JPS62278689A - Word retrieving system - Google Patents

Abstract

PURPOSE:To fetch a desired word even if a wrong character or an unreadable character is contained in several retrieval keys, by generating plural pieces of retrieval keys from a partial character string generated from adjacent characters in an input character string. CONSTITUTION:A retrieval key converting part 3 fetches a possible partial character string in an input character string stored in an input character string register 2, converts it successively to a key value and stores it in a retrieval key register 4. An index control part fetches successively the retrieval key value from the retrieval key register 4, fetches a dictionary pointer group to a word containing the retrieval key value from a dictionary pointer storage part 7 by using an index table 6, and transfers it to a dictionary pointer counting part 8. The dictionary pointer counting part 8 transfers plural pieces of different dictionary pointers in a dictionary pointer register 9 and its counting value to a candidate pointer selecting part 10. The candidate pointer selecting part 10 reads out successively a dictionary pointer and its counting value from the dictionary pointer register 9, and when a counting value T exceeds a threshold T, it is decided to be the dictionary pointer of a candidate word, and transferred to the index control part 5.

Description

[Detailed Description of the Invention] 3. Detailed Description of the Invention (Field of Industrial Application) The present invention is a word search method, in particular, a word search method in which an input character string is read by an optical character reading device (hereinafter referred to as OCR). The present invention relates to a word search method for searching a word dictionary for a plurality of words using partial strings of an input string when a character string includes misread characters or unreadable characters.

(Prior Art) Katakana, alphabets, etc. are often used in the form of proper nouns or common nouns, such as names of people, places, and products, in mail, forms, etc. that are read by character reading devices. Unlike numbers, these characters often have strong dependencies between characters within a word, and have sufficient redundancy. Therefore, if words are recognized as units, by utilizing their dependencies and redundancies, it is possible to correct misread characters and recover unreadable characters, and the recognition rate can be considerably improved. Such word-by-word recognition will hereinafter be referred to as word recognition. Generally, in such word recognition, a word dictionary storing personal names, place names, etc. is stored in a memory, a disk, or the like.

The problem in this case is that as the number of words included in the word dictionary increases, the time required to search them increases, resulting in a significant reduction in recognition speed. In order to reduce the problem of slow recognition speed when performing word recognition using such word dictionaries, we can perform simple conversion using a specific substring of the input string output from OCR as a search key. A commonly used method is to perform an operation and use a hash table to search for a plurality of words that match a search key.

Here, if the input string contains an error character ejection state (unreadable or multiple candidate characters are output), the specific substring mentioned above will contain a word that matches the input string. It may be difficult to search for multiple words in certain situations.

In order to solve the above problems, conventional methods have been used, for example, Proceedings of the 20th National Conference of the Information Processing Society of Japan in 1978, Nos. 487-4.
As shown in page 88, 6I'-31'-0CR's Bad Word Recognition, a Confujo/・Matrix/((Confusi
There is a word dictionary search method that converts a specific partial character string into a search key based on ``on Matrix''. In addition, for example, the technical magazine "EEK Tran
saction on Compute-srs) J's V
ol C-27+ 48+ August" 1978, "A Multi-Font Word Recognition System for Vostal Address Reading (
a multifont word rscogni-
tion system for postal ad
128th to 13th entitled dress reading J
As shown on page 0, there is a search method that uses two specific partial strings, the first half and the second half of an input string, as search keys. In addition, with the above-mentioned method, if a predetermined partial character string selected as a search key contains a character cutout error such as one character changing to two characters, it becomes difficult to perform a word search. As described in Japanese Patent Application No. 56-103554 "Word Search Device" filed by J.D., there is a method to deal with this problem by preparing in advance a character category string in which character segmentation errors may occur.

(Problem to be Solved by the Invention) As mentioned above, even if a specific substring is set and a conflation matrix is used, the length of the substring changes due to character extraction errors, etc. It becomes difficult. In addition, with the method of preparing in advance combinations of character categories that may occur due to character extraction errors, when handling mail items that use printed characters, etc. that include various types of characters, it is necessary to Since the patterns are different and also depend on the character extraction method used in OCR, it is difficult to prepare in advance possible combinations of character categories. SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, it is an object of the present invention to provide information on the relative position of possible substrings generated from adjacent characters (for example, subcharacters including the starting character of an input string). (such as partial strings that do not include the start or end characters of a column or input character string) are converted into search keys, and the same word is searched for among multiple words that can be searched using the multiple search keys mentioned above. By selecting the word corresponding to the input string based on the number of search keys, the desired word or multiple words containing the desired word can be selected even if some of the search keys contain incorrect characters or select characters. The purpose of the present invention is to provide a word search method that allows the selection of words.

(Means for Solving the Problems) In order to solve the above-mentioned problems and achieve the above objectives, the word search method provided by the present invention uses partial strings in an input string to be stored in a word dictionary. A word search method for searching for a predetermined word includes means for converting a partial string of adjacent characters of the input string into a search key including relative position information; pointer storage means for storing a pointer indicating a storage position of a word included in a block for each different search key; and a pointer storage means for storing a block storage position of a block for each search key within the pointer storage means at a predetermined location determined by the search key. A plurality of possible search keys are generated from the index table stored in the address and the input character string, and a plurality of blocks corresponding to the plurality of search keys are generated from the pointer storage means using the index table. and means for selecting a word corresponding to the input character string based on the number of pointers indicating the same word included in the plurality of blocks and the number of the input character strings. It is characterized by

(Operation) In the present invention, in order to generate a plurality of search keys from possible substrings generated from adjacent characters in an input string, some of the search keys include incorrect characters or unreadable characters. It is possible to extract the desired word even if it is included.Here, by using multiple search keys, the desired word can be extracted accurately, but it also includes a group of words that are more than necessary.Therefore, In the present invention, when converting possible substrings generated from an input string into search keys, the search keys are generated including the relative position information of the substrings, and the search keys are Among a plurality of searchable words, it is possible to select only a group of words similar to the input character string based on the number of search keys that searched for the same word.

(Example) Hereinafter, an example of the present invention will be described with reference to the drawings. FIG. 1 shows an example of a method for calculating a series of search keys from an input character string. FIG. 1(a) shows an example in which 26 alphabetical characters are associated with unique code values. In Figure 1(a), Alphabella)A has a code value of 0, Alphabella)B has a code value of 1, etc.
..., so when converting any substring of an input string into a search key, you can convert it into a substring by a simple calculation using the - code value mentioned above. A uniquely corresponding search key value (hereinafter referred to as search key value) can be obtained. In the following description, words and input character strings are assumed to be 26 characters of the alphabet, but it goes without saying that the same applies to katakana and special symbols, or even if they are used at the same time.

Figure 1(b) shows the input character strings a, a,...al(
However, al is Alphapella F Am B+ c+
...any song z, 1=1.・・・・・・n
) using two adjacent characters to generate a partial character string and calculate a search key value. Furthermore, Figure 1 (
In b), two adjacent characters are shown, but it goes without saying that three characters can also be used. Figure 1
In (b), the symbol PI is the value for the position information of the two adjacent characters aiai+1 for which the search key value is to be calculated. For example, if the character material of the two adjacent characters "i ai+1" is the first character, the symbol P is the value 0, two adjacent characters a1a1+,
If the character ``al'' is the last character of the input string, the symbol EX has a value of 2, and the two adjacent characters ``a1a1'' are both
If the character is neither the first character nor the last character of the input character string, a value of 1 is given as the symbol P. Therefore, adjacent 2
The character ai ai+1 is the conversion formula P・26″+ai・
26 + al4. By using , a search key value can be obtained at will, and even if two adjacent characters aiai+1 are composed of the same characters, for example, a different search key can be generated depending on whether the character a1 is the first character or not. Become.

FIG. 2 shows an example of a method for retrieving a desired word using a search key generated from an input character string.

In the figure, the input string? ? BCD1! Word ABCDRIF corresponding to K (? indicates unreadable state)
(For example, when OCR reading results in character A becoming ?? due to a character cutting error and character F being mistakenly recognized as character E), the method for extracting the character is as follows.

First, the conversion formula PI・262+吋・'l (i+ai++ (in FIG. 2, 31=B,
c.D.

”' ai+1 = C+ D+ K, W)
Using , two adjacent characters BC, CD, DB,
I! J! : is converted to a search key. For example, two adjacent characters BC are l・26t+1・26+2=704 (11
=l, al is the letter B and the value is 1+ ai+1 is the letter C and the value is 2), and the value 704 is used as the search key.
is obtained. In the same way, two adjacent characters CD, DJ K
W also has a search key value of 731,758.14.
60 is obtained. Note that if a plurality of recognition results for one character image are obtained from OCR, it is assumed that search key values have been generated for all combinations of two characters by the method described above.

Next, the two adjacent characters BC, CD, and DI mentioned above.

The index table is accessed using the search key values 704, 731, 758°1460 for KBli.

In the index table, the starting position of a block of a word group containing two adjacent characters uniquely determined by the search key is stored in the dictionary pointer storage at an address corresponding to the search key value.

For example, in the search key value 704 for two adjacent characters BG in the input string, the address 704 of the index table has the following information:
Pointer P(1, BC) (However, 1 indicates the value of relative position information P described in FIG. 1, and pointer P(1, BC)
, BC) indicates the start position of block BIC (1, BC) of a word group having two adjacent characters BG with relative position information P = 1, so block BK in the dictionary pointer storage unit is stored. (1, BC) can be directly obtained using the index table. Here, search key value 70
For example, in the word dictionary shown in FIG.
Since storage locations such as F, FBCAAA (hereinafter referred to as dictionary pointers) P, ,P, etc. are stored, by taking out the dictionary pointers P, ,P, the input character string 7'1
Word AEC containing two adjacent letters BC in JBCDKK
DIItF, FBCAAA, etc. are obtained. Similarly, the dictionary pointer P is used for two adjacent characters CD in the input string.
,,F,,p,.

P, Ka, two adjacent letters D]l! ! For, the dictionary pointer P,, P,, p, has two adjacent characters 11! A dictionary pointer po, p is obtained for K.

In addition, in FIG. 2, the word F! Adjacent 2 to FF and CCD
Does the input string include CD as a character? ? Although BCDKK also includes two adjacent characters CD, in the present invention, the search key value 731 is calculated in consideration of the relative position information PI of the input character string. Therefore, the word KFFCCD is not extracted due to the two adjacent characters CD of the input character string, and it is possible to restrict extraction of more words than necessary.

Next, the input string? ? BCDKK possible search key values 7
Blocks BK (1, BC), BK (1° CD), 'B extracted using 04,731,758.1460
K (1, DB), dictionary pointer p0. in BK (2, 1nK). The number of p,, p,, p,, p, is calculated. That is, the dictionary pointer P0 is used once, the dictionary pointer Pl is used three times, the dictionary pointer p is used three times, the dictionary pointer P is used twice, and the dictionary pointer P is used three times.

will be once. Input character string 77BCDF here! The candidate word for K is the dictionary pointer p0. p.

Count values of p,, P,, P, 1. 3. 3. 2.
1 and the number of characters jin=7 in the input character string.

As an example, a threshold value is calculated as T=α·(tin-1) (0<α≦1), and a dictionary pointer having a count value of 7 or more is selected. For example, in the example above,
When α=0.5, the threshold T is 3, and the dictionary pointers P, , P2 with a count value of 3 or more, that is, the candidate word ABCD
II! F, CDCD1iK will be selected. As another example, the number of characters in the input character string is not used, and the candidate word A B is simply retrieved by taking out the dictionary pointer, that is, the word that has the maximum count value of the dictionary pointer.
CDKF.

You can also take out the CDCDIK.

As shown in the above example, if the word corresponding to the input character string cannot be uniquely determined, for example, Japanese Patent Application No. 55-126244 "Difference Detection Device" filed by the applicant of the present application. This can be uniquely determined using a word matching device that matches input character strings and words as described in . Furthermore, it goes without saying that the dictionary pointer storage section and the index table described above can be created in advance using an ordinary microprocessor or computer once the word dictionary to be used is determined.

FIG. 3 is a logical block diagram showing one embodiment of the present invention. In the figure, 1 is a character reading device (OCR),
The character strings read by 0CR1 are sequentially stored in the input character string register 2. 3 is a search key conversion unit, which extracts possible substrings from the input string stored in the input string register 2 and sequentially converts them into search key values, as explained using FIG. Store in search key register 4. When a plurality of search key values have been set in the search key register 4, the index control unit 5 sequentially retrieves the search key values from the search key register 4, and uses the index table 6 as shown in FIG. A group of dictionary data for words including the search key value (partial character string in the input sentence 11) is taken out from the data storage section 7 and transferred to the dictionary pointer counting section 8. The above operation is performed on all search key values stored in the search key register 4. Next, if the dictionary data transferred from the index control unit 5 is the same as the dictionary data already transferred, the dictionary pointer counting unit 8 counts up the count value for the dictionary data. By calculating the count value for different dictionary pointers, the dictionary pointer register 9
Then, a plurality of different dictionary pointers and their counts are transferred to the candidate data selection unit io.

The candidate pointer selection unit 10 selects II! from the number of characters lin of the input string stored in the input string register 2 and a predetermined coefficient α (0<α≦1). 4-value T(=α・(zi
n-1)). Next, 1 candidate pointer 1 selection part 1
0 reads the dictionary data and its count value sequentially from the dictionary pointer register 9, and if the count value is equal to or greater than the threshold value 7, it is determined that the dictionary pointer is a candidate word and is transferred to the index control unit 5. Note that the candidate word selection unit 10 determines the dictionary pointer with the largest count value among the count values of the plurality of dictionary pointers stored in the dictionary pointer register 9 as the dictionary pointer of the candidate word. You may use a method. When the dictionary pointer of the candidate word is transferred to the index control unit 5 from the candidate pointer selection unit 10, the index control unit 5 retrieves the word corresponding to the transferred dictionary pointer from the word dictionary storage unit 11 and stores it in the candidate word register. Transfer to 12.

As described above, the word corresponding to the input character string stored in the input character string register 12 is set in the candidate word register 12.

Note that when multiple words are set in the candidate word register 12, the input character string is compared with the word using a normal word matching device (not shown in the figure), and the word is determined to be a specific word. .

(Effects of the Invention) As explained above, according to the present invention, when extracting a desired word using a partial string in an input string as a search key,
Even if some of the search keys have erroneous characters or are unreadable, it is possible to stably and efficiently select a desired word or a plurality of words containing the desired word.

[Brief explanation of drawings]

FIG. 1 is a diagram illustrating an example of converting an input character string into a series of search keys. FIG. 2 is a diagram showing an example of a method of indexing a desired word using a series of search keys. FIG. 3 is a logical block diagram illustrating one embodiment of the present invention. In the figure, 1 is 0CR12 is the input character string register, 3
1 is a search key conversion unit, 4-digit search key register, 5 is an index control unit, 6 is an index table, 7 is a dictionary boiler storage unit,
8 is a dictionary pointer counting section, 9 is a dictionary pointer register, lo is a candidate data/data selection section, 11 is a word dictionary storage section, and 12 is a candidate word register.

Claims (1)

[Claims] In a word search method that searches for a predetermined word stored in a word dictionary using a substring in an input string, a search key is used to search for a substring consisting of adjacent characters of the input string, including relative position information. a pointer storage means for storing a pointer indicating a storage position of a word containing a partial character string corresponding to the search key in blocks for each different search key; The pointer storage means generates an index table in which the block storage position of each block is stored at a predetermined address determined by the search key, and a plurality of possible search keys from the input character string, and uses the index table to generate a plurality of possible search keys. means for extracting a plurality of blocks corresponding to the plurality of search keys from the plurality of search keys, based on the number of pointers indicating the same word included in the plurality of blocks and the number of the input character strings, A word search method comprising means for selecting a word corresponding to the input character string.