JPH01276285A - System for collating word - Google Patents

Abstract

PURPOSE:To improve a correct answer rate by introducing a method to permit a retrieval only once to a recognition time position at one point. CONSTITUTION:A position history table 8 to record the retrieval history of the recognition time position for each word code is provided, and by inputting the recognition time positions of the word code and a character code to select the word code, whether an overlapping retrieval exists or not is tested through the use of a position testing part 7 to output the word code and an evaluation value to be unretrieved as to the recognition time position of the word code. Thus, at the time of checking the character in one or more in-dictionary positions as to one recognition position because plural candidate characters exist in one recognition time position and the word exists which is composed of the combination of the candidate characters, since the retrieval history is referred to by the position history table 8 from the word code and recognition time position at the position detecting part 7 even to the case of outputting the evaluation value my mistake.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a word matching method, and in particular, when there are multiple candidate characters for the character position of an input character string, even if there is an error in the input character, The present invention relates to a word recognition method for efficiently recognizing words.

[Conventional technology]

In conventional word recognition methods, when recognizing words from an input character string represented by a sequence of character codes, there are multiple candidate characters for each character position in the input character string, and there are errors in the input character string. Assuming that the input character string contains omitted characters or redundant characters, it may be necessary to match each candidate character at each position of the input character string with characters at one or more positions constituting an arbitrary word.

FIG. 6 is a block diagram showing an example of such a conventional word recognition device.

In FIG. 6, 1 is a word selection table.

2 is a selection processing unit, 3 is a judgment value table, and 4
is a judgment value accumulation processing section, 5 is a sorting processing section,
6 is an evaluation value table. for example.

When "Decotar" is entered as an input character string (assuming that the ko in this text was mistakenly entered as ko), by referring to the word selection table 1 of the 1 selection processing unit 2,
A word code and a judgment value are output for each of the first character, second character, etc., and these are transferred to the judgment value accumulation processing section 4 at the next stage. The judgment value accumulation processing section 4 determines an evaluation value for each input word code by referring to the judgment value table 3, and transfers it to the next stage sorting processing section 5. The sort processing unit 5 ranks and outputs candidate words using an evaluation value table. The one with the highest rank is the recognized character.

This word recognition device uses a word dictionary in which the arrangement of each character code corresponds to each word, and in order to compare the arrangement of character codes of the input string with each word, Enter the character codes that make up the input string in . In other words, as mentioned above, when you input "dekotal", digital Wll, data W12, and decoder W13 are output for the first character "de", and the second character "de" is output.
Decoder W13 and record W2B are output for coj. Thereafter, in the same way, a word code and a judgment value are output for each input character. In this manner, the selection processing unit 2 outputs a word code containing a character of the character code at each character position using the word selection table 1. In order to deal with typos, redundant characters, and omissions, this word recognition device uses character positions in the word code in the dictionary (hereinafter referred to as dictionary positions) and character positions in the input string (hereinafter referred to as recognition positions). ) do not match, the word selection table is searched at one or more positions in the dictionary for each candidate character at one recognition position.

Regarding the conventional word matching method, please refer to the specification of Japanese Patent Application No. 1983-248415 and the drawings (title of invention: word recognition device), for example.

[Problem to be solved by the invention]

In such a word recognition device, there is one or more candidate characters at one recognition position, and the word selection table is searched for one or more positions in the dictionary for that candidate character. An erroneous evaluation value as described below will be output. Therefore, such erroneous output should be corrected!ii! A problem arises.

FIG. 7 is a diagram showing a specific processing process by a conventional word recognition device.

In character recognition, as shown in FIG. 7, one or more character candidates are output at one recognition position. In other words, if the word dictionary contains Tokyo, Kyoto, and Tokyo, and you input "Kyoto" as the input character string, the three characters Higashi, Kyoto, and Toru will be selected as candidate characters in position cl (gear lag 91st character) at the time of recognition. , C2, the three characters Miyako, Gun, and Gun are output as candidate characters, respectively. In other words, there are cases where candidate characters within the same recognition position can form a word ([i'Tokyo]).

In this case, if the above method is used, the selection processing unit 2 in FIG.
, when the first character at recognition position c is considered as the first character in the dictionary, the candidate character "Tokyo" is extracted from "Higashi J" in ■, and when cl is considered as the second character in the dictionary, the candidate character is in ■ ``Tokyo'' is removed from Kyoto J. At this time, 1
Even though only one candidate character can be used at the recognition position of a place, when searching for "Tokyo", two candidate characters are used at one recognition position, giving an incorrect evaluation value. It turns out. In other words, in the judgment value accumulation section, "
Tokyo J is 1.7, ``Kyoto'' is 1.6, and ``City Center'' is 0.8.
In the next sort processing section, "Tokyo" is evaluated as 1.
7. ``Kyoto J'' is 1.6, and ``Tokyo Center'' is 0.8. Therefore, the conventional method does not perform accurate processing.

Due to this, in the example of FIG. 7, "Kyoto", which should originally be extracted as the first place, is output as the second place.

An object of the present invention is to solve such conventional problems, and to avoid outputting evaluation values by mistake when matching characters at one or more dictionary positions for one recognition position, and to avoid inputting To provide a word matching method capable of recognizing words with high precision even if there are misspellings, omissions, or extra characters in characters, and also capable of recognizing words with high precision even when parallel processing is performed. It is in.

[Means to solve the problem]

In order to achieve the above object, the word matching method of the present invention is a word matching method in which an input character string in which one-character codes are arranged corresponds to which word among preset words. It includes a word selection table that associates the characters used with words for each position of the constituent character array (hereinafter referred to as the dictionary position), and contains a word selection table that matches the characters used and words, and ) candidate character character code 1
- and the dictionary's own position as input, and a selection processing means that outputs the corresponding word code, and a position history table that records the position at the time of recognition of the character code used to select the word code for each word code, position verification means that receives as input the code and the position at the time of recognition of the character code used to select the word code, and outputs a search history indicating whether or not the position at the time of recognition of the word code has been searched with other candidate characters;
For each word code output from the position verification means,
It has a judgment value accumulating means for adding an evaluation value indicating certainty, and a sorting means for extracting candidate words based on the magnitude relationship of the judgment values of the judgment value accumulating means, and each recognition of the input character string is performed by the selection processing means. Regarding the candidate characters for each position, a word code corresponding to the character code is selected from the word selection table for each one or more dictionary positions, and the search history of the word code and the position at the time of recognition is checked by the position verification means. Read from the position history table, set the search history to the position history table only if it is not yet searched,
The word code is sent to the judgment value accumulation means, and the judgment value accumulation means cumulatively adds the evaluation value of the word code output from the position verification means to obtain a judgment value for the input character string, and performs the sorting. The feature is that candidate words are extracted by means. Another feature is that a measure of certainty obtained from the recognition results of each character in the input character string is written as the search history in the position history table.

[For production]

In the present invention, the accuracy rate is improved by introducing a method that allows only one search for one recognition position. That is, a position history table that holds the score and position at the time of recognition for each candidate character for each position at the time of recognition of the input character string, and records the position at the time of recognition of the character code used to select the word code for each word code. , enter the word code and the position at the time of recognition of the candidate character used to select the word coat, and for that word code, check the word code whose position at the time of recognition has not been searched and the evaluation value of that word code. The output position verification unit determines whether the word code obtained by the selection processing unit and the recognition position of the candidate character for which the word code is selected has already been selected from the same recognition position. Only when a selection is made for the first time, the word code and evaluation value are output to the judgment value accumulator, and at the same time, the status of verification is recorded in the position verification table.

As a result, there is one or more candidate character recognition at each recognition position of the input string, and there are typos, omissions, and errors in the input string.
Even if redundant characters are included, the position history table can be used to prevent incorrect evaluation of word codes that occur when matching one or more candidate characters at one recognition position with one or more characters at one or more dictionary positions. By using and avoiding this method, highly accurate word recognition becomes possible.

〔Example〕

EMBODIMENT OF THE INVENTION Below, one embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram of a word matching device showing an embodiment of the present invention.

The word matching device of the present invention includes a selection processing section 2 including a word selection table 1, and a position verification section 7 including a position history table 8.
, a judgment value accumulation section 4 including an evaluation value table 3, and a sorting processing section 5 including a candidate word table (not shown).

FIG. 2 is a process diagram showing an embodiment of the word matching method of the present invention.

The operating principle of the present invention will now be described for a case where the same character recognition results are obtained for the same input character string as in FIG. 7.

The selection processing unit 2 performs the following selection processing for each position at the time of recognition. First, the word selection table 1 is searched using the character code of each candidate character at the position at the time of recognition, and the word code corresponding to the character code is read out. This word selection table 1 shows a correspondence table between character codes for each dictionary position and word codes for which the character code exists at this dictionary position.

The word recognition device selects a word when the recognition position 1001 is the first character in the dictionary, and when the second character position is ■. In ■, the candidate characters 'Higashi' to 'Tokyo' and the candidate characters 'Kyo' to 'Kyoto' are read out, respectively, and 'Tokyo', the evaluation value 0.9, and the position at the time of recognition c1
, ``Kyoto'', evaluation value 0.8., and cl are output in combination. Also, in ■, ``Tokyo'' is read from the candidate character ``Kyo'', and ``Tokyo'', 0.8, and The position at the time of recognition c1 is combined and output. Also, the position at the time of recognition c
Regarding 2, according to the word selection table 1 whose dictionary position is the first character and the second character, 'inside', 0.8 and c2,
Also 'Kyoto', 0.8. and c2 are combined and output. By sequentially processing in this way, the selected word code, the position at the time of recognition, and the evaluation value are combined and transferred to the position verification section 7 at the next stage.

In the position verification section 7, based on the word code sent from the word selection section 2, the search history regarding the position at the time of recognition of the word code is retrieved from the position history table 8, and the search history regarding the position at the time of recognition sent from the word selection table 1 is retrieved. Test whether the search for has been performed before. The position history table 8 shown in FIG. 2 is an example in which the position at the time of recognition is up to the second character, and for each word code, the processed (1cl, c2) position at the time of recognition is
) or unprocessed (0). As an example of the position verification process using this position history table 8, the position history table 8 is as shown in FIG.
The search history obtained by testing the position of ■ is being recorded, and next, in ■, the result with the word code 'Tokyo', the evaluation value 0.8, and the position at the time of recognition is cl was sent. Let me explain the case. From the position history table 8, it is determined that the word Tokyo is duplicated from the word code ``Tokyo'', and the recognition position C1 is processed (1) and the recognition position [c2 is unprocessed (0). You can get that. From this, in the word code 'Tokyo', it can be detected that the recognition position C1 sent from the selection processing unit 2 has been processed, so the word code 1 Tokyo' is
Not sent to. In addition, in the example of position testing for ■, for the word code 'Kyoto', it is detected that the recognition position c2 sent from the selection processing unit 2 is unprocessed, and the word code 'Kyoto' has an evaluation value. It is also sent to the judgment value accumulator 4. At this time, 1Kyoto' and c2 of position history table 8
Set searched (1) in the field specified by .

The judgment value accumulator 4 evaluates the same word code among the word codes output by the 1-position verification unit 7.

Add evaluation values cumulatively. In the above example, the word code 'Tokyo' is only sent in the case of ■ in Figure 2, so
The evaluation value is 0.9, which means that a correct value has been given as the evaluation value of the word obtained from the recognition candidate characters. After that, the sorting processing unit 5 compares the evaluation values. In the example of FIG. 2, this word recognition device outputs 'Kyoto' as the first candidate.

FIG. 3 is a diagram of a position history table at the time of expansion according to the present invention.

That is, FIG. 3 shows a position history table when the recognition position is extended to the Nth character. The position history table 8 is composed of columns for recording whether the word code W and each arbitrary recognition position C are processed (1) or unprocessed (0).

When searching the position history table 8, 110 can be obtained using the word code and the position at the time of recognition currently being processed. For this purpose, when the word code Wk is selected in the 1 selection processing section 2 by inputting the character code CDj of the candidate character in the recognition position [Ci, W in the position history table 8
The value in row k and column Ci is read out, and it is tested whether or not the word code is to be output to the judgment value accumulator 4. If the result of the test is that it has not been processed, record searched (1) in the Wk row and Ci column, confirm that all recognition positions C1 to Cn have been processed, and add the word code and evaluation. The value is sent to the judgment value/accumulator 4. Furthermore, if the input word code has already been processed from the table, that word code is not output to the judgment value accumulation unit 4.

The configuration shown in FIG. 1 will be explained in further detail.

In the configuration shown in FIG. 1, a set of candidate characters, recognition positions, and evaluation values at each recognition position of an input character string is input, and the most likely word is output. For each candidate character code at each recognition position, the search circuit 11 uses the word selection table 1 in memory to search for a word code corresponding to the character code for one or more positions in the dictionary, and then searches for the word code corresponding to the character code for one or more positions in the dictionary. It is sent to the determination circuit 12. The determination circuit 12 refers to the position history table 8 in the memory and verifies whether the word code has been searched repeatedly. If no duplicate searches have been made, the search is recorded in the position history table 8, and the word code and evaluation value are sent to the addition circuit 13. The addition circuit 13 detects the current judgment value of the word code from the evaluation value table 3 in the memory, updates the word code judgment value by adding the evaluation value, and returns the value to the evaluation value table 3. . all! ! After the processing using all the dictionary self-positions of the recognition time position is completed, the sorting circuit 9 sorts the judgment value table 6 (not shown) and outputs the result. That is, the word code and the judgment value are output.

FIG. 4 is a block diagram of a word matching device showing another embodiment of the present invention, and FIG. 5 is a diagram of a position history table used in FIG. 4.

FIG. 4 shows an example of the configuration for parallel verification.

In FIG. 4, 1 selection processing/position verification units 10a to Io
n search circuits 11a to lln and determination circuits 12a to 1
2n in parallel for each position in the dictionary to create position history tables 8a to 8n for each position in the dictionary.

Thereafter, an adder circuit 13 calculates a judgment value, a sorting circuit 9 compares the judgment values, and outputs the result. At this time, 1 position history table 8 and 8a to 8n
As shown in FIGS. 5(a), 5(b), and 5(c), instead of the flag 110, an evaluation value is recorded in the column of each recognition position of the word code.

After the processing for each position in the dictionary is completed, the comparison circuit 14 inspects each position history table 8a to 8n, and if there are duplicates, the evaluation values are compared and the higher evaluation values are prioritized. Avoid false evaluations due to duplicate searches. This makes it possible to match candidate characters with high scores during recognition.

In Fig. 5, the word code W of the one-sentence lag in the dictionary is
The first character at the recognition position of No. 1 overlaps with the first character at the recognition position of the M-th word code W1 in the dictionary. Therefore, by comparing the two, the highest score of 0.8 is set as the score of one sentence lag in the dictionary position in the w1 row and 1st column.

Note that matching a character with one or more dictionary positions for one recognition position is a problem that occurs during recognition because the punctuation of characters or syllables is unknown, such as handwritten characters with arbitrary pitches or continuous speech. It is also necessary when correcting the discrepancy between the position and the dictionary's own position.

In this way, in the present invention, a position history table is provided in which the search history of the position at the time of recognition is recorded for each word code, and by inputting the word code and the H position of the character code from which the word code is selected, It is tested whether or not there is a duplicate search using a position verification unit that outputs unsearched word codes and evaluation values regarding the position at the time of recognition of the word code. As a result, there are multiple candidate characters at one recognition position, and some words can be formed by combining these candidate characters, so when matching characters at one or more dictionary positions for one recognition position, Furthermore, even if an evaluation value is output by mistake, the 1-position detection unit refers to the search history using the position history table based on the word code and the position at the time of recognition, so that erroneous evaluation can be prevented. In addition, even when performing parallel processing, by providing a position history table for each processor and using the value recorded in the position history table as the evaluation value of the character candidate at each recognition position, high accuracy can be achieved even with parallel processing. can recognize words.

〔Effect of the invention〕

As explained above, according to the present invention, even if there is a character candidate of one code or more at each character position of an input character string, and there are typos, omissions, or extra spaces in the input character string, high accuracy can be achieved. can recognize words. Further, even when parallel processing is performed, words can be recognized with high accuracy by using the values recorded in the position history table as evaluation values of character candidates.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a word matching device showing an embodiment of the present invention, Fig. 2 is a process diagram when performing recognition processing using Fig. 1, and Fig. 3 is a position history table in Fig. 1. 4 is a block diagram of a parallel processing word matching device showing another embodiment of the present invention, and FIG. 5 is a diagram showing an example of the configuration of a position history table used in the device of FIG. 4. , Figure 6 is a block diagram of a conventional word recognition device, and Figure 7 is a block diagram of a conventional word recognition device.
It is a figure which shows the operation process of the apparatus of figure. 1: word selection table, 2: selection processing section, 3: judgment value table, 4: judgment value accumulation section, 5: sorting processing section, 6: position history table, 7: position verification section. 8.8a to 8n: position history table, 9: sorting circuit,
10a to 10m: selection processing/position verification section, 11: search circuit, 12: determination circuit, 13: addition circuit, 14: comparison circuit. Patent applicant: Nippon Telegraph and Telephone Corporation Figure 1 High power, word code, judgment value Figure 3 0: Unprocessed, 1: Processed

Claims (2)

[Claims] (1) In a word matching method that matches an input character string in which character codes are arranged to which word among preset words, each position (
It includes a word selection table that matches characters used with words for each position in the dictionary (hereinafter referred to as the position in the dictionary), and the character code of the candidate character at any character position in the input string (hereinafter referred to as the position at the time of recognition). It includes a selection processing means that receives a position in the dictionary as an input and outputs a corresponding word code, and a position history table that records the position at the time of recognition of the character code used for selecting the word code for each word code. a position verification means that receives as input the position at the time of recognition of the character code used to select the word code and outputs a search history indicating whether or not the position at the time of recognition of the word code has been searched with other candidate characters; For each word code output from the testing means, the method includes a judgment value accumulating means for adding an evaluation value indicating certainty, and a sorting means for extracting candidate words based on the magnitude relationship of the judgment values of the judgment value accumulating means, The selection processing means selects a word code corresponding to the character code from the word selection table for each one or more dictionary positions for each candidate character in each recognition position of the input character string, and the position verification means selects a word code corresponding to the character code for each position in the dictionary. The word code and the search history of the position at the time of recognition are read from the position history table, and only if the search history has not yet been searched, the search history of the searched position is set in the position history table, and the word code is stored in the judgment value accumulating means. and the judgment value accumulating means cumulatively adds the evaluation values of the word codes output from the position verification means to obtain a judgment value for the input character string, and the sorting means extracts candidate words. A word matching method that uses (2) The word matching method according to claim 1, wherein a measure of certainty obtained from the recognition result of each character of the input character string is written as the search history in the position history table.