 (selected character string). The evaluation value of the phrase in the modified character string 11 is E _i (i = 1,
B). However, B is the number of clauses in the entire document.

The keyword extraction unit 12 is a phrase selection unit 10
The keyword of the document to be recognized is extracted from the corrected character string 11 output by the above, and the keyword set 13 is created. For this, the evaluation value of the clause in the modified character string and general frequency information of the word are used. For example, ease K _w becomes to keywords for word w can be determined by equation (2).
That is, the difference between the evaluation value and the general frequency of the word w is calculated for all the phrases including the word w in the document, and the sum of the difference and the larger value of 0 is calculated.

[0017]

[Equation 2]

However, S _i represents a character string included in the phrase i, and F _w represents a general frequency of the word w. Ease of becoming a keyword that can be calculated by (Equation 2) K _w
Are extracted in order from the largest one. Table 1 shows the result of extracting the keywords from the document related to the neural network according to the present embodiment.

[0019]

[Table 1]

Table 1 shows words arranged in descending order of the ease of becoming a keyword K _w . As can be seen from this table, the keywords contained in the document can be extracted. Further, when only the frequency information of words is used, frequent words such as “koto” and “mono” are extracted. However, since extraction is performed by obtaining the easiness of becoming a keyword K _w , such “ It can be seen that frequent words such as “thing” and “thing” are not extracted as keywords, but only words that represent the content of the document are extracted as keywords, and the obtained keywords well represent the characteristics of the sentence.

The keyword partial match search unit 14 performs a partial match search between the obtained keyword and the candidate character string set. For example, if “recognition” is extracted as a keyword, “recognition *” and “* knowledge” in the correction character string are extracted as partial matching characters.

The candidate word adding section 15 adds the partially matched keywords to the candidate word set 5 as candidate words of the candidate character string. In the above-mentioned example, the partially matching “recognition *” and “* knowledge” are added to the candidate word set 5 as “recognition”. As a result, a character that has not been output from the character recognition unit 1 can be included in the candidate characters. Since this uses a keyword indicating the feature of the sentence, it is possible to perform processing according to the feature of the sentence.

Further, the keyword weighting unit 16 increases the evaluation value of the candidate phrase according to (Equation 3) if there is a keyword in the phrase. This enhances the evaluation of candidate clauses using keywords that represent the characteristics of the sentence. This makes it possible to make corrections using the characteristics of the text.

[0024]

[Equation 3]

The above character recognition unit 1 constitutes the candidate character recognition means, the word search unit 3 and the phrase search unit 6 constitute the word phrase search unit, and the phrase evaluation value calculation unit 9 constitutes the evaluation value calculation unit. The phrase selection unit 10 constitutes a phrase selection unit, the keyword extraction unit 12 constitutes a keyword extraction unit, the keyword weight addition unit 16 constitutes a keyword weighting unit, and the keyword partial match search unit 14 forms a partial match search unit. And the candidate word adding unit 15 constitutes a candidate word adding means. The part excluding the character recognition unit 1 described above constitutes a character correction unit.

Next, the operation of the above embodiment will be described.

First, the document to be recognized is processed by the character recognition unit 1 to obtain a candidate character set 2 of, for example, the 10th place. Next, the word search unit 3 selects a combination of candidate characters that match a word existing in the word dictionary 4, and the phrase search unit 6
Then, referring to the grammar dictionary 7, a combination of words that can become a clause is selected. Then, the phrase evaluation value calculation unit 9 calculates the evaluation value of the phrase searched by the phrase search unit 6. The phrase selection unit 10 selects a correct phrase from the candidate phrases based on the calculated evaluation value, and outputs the corrected character string 11 of the first pass.

Next, the keyword extracting unit 12 calculates the easiness of becoming a keyword K _w and extracts the keyword.

The keyword partial match search unit 14 performs a partial match search between the obtained keyword and the candidate character set 2. In order to reduce the amount of calculation, the candidate character set 2 may include only characters higher than the character recognition unit 1, for example, only the first character. This is not a problem as long as the character recognition unit 1 has a high recognition rate.

Next, the candidate word addition unit 15 sets the keywords obtained by partial matching as candidate words and sets the candidate word set 5.
Added to. Evaluation value E _a of the character recognition of the corrected character
Is suitable to give the lowest evaluation value here.

It is preferable to use the tenth place as the keyword to be used. However, if the text is long, it is better to use more keywords.

Again, the phrase retrieval unit 6 and the phrase evaluation value calculation unit 9
Then, the candidate phrase is selected using the candidate word set 5 to which the keyword is added as the candidate word, and the phrase evaluation value of the newly selected candidate phrase is obtained.

Next, the keyword weight adding unit 16
Correct the clause evaluation value of (Equation 3) according to the characteristics of the sentence.

Finally, the phrase selecting unit 10 selects a phrase having a large evaluation value among the phrase candidates and outputs the final modified character string 11 of the second pass.

Part of the result of character recognition according to this embodiment is shown in FIGS.

FIG. 2 shows the effect of the keyword weight adding section 16. In FIG. 2, seven clauses were selected as candidates for the character string “on-site learning function”. In the 1-pass recognition result, the phrase "learning" was not selected, but since the word "learning" is included in the keyword information, the evaluation value of the phrase "learning" increases in the 2-pass recognition result, and the character Could be fixed. In this way, the phrase is selected from the lexically and grammatically correct ones by using the evaluation values in both the 1-pass recognition and the 2-pass recognition, and the keyword information is used in the 2-pass recognition to further match the character string to the content of the document. Was able to choose.

FIG. 3 shows the result of applying this embodiment to a sentence showing the content of "post-processing effect of character recognition". Using this result, the keyword partial match search unit 14 and the candidate word addition unit 15 are used. Show the effect. In the output from the character recognition unit 1, the second “recognition” knowledge is not included in the candidate characters. Therefore, the word cannot be corrected by the conventional method. In this method, as shown in FIG. 3, the keyword based on the feature of the sentence can be extracted, so that “recognition” can be included in the candidates by using the keyword partial match search. Therefore, the correct sentence could be obtained after the correction of the second pass.

In the above-described embodiment, the keyword is extracted from the character string once processed by the character correction unit, and the two-pass recognition in which the character correction unit processes again using the extracted keyword information is performed. A keyword may be extracted using the result of the recognition, and the n-pass recognition may be performed again by repeating the process once or more in the character correction unit. in this case,
Further, correct keyword information is obtained and the recognition rate is improved.

Further, in the above embodiment, for the character extracted from the outside of the candidate (that is, the character obtained by adding the keyword as the candidate word to the candidate word set), the evaluation value E _a of the character recognition unit 1 Gives the lowest value, but sends the extracted character to the character recognition unit 1 and evaluates the character E _a for this character.
May be recalculated to obtain the evaluation value E _a . by this,
Furthermore, accurate evaluation becomes possible.

Further, in the above embodiment, the keyword is extracted by using the evaluation value of the phrase and the general frequency information of the word. However, the present invention is not limited to this, and the candidate character string which can be a candidate word is recognized in the recognition target sentence. The keyword may be extracted using the number of appearances and the appearance frequency dictionary of the candidate word in a general sentence. Alternatively, other extraction methods may be used.

In the above embodiment, the character is recognized by using both the partial keyword matching search and the weight addition. However, the present invention is not limited to this, and only one of the keyword partial matching search and the weight addition is performed. It may be configured to recognize a character by using it.

[0042]

As is apparent from the above description, the present invention has an advantage that the character recognition rate can be increased.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a character recognition device according to an embodiment of the present invention.

FIG. 2 is an output diagram of an experimental result showing an effect of keyword weight addition of the embodiment.

FIG. 3 is an output diagram of experimental results showing the effects of the keyword partial match search and the candidate word addition unit of the embodiment.

FIG. 4 is a configuration diagram of a conventional character recognition device.

[Explanation of symbols]

 1 character recognition unit 3 word search unit 4 word dictionary 6 phrase search unit 7 grammar dictionary 9 phrase evaluation value calculation unit 10 phrase selection unit 12 keyword extraction unit 14 keyword partial match search unit 15 candidate word addition unit 16 keyword weight addition unit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Taiji Yuki, 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd.

Claims (3)

[Claims] 1. A candidate word group is obtained by a candidate character recognition means for reading a character string of a recognition target sentence to obtain a candidate character group, and a candidate character string created from the word dictionary and the candidate character group. And a word phrase search means for obtaining a candidate phrase group from a grammar dictionary, an evaluation value calculation means for calculating, for each phrase, an evaluation value of at least the lexical and grammatical correctness of the candidate phrase group, and the calculation thereof. A phrase selecting unit that selects a phrase from the candidate phrase group according to the result and outputs a selection character string created by the selected phrase, and a keyword based on a predetermined criterion from the output selection character string And a keyword weighting unit that weights the evaluation value for a phrase including the keyword,
A character recognition device characterized by recognizing the read character string using the weighting. 2. A candidate word group is obtained by a candidate character recognition means for reading a character string of a recognition target sentence to obtain a candidate character group, a word dictionary and a candidate character string created from the candidate character group, and the candidate word group. And a word phrase search means for obtaining a candidate phrase group from a grammar dictionary, an evaluation value calculation means for calculating, for each phrase, an evaluation value of at least the lexical and grammatical correctness of the candidate phrase group, and the calculation thereof. A phrase selecting unit that selects a phrase from the candidate phrase group according to the result and outputs a selection character string created by the selected phrase, and a keyword based on a predetermined criterion from the output selection character string A keyword extraction means for extracting a partial match search means for performing a partial match search between the candidate character group and the keyword, and the partially matched keyword as a candidate word And a candidate word adding unit that adds the candidate word group to the candidate word group, and recognizes the read character string using the candidate word group. 3. The character according to claim 1, wherein the predetermined criterion is the number of appearances of the candidate word in the recognition target sentence and the appearance frequency of the candidate word in a general sentence. Recognition device.