JPH1021262A - Information retrieval device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To execute fuzzy retrieval of a desired character string by accessing a dictionary by means of an address with the group of character strings described in an inter- character transition information table and displaying the group of the character string in a dictionary corresponding to the address as a retrieval result. SOLUTION: An inputted handwritten pattern is character-recognized a3 by matching with a character pattern in a recognition dictionary a3. In order to optimize the candidate precedence of a candidate character obtained as the result of character recognition, a character transition and probability table a4, which is prepared by learning and describes transition probability between characters and pointers of candidate words in a word dictionary including character transition, and a character appearance probability table d8 are referred to and the candidate precedence is optimized in order to improve appearing probability as a character string, and a pointer to a word in the word dictionary to which the optimized character string corresponds to (a5). The obtained word in the word dictionary a6 and the candidate character string are collated to obtain (a7) the pertinent word and information before/after it from hierarchical information in the word dictionary to display the result on LCDa8.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an information retrieval apparatus for retrieving a character string registered in advance based on a key character input later.

[0002]

2. Description of the Related Art Conventionally, when searching words in a word dictionary,
It is common to search by inputting the whole word or the character string at the beginning of the word. For example, in the Japanese syllabary search for a car navigation system, when you start typing characters from the beginning of a word,
The candidates are narrowed down to words that include the input character at the beginning of the word.

[0003] In another conventional example, input of an address or a fixed phrase is often used in slip processing of a PenPC (pen input computer) or the like. Conventional methods of inputting addresses and fixed phrases include (1) menu selection, and (2) a combination of character recognition and menu, when a zip code is input, a relevant address is displayed in a menu and selected.
(3) A method of recognizing a handwritten input character and optimizing the candidate with a word dictionary is disclosed.

The method using the menu (1) is, for example, as follows:
Reference "Frameless Handwritten Address Recognition Allowing Character Position Shift"
(Transactions of the Institute of Electronics, Information and Communication Engineers D-2, January 1994)
As disclosed in the above, in data hierarchized like an address, a method of selecting sequentially from the upper hierarchy to the lower hierarchy (for example, "Ibaraki prefecture" → "Hitachi city" → "Omika town") Is common. Therefore, if you enter "Hitachi City", but you do not know in which prefecture "Hitachi City" is located, for example, "Ibaraki Prefecture" or "Tochigi Prefecture", there is a problem that you cannot easily select "Hitachi City". is there.

[0005] In the method of (2), when the postal code is input, the relevant address is displayed on the menu by inputting only the postal code as a character, and after that, the place name determined from the postal code can be selected from the menu. It is possible. However, it is difficult to memorize the postal code of his / her address, but also to other postal codes.

A method of (3) recognizing a handwritten input character and optimizing the candidate with a word dictionary will be described with reference to the drawings. FIG. 3 is a configuration diagram of conventional character recognition. The handwritten pattern input from the tablet a1 is subjected to character recognition a3 by pattern matching with the recognition dictionary a2, the obtained candidate characters are subjected to word matching b7 with the word dictionary a6, and the corresponding words are displayed on the LCD a8.

FIG. 4 is a schematic diagram showing the operation when an address is input using conventional character recognition. For example, to enter "Omika Town, Hitachi City, Ibaraki Prefecture," enter "
"Omika-cho" is input by hand in the predetermined address input area b1. Then, the character recognition a3 is performed, and the obtained candidate characters are word-matched from the highest hierarchy prefecture using the word dictionary a6, the candidate characters are optimized, and the result is output as a candidate character.

Conventionally, hierarchical data such as an address is accessed from the highest hierarchical level because, generally, the higher the hierarchical level, the smaller the data amount, and if a higher level is determined, lower candidates can be narrowed down. If you search the word dictionary by inputting the keyword "Omika" in the conventional system, you cannot know which level it is in, so you have to do a full-text search for a word dictionary of about 1.5MB. This is because it is not practical when responsiveness is required as in character recognition.

((Example of address word dictionary) Prefecture: about 50 × about 3 characters × 2 bytes = about 300B Municipalities: about 4000 × about 3 characters × 2 bytes = about 2.5KB Less: about 160,000 X about 4 characters x 2 bytes = about 1.3 MB Total: about 1.5 MB) However, as in the conventional method, for example, "Imaki-cho, Hitachi City, Ibaraki prefecture" If not, it would be difficult for the user.

[0010]

In the conventional method of inputting characters from the beginning of a word and narrowing down candidates (front matching method),
Without knowing the beginning of the word, the desired word cannot be searched. For this reason, there is a problem in that even if one knows the middle or the end of a word by learning the information, the information is not used.

[0011] In addition, as a problem common to the above-mentioned prior arts (1) to (3), in the case of using character recognition, it is necessary to input all addresses and fixed phrases by hand, which is troublesome for the user. Searching a word dictionary from a hierarchy other than the highest level requires a huge amount of time to search, and when selecting from a menu, if the hierarchy is like an address, if the upper level is not known, the lower level is selected. There is a point that cannot be done.

An object of the present invention is to provide an information retrieval apparatus capable of performing an ambiguous search for a target character string by using not only the head or end characters of the character string to be searched but also information on spelling as a key input.

Still another object of the present invention is to provide an information retrieval apparatus which can add key input information without concern for the order when key characters are input and candidates cannot be narrowed down.

[0014]

According to the present invention, there is provided an information retrieval apparatus for retrieving a character group including an input character string, a dictionary for storing a plurality of character string groups in advance at respective unique addresses, A character transition information table storing at least transition information from one character to another character and an address of a character string group including a combination of the characters having the transition relationship among the plurality of character string groups, The dictionary is accessed by an address described in the character transition information table for a character string group corresponding to the input character string, and a character string group in the dictionary corresponding to the address is displayed as a search result. That's what I did.

Thus, a desired character string can be obtained at a high speed by appropriately inputting a part of the character string to be searched.

[0016]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of the present invention applied to a device for inputting an address by handwriting and an example applied to an ambiguous word search.

[Handwritten Address Input Device] FIG. 2 is a schematic diagram showing the operation of a handwritten input address recognition device according to one embodiment of the present invention. In the address input area b1, with the stylus pen a9,
Hand-write the character "Omika" which is the key of the address. Then, the handwritten stroke (handwriting) is recognized by character recognition a.
3 to obtain a keyword candidate character, a first candidate "Omika", and a second candidate "Inukaru". The candidate characters are determined by referring to the transition probability between characters (probability following (character A) → (character B)) in the character transition probability table a4 and the character appearance probability table a8. When combined, it is determined whether the probability of appearance as a character string is high, and so that the probability of appearance as a character string is high,
Optimize the order of candidate characters. Further, referring to the candidate word No. in the character transition probability table, a pointer to a word containing a character transition in the candidate character string is obtained. In the example of the figure, the appearance probability increases when "Omika" is combined with the candidate character. From the character transition probability table, (Omi → Mi) and (Mi →
The pointer No. A11 of the word containing the word "?" In this method, a target word can be accessed at high speed by newly providing a pointer to a word in the word dictionary in the transition probability table used for post-processing of character recognition. And if you find the corresponding word in the word dictionary,
The upper hierarchy is uniquely obtained from the hierarchy information of the word dictionary. In this example, "Omika-cho" at the lowest level is uniquely obtained, so that "Hitachi City, Ibaraki Prefecture" can be uniquely obtained at the higher level. Therefore, even though only "Omika" was handwritten, "Omika town, Hitachi City, Ibaraki Prefecture"
The address can be displayed as a candidate from the beginning and input.

FIG. 1 is a functional block diagram of a handwritten input address recognition apparatus according to one embodiment of the present invention. The operation will be briefly described with reference to FIG. By sliding the stylus pen a9 on the tablet a1, a handwritten character pattern is input. The input handwritten pattern is converted into a recognition dictionary a
Then, character recognition a3 is performed by matching with the character pattern in step 2. In order to optimize the candidate rank of candidate characters obtained as a result of character recognition, pointers of candidate words in the word dictionary, which include learning and transition probability between characters and character transitions, which are created in advance by learning. With reference to the character transition probability table a4 and the character appearance probability table a8, the candidate order is optimized so that the appearance probability as a character string is increased, and the optimized character string is added to the word in the corresponding word dictionary. A pointer is obtained (a5). The words in the word dictionary a6 obtained by the above method are compared with the candidate character strings, and the corresponding word and information before and after the word are obtained from the hierarchical information of the word dictionary (a7). The obtained result is displayed on LCDa8.

By using the [1] method of creating the character transition probability table a4 and [2] the character transition probability table a4, which are the features of the present invention, the probability of occurrence of candidate characters as a character string is increased. A method of optimizing the candidate order and obtaining a pointer to a word including a character transition in the optimized character string, [3] a character string including a handwritten input word (keyword) using the obtained word pointer For how to guess, and [4] how to display the guessed string,
This will be described later in detail with reference to the drawings.

FIG. 5 is a hardware configuration diagram of a handwritten input address recognition device according to an embodiment of the present invention. Function to match input pattern and dictionary pattern, function to optimize candidate characters obtained by pattern matching, find pointer to word containing candidate character, function to match candidate characters to words, and search for information before and after Is, for example, RO
CPUc3 reads the program stored in Mc5,
It is realized by executing. Each function will be described later in detail with reference to the drawings.

[1] Method of Creating Character Transition Probability Table a4 FIG. 6 is a flowchart illustrating a method of creating the character transition probability table a4. First, d1 for reading out the word dictionary a6 for learning. For the read word, the number of appearances of the characters included in the word is counted (d2), and recorded in the appearance number table d3. Subsequently, the number of character transitions is counted (d4) and recorded in the transition number table d6. continue,
D5 for recording the word No. (word pointer) of the original word including the character transition in the transition count table d6. For example, in the example of the word "Omika", the number of occurrences of a character is
"Large", "mi" and "ka" are counted once each, and in the number of character transitions, "large → mi" and "mi → ka" are counted once each. In addition, the word No. of the word "Omika" is recorded in the candidate words No. of "O-mi" and "Mi-oka". Such processes d1 to d6 are repeatedly performed for all the words in the word dictionary a6. Thereafter, referring to the character appearance frequency table d3 and the character transition frequency table d6 created by the above-described processing, a character transition probability table a4 describing information on transition probabilities between characters and candidate words No. including character transitions is created. (D9). Further, if necessary, a character appearance probability table d8 describing the appearance probability of each character is created with reference to the character appearance frequency table d3 (d7). The configuration of each table will be described below with reference to the drawings.

FIG. 7 is a schematic diagram of a word dictionary. Here, as an example, an address word dictionary a6 is shown. This dictionary is
It has a hierarchical structure of prefectures, municipalities, and lower levels, and each element is composed of a word No. (word position information), an upper word No., and a word (character string). Each word can be accessed by the word No. The hierarchical relationship between words can be known from the upper word No.

FIG. 8 is a schematic diagram of a character appearance frequency table. The character appearance frequency table d3 stores the number of times each character appears in the learning word dictionary, and is used to count the total number of characters in the learning word.

FIG. 9 is a schematic diagram of a character transition number table. The character transition number table d6 is used to record the number of times each character transition appears in the learning word dictionary and the No. (candidate word No.) of the word containing the character transition.

FIG. 10 is a schematic diagram of a character appearance probability table. For example, the appearance probability of the character “day” can be represented by appearance probability (day) = number of appearances (day) / total number of characters. Therefore,
In order to create the character appearance probability table d8, the number of appearances of each character may be obtained by referring to the appearance frequency table d3 and recorded in the character appearance probability table d8.

FIG. 11 is a schematic diagram of a character transition probability table. For example, the transition probability of the character transition "day → standing" is: transition probability (day → standing) = number of transitions (day → standing) / number of appearances (day)
Can be represented by Therefore, in order to create the character transition probability table a4, the appearance number table d3 and the transition number table d6
, The transition probability of each character may be obtained and recorded in the character transition probability table a4.

With the above processing, a character transition probability table a4 with a pointer to a word in the word dictionary can be created. In this embodiment, the word No. is used as word identification information for identifying each word in the word dictionary. However, the present invention is not limited to this, and may be code information. Similarly, the upper word No. is used as the hierarchical information indicating the hierarchical relationship between the words in the word dictionary. However, the present invention is not limited to this. It may be.

[2] Post-Processing Based on Transition Probability (Optimization of Candidate Characters and Determination of Candidate Word Pointer) FIG. 12 is a schematic diagram of character recognition post-processing performed using a transition probability table. As shown in the figure, it is assumed that the candidate characters of the input pattern "Omika" are the first candidate "Omika" and the second candidate "Inumika". It should be noted that, here, in order to show that even if there is an error in the character recognition result, it is possible to cope with the optimization of the candidate character based on the transition probability. I do. There are eight possible combinations of candidate characters shown in the figure. For the set of these character strings, the character transition probability table and the pointer to the word candidate word in the word dictionary are obtained with reference to the character transition probability table. A combination of candidate characters having a high appearance probability of a character string is an optimal combination (path). Therefore, the character set is used as the result of the transition probability post-processing together with the candidate word No.

First, a method of calculating the appearance probability of a character string will be described.

In general, the appearance probabilities of the character strings S1 S2... Sn can be approximated by the following equations.

[0031]

## EQU1 ## Appearance probability (S1 S2... Sn) = appearance probability (S1) .times.transition probability (S1.fwdarw.S2) .times..times.transition probability (Sn-1.fwdarw.Sn)... In the example of “Omika” in .3, it is as follows. Appearance probability (Omika) = Appearance probability
(Large) × transition probability (large → mi) × transition probability (mi → ka) Specifically, the character transition probability table a4 and the appearance probability table a8
The value is obtained by the following equation.

[0032]

[Number 2] occurrence probability (Omika) = (18/89274) × (1/18 ) × (2/15) × (2/198) = 14940 × 10 ~ 8% ... ( number 2) Next, the candidate word No. (A method of obtaining a pointer 0 of a word candidate in a word dictionary will be described. A candidate word No. of a character string S1 S2... Sn can be obtained as in the following equation.

[0033]

[Equation 3] Candidate word No. (S1 S2... Sn) = Candidate word No. (S1 → S2) ∩ ∩ {Candidate word No. (Sn-1 → Sn) ... (Equation 3) In the case of "Omika", the following expression is obtained from the character transition probability table a4.

[0034]

(Equation 4) Candidate word No. (Omika) = {Candidate word No. (Obvi → Mi)} Candidate word No. (Omika) = (A11) ∩ (A11, A12) = A11 ... (Equation 4) In the example, the candidates are narrowed down to “Omika-cho” (A11) by taking an AND (∩), but the candidates may be “Omika-cho” and “Mikanohara-machi” by taking OR (∪). In this case, considering the number of matches, “Omika-cho” can be the first candidate. Details will be described later with reference to the drawings.

According to the above method, the candidate character obtained by the character recognition is optimized so that the probability of appearance as a character string is increased, and a candidate word No. (pointer to a word) of the optimized candidate character string is obtained. be able to. According to the above method, for example, even if only "Mika" which is a part of "Omika-cho" is input, "Omika-cho" and "Mikahara-cho"
Can be a candidate.

[3] Words (Keywords) Input by Handwriting
Next, a method of estimating the entire character string including the optimized candidate character string (keyword input by handwriting) will be described with reference to the drawings.

FIG. 13 is a schematic diagram for estimating a character string based on the post-processing result. In this example, the keyword (candidate character) “Omika” and the candidate word No.
11 shows the case where 11 was obtained. As you can see from this figure,
The keyword “Omika” has only one in the word dictionary and is the lowest level, so the high-order character string “Hitachi, Ibaraki” is uniquely determined. Therefore, as a result of estimating the entire handwritten character “Omika”, the candidate character string “Hitachi
Omika-cho "is displayed.

FIG. 22 is an example similar to FIG. 13, but as a candidate character string for the handwritten character “Omika”, the character string guess result “Omika-cho, Hitachi City, Ibaraki Prefecture” and the character recognition result without the guess result are shown. The feature is that "Omika-cho" is selected as both candidates so that the user can select it. In this way, when both the character string guessed from the keyword and only the keyword without guess are set as candidates, there is an advantage that the user can select a favorite one.

FIG. 14 is a schematic diagram for estimating a character string based on the post-processing result. In this example, a case where “Hitachi” (candidate word No. A1), which is a word in the middle hierarchy of the word dictionary, is obtained as a result of the post-processing by the transition probability. As can be seen from the figure, the keyword “Hitachi” has 1 in the word dictionary table.
Since there is only one, the upper-level character string "Ibaraki" is uniquely determined. However, since there are multiple lower layers, for example,
A plurality of candidates may be displayed as "Omika-cho, Hitachi-shi, Ibaraki", "Mikahara-cho, Hitachi-shi, Ibaraki", and the user may select one.

FIG. 15 shows a word "Hitachi" (candidate word No. A) which is a word in the middle hierarchy of the word dictionary as a result of post-processing as in FIG.
1) In this example, there are a plurality of lower-order candidates obtained by the transition probability. In the example of this figure, the word dictionary holds population information for each city, and displays a town with a large population in a plurality of lower layers in a priority order with higher priority.
Operability has been improved.

FIG. 16 is similar to FIG. 15, but the word dictionary has transition probabilities between words instead of population information. The operability can be improved by preferentially displaying words having a high transition probability in a high candidate order among a plurality of lower layers. The transition probability between words can be learned from a text or the like where a place name frequently appears. Further, simply, the frequency may be determined based on the population of the city, and the learning may be performed.

FIG. 21 shows a keyword "Hitachi" (word No. A) in the middle hierarchy of the word dictionary as a result of post-processing as in FIG.
1) is obtained, and lower candidates are not uniquely determined. The difference from FIG. 14 is that only “Hitachi City, Ibaraki Prefecture”, for which a candidate has been uniquely determined, is displayed, and for cities, towns and villages that are not uniquely determined, the user inputs again.

As can be seen from the above, the hierarchy higher than the keyword can be automatically and uniquely guessed, and if there are a plurality of candidates lower than the keyword, it cannot be uniquely determined. You only need to input that part.

[4] Method of Displaying Guessed Character String Next, a method of displaying the character string candidates obtained by the above method will be described. FIG. 17 to FIG. 20 are diagrams showing display forms of the estimated character strings. In each case, "Omika-cho" is entered and the character string of the guess recognition result is displayed, "Omika-cho, Hitachi City, Ibaraki Prefecture".

In FIG. 17, the part of "Omika-cho" in the candidate character string "Omika-cho, Hitachi-shi, Ibaraki" is displayed near the handwritten "Omika-cho". This display method has an advantage that it is easy to compare handwritten handwriting with a candidate for a recognition result.

FIG. 18 shows "Omika-cho, Hitachi City, Ibaraki Prefecture" from near the handwritten position of "Omika-cho". In this display method, since the heads of the handwritten characters and the candidate character strings are aligned, there is an advantage that the display start position of the result can be easily recognized.

In FIG. 19, a candidate character string "Omika-cho, Hitachi-shi, Ibaraki" is displayed centering on the position handwritten as "Omika-cho". This display method has an advantage that the positions of the handwritten character and the candidate character string are the closest as a whole, and the viewpoint does not need to be moved much.

In FIG. 20, it is displayed that the position of the end handwritten as "Omika-cho" and the position of the last character of the candidate character string "Omika-cho, Hitachi City, Ibaraki Prefecture" are aligned. This display method has an advantage that a candidate character string is not hidden under a hand holding a pen, especially for right-handed users.

As shown in FIGS. 17 to 20, there are characteristic display methods, but the system may appropriately determine any one of the display methods. You may make it selectable.

In FIGS. 17 to 20, the keyword "Omika" in the candidate character string is displayed in reverse video for easy understanding. However, other methods such as changing the font may be used. In addition, it is not necessary to perform the reverse display or change the font.

As described above, according to the present embodiment, when inputting hierarchical data such as an address, a fixed phrase, or the like, the entire character string can be guessed and displayed as candidates only by inputting a key character. Can be. Further, the word dictionary can be accessed at a high speed to perform word collation, and even when the word dictionary is hierarchized as in the above-described embodiment, the word dictionary can be accessed at a high speed regardless of the upper or lower level of the hierarchy.

[5] Fuzzy Search for Key Input Characters Next, a fuzzy search using input characters as keys will be described. Here, the ambiguous search means that a search is performed using ambiguous information that does not completely match the key input, is partially missing, or has a partially different order of key characters as a key. .

In FIG. 23, "Omika" is input as a key. When this is decomposed into character transitions, as shown in the figure,
It is divided into “Large → Mi” and “Mi → Ka”. The characters that include these character transitions are:
"Mikakamachi" and "Mikanoharamachi" are candidates for "Mikaka". Since both of these are place names under “Municipalities and below,” they are not linked together in a hierarchy. Therefore,
Next, the score of each of the two candidates (the number of matching key characters and the number of matching character transitions) is calculated. For character matching,
"Omika-cho" has 3 points because the key character "Omika" matches 3 characters, and "Mika Hara-machi" has 2 points because the two characters "Mika" among the key characters "Omika" match. Count. In the character transition, "Large → Mi" and "Mi → Ka" correspond, so two points, "Mika Haramachi" is "Mi → Hara"
Since only "?" Is applicable, it can be calculated as one point. FIG.
In the example of No. 3, priority was given in the order of (the number of matching key characters> the number of matching character transitions of the key characters> the order of the corresponding key characters> the key character matching position). Therefore, "Omika-cho", which has the highest priority and the highest score of the number of matching characters, is ranked first, and "Mikahara-cho" is ranked second. Referring to the word dictionary a6, the first is:
"Omika-cho, Hitachi City, Ibaraki Prefecture", 2nd place: "Mikanohara Town, Hitachi City, Ibaraki Prefecture".

As described above, when searching for an address that includes the input key character "Omika" (or even a part thereof), display priorities are assigned to the search results, and a likely result is displayed first, so that operability is improved. It is necessary to improve. The priority order is the number of matching key characters, the number of matching character transitions of key characters,
At least one of the hierarchical order of the key characters, the position where the key characters match, the arrangement order of the key characters, the distance between the key characters, and the frequency of use of the words may be considered. In the example of FIG. 24, "Mika" and a key character are input. As in FIG. 23, the place names including the character transition “Mika” are “Omikacho” and “Mikaharamachi”. In this example, both "Omika town" and "Mika Hara town"
Transition 1 including 2 key characters (2 points), same as key character
(One point), no priority is given from the number of character matches and the number of character transition matches. In addition, the hierarchy, both levels, etc. are at the same level of “Municipalities and below”. Therefore, in general, a character string including a key character at the beginning is often input. Therefore, “Mika Haramachi” including the key character “Mika” at the top is ranked first. The second place is Omika-cho. As the search results, the first place: “Mikahara-cho, Hitachi City, Ibaraki Prefecture” and the second place: “Omika-cho, Hitachi City, Ibaraki Prefecture” are displayed with reference to the word dictionary a6.

In the example of FIG. 25, "Mayumi Hitachi-Ota-shi" is input as a key input character. Breaking down "Hitachi Ota City Mayumi" into character transitions, "Hitachi""Rikuta""Ota""Taichi"
"Ichima" and "Mayumi". In the example shown in the figure, words including these character transitions include “Hitachi”, “Rikuta”, “Ota”, and “Taichi”, and “Hitachi-Ota-shi” includes words that include “Ota” and “Taichi”. "Ota City" and "Ota City" do not contain the words "Ota" or "Ichima" that contain "Ota", but "Mayu (Hitachi-Ota City)" or "Mayumi (Ikoma City)" as words that contain "Mayumi" ) "And" Mayumi Minami ". By the way, it can be seen from the word dictionary a6 that “Hitachi Ota City” − “Mayumi (Hitachi Ota City)” and “Ota City” − “Ota” have a hierarchical relationship. Therefore, "Mayumi Hitachi-Ota City" and "Ota City, Ota City" can be considered by connecting them.

The score based on the number of character matches of these words is:
"Hitachi Ohta City Mayumi" 7 points, "Ota City Ota" 3 points,
"Hitachi Ota City" has 5 points, "Ota City" has 3 points, "Ota City" has 3 points, "Mayumi (Ikoma City)" has 2 points, and "Mayumi Minami" has 2 points. The score based on the number of character transition matches of these words is 4 points + 1 point = 5 points for "Mayumi Hitachi-Ota City", assuming the sum of the number of character transitions of key characters included in these words. However, the key character "Ota" is regarded as having appeared twice, although the key character "Ota" has been written only once. Therefore, the weight is reduced, for example, 2 points + 0.1 points = 2 .1 point. For other words, score the number of character transitions of the key characters included in the word, 4 points for "Hitachi Ota City" and 2 points for "Ota City"
Point, "Ota City" also has 2 points, "Mayumi (Ikoma City)" has 1 point,
"Mayumi Minami" is one point. Therefore, priority is given by the number of character matches, priority is given by the number of character transition matches for characters with the same number of characters, and priority is given in the order of hierarchy to those that cannot be given priority. Rank: "Mayumi Hitachi-Ota City", 2nd place: "Hitachi Ota City", 3rd place: "Ota City, Ota", 4th place: "Ota City", 5th place: "Ota City Town", 6th place
Rank: Mayumi (Ikoma City), 6th: Mayumi Minami.
(The 4th and 5th ranks are given priority in order of hierarchy.) The search result refers to the word dictionary a6, and the 1st rank: "Ibaraki Prefecture
Mayumi-cho, Hitachi-Ota City, 2nd place: Yamashita-cho, Hitachi-Ota-shi, Ibaraki, 3rd place: "Ota-shi, Ota-shi, Gunma", 4th place: "Honmachi, Ota-shi, Gunma", 5th: "Kashihara-shi, Nara Prefecture" Ota City ”, 6th place:“ Mayumi, Ikoma City, Nara Prefecture ”, 6th place:“ Mayumi Minami, Ikoma City, Nara Prefecture ”.

In FIG. 48, "Mayumi Hitachi Ohta" is input as a key input character. Here, a method will be described in which priorities are assigned based on the total score of (key character matching number (a)> key character corresponding hierarchy (b)> key character matching position (c)> remaining character number (d)). By the way, if you break down "Hitachi Ota Mayumi" into character transitions, you can get "Hitachi", "Rikuta", "Ota", "Tama", "Mayumi"
The words containing these character transitions are "Mayumicho, Hitachi-Ota City", "Hitachi-Ota City", "Ota, Ota", and "Ota City".
It is assumed that "Ueota", "Mayumi (Ikoma City)" and "Mayumi Minami" are obtained. The score (a) based on the number of character matches of these words is:
Of the key characters, 1 point if included in the candidate word, 0 points if not included, 6 points for "Mayumi-cho, Hitachi-Ota-shi", 4 points for "Hitachi-Ota-shi", and 2 points for others Becomes The score (b) according to the hierarchy is 3 points at the prefecture level,
If the city / county level is 2 points and the level below the towns and villages is 1 point, and there are multiple levels, add them. Then, “Mayumi-cho, Hitachi-Ota-shi” and “Ota, Ota-shi” including cities, counties and municipalities are 2 + 1.
Points = 3 points, "Hitachi Ota City" and "Ota City" of cities and counties are 2 points,
"Kamiota", "Mayumi (Ikoma City)" and "Mayumi Minami" will be one point. Next, in the evaluation value (c) based on the key character matching position,
Scores are calculated from the beginning of the word, such as 0 point, from the second character, -1 point, from the third character, -2 points, and so on. Then, "Ueota" is -1 point, and the other points are 0 points. In the remaining character score (d), the number of characters other than the key characters included in the candidate word is counted. If N characters are included, -N is counted. As shown in the figure, "Mayumi-cho, Hitachi-Ota City" has -2 points, and "Hitachi-Ota City"
"Ota City Ota""OtaCity""Kamiota""MayumiMinami" is -1
The point "Mayumi" becomes 0 point. Finally, these total scores (e) are given priority in order of (number of key character matches (a)> key character corresponding hierarchy (b)> key character match position (c)> number of remaining characters (d)) Here, (e) = 1000 ×
(A) + 100 × (b) + 10 × (c) + (d). The weight given to each evaluation value may be increased as the evaluation value has a higher priority. Further, in order to make the priority absolute, a value larger than the maximum value of the evaluation value with a low priority may be weighted (for example, if the maximum of the evaluation value (d) of the number of remaining characters is 0, The evaluation value (c) based on the key character matching position is 1
That's fine. ). The priority is determined based on the evaluation value thus determined, and a search result is obtained by referring to the word dictionary a6.

By using the technique of decomposing into character transitions and searching as described above, key input characters are partially missing or partially erroneous, or the order of key characters is ambiguous. Even if there is, search is possible. Here are some examples.

In FIG. 26, key characters are input in reverse order of "Mayumi-Ota" to search for the same place name "Mayumi-cho, Hitachi-Ota-city, Ibaraki-ken" as searched in FIG. However, when the key characters are decomposed into character transitions, "Mayumi (Hitachi Ota City)", "Mayumi (Ikoma City)", and "Mayumi Minami" are obtained as words including "Mayumi", as in FIG. No word contains "Yuta", but as a word containing "Ota"
"Hitachi Ota City""OtaCity""Ota City Town""Ota" is obtained. After that, as in FIG. 25, words connected in a hierarchy are connected to “Mayumi Hitachi-Ota-shi” and “Ota-shi Ota”, and if the score is calculated in the same manner, the first place “Mayumi-cho, Hitachi-Ota-shi, Ibaraki” is obtained. Even if the order of the key characters is reversed, as in this example, a valid result can be obtained by using character transition (two-character set) in the pointer (index) to the word dictionary. (It can also be seen from the fact that there are many two-character idioms). If this is
Assuming that each character is indexed, it is easy to imagine that if you search without regard to order, unexpected and unexpected search results will be included (for example, if you enter “Yamagata”, Is included).

FIG. 27 shows a case where "Mayumi-cho, Hitachi-Ota-shi, Ibaraki-ken" is to be searched, and "Mayumi, Hitachi-Ota-shi" should be input, and "Mayumi, Hitachi-Ota-shi" and "Ota"-"Ota" are entered. This is an example of incorrect input. In terms of words, "Hitachi Ota City" and "Hitachi Ota City" are different, but if key input characters are decomposed into character transitions, "Hitachi""Rikuta""Ota""Taichi""Ichima"" Mayumi, and from Hitachi and Taichi,
"Hitachi-Ota City" can be included in the candidate. From the word dictionary a6, words connected in a hierarchy are connected, and if the score is calculated in the same manner as described above, the character transition of “Mayumi Hitachi-Ota-shi” becomes the character transition of the key character “Mayumi Hitachi-Ota-shi”. Daejeon "," Taichi "," Ichima "," Mayumi "has the highest number of matches,
"Mayumicho, Hitachiota City, Ibaraki Prefecture" is ranked first. In this way, even if there is an error in a part of the key input character or the like, a desired character string can be searched using the remaining characters as a key.

By the way, in FIGS. 25 to 27 and the like, the number of search results of the fuzzy search increases because the number of key input characters has increased. Sometimes it's nice to see all of the fuzzy search results, but it's hard to see them all, and sometimes you want to narrow down the candidates because you've put a lot of key letters. In this case, as shown in FIG. 28, if the user can specify the display condition of the search result, the user can select necessary information according to his / her preference. In the example of the figure, it is specified that only the search results including 80% or more of the key input characters are displayed. Therefore, the search system calculates what percentage of the key characters are included in the address candidates obtained by the search, and displays only the search results that meet the conditions specified by the user.

In the example of FIG. 28, the user is caused to select how much the result of the fuzzy search is to be output, based on an index such as a matching ratio with the key input character. In addition, the user may be allowed to select the output of the result by, for example, (displaying a result that can be completely ANDed with a key input / displaying an OR result of the key input).

Further, as shown in FIG. 29, the user is allowed to specify how to prioritize fuzzy search results.
Information desired by the user may be preferentially displayed. In the example shown in the figure, since there is a designation to give priority in the order of (hierarchy> number of character matches), among the fuzzy search results including key characters, those with higher hierarchies are displayed with priority. [6] Fuzzy Search of One Character By now, the fuzzy search using the character transition information has been described. The character transition information is information indicating a relationship between two characters, and is generally effective for words having two or more characters. However, some words have one letter (eg, "Izumi", "I", "B", "C", etc.)
Therefore, it is necessary to extend the fuzzy search based on the character transition information to one-character words. Next, using character transition information, 1
A method for ambiguous search for words in letters is described.

FIG. 30 shows an outline of a transition information learning method and a retrieval method for retrieving even one character information based on the character transition information. The point of one character search is 1
The word W of a character is represented by a character transition of W → Null. For example, in the example in the figure, "Mine, Kawaguchi-shi, Saitama"
“Sai → Tama” “Tama → Prefecture” “River → Mouth” “Mouth → City” “Mine → Nul”
I learned it assuming the character line transition "l".

By learning in this way, for example, if a keyword for searching is entered by separating “Kawaguchi, Mine”, “Kawaguchi → Mouth” and “Mine → Null” (or
"Mine → * (anything)"
As shown in the figure, “Kawaguchi Ichimine” can be selected as the first candidate.
(While learning the word w of one character as “w → Null”, it is sometimes convenient to regard it as “w → * (anything is acceptable)” when searching. The details are described in 31.) It is sufficient if you enter the keyword of one character separated like “Kawaguchi, Mine” mentioned above, but it should be input without being separated like “Kawaguchi Mine”. Is also conceivable. In this case, "Kawaguchimine" is regarded as a two-character transition ("river → mouth", "mouth → mine") as in the past, and as shown in FIG. River → *, Mouth → *, Mine →
*)), It is possible to search for the one-character word “mine” without missing it.

By the way, the word w of one character is changed to “w → Nu”.
Although it has been described that it may be more convenient to regard “w → * (anything)” in the case of searching for a search after learning “ll”, such an example will be described with reference to FIG.

The example of FIG. 31 shows an example of learning and searching for an address "A, Asahi-shi, Chiba". At the time of learning, as in the example of FIG. 30, the character transitions “1000 → leaf” “leaf → prefecture” “Asahi → city”
Learn by breaking down into "I → Null".

Next, the search will be described. In the case of a place name such as "Asahi City" where the unique place name is one letter of "Asahi" and followed by "City" indicating the unit and becomes two letters, "Asahi" which represents the unique place name at the time of search Often you want to enter only one character. For example, this is a case where the user wants to search for "Asahi-shi, Chiba-ken", and inputs "Asahi-a". In such a case, if "Asahi, I" is regarded only as "Asahi → Null" and "I → Null", the word "I" can be searched because it has been learned as "I → Null". , "Asahi City" cannot be searched because "Asahi → City" has been learned. In order to prevent such a problem, when learning, the word w of one character is changed to the character transition “W
"→ Null", but when searching, it is better to regard the one-character word (or character) W as "W → * (anything is acceptable)". However, in this case, since all character transitions preceded by W are included, the number of unintended candidates is expected to increase rapidly. In such a case, a unique place name such as "Asahi City" has one letter at the beginning, followed by "prefecture / city / county / town / village" indicating the unit and has two letters. Should be given priority. To do this, first, "W → capital", "W → road", "W → prefecture" ... "W → village" are searched preferentially, and if there is room in the number of search results presented to the user, , "W → * (anything is fine)" may be executed.

Alternatively, first, "W → capital" and "W → road"
"W → fu" ... "W → village" is searched preferentially, the result is shown to the user, and if there is a desired search result, the search is terminated. If not, "W → * (anything is fine)"
There is also a way to execute.

Alternatively, in the user's search condition setting panel as shown in FIG. 29, “W → Tokyo”, “W → Road”, “W → Fu”
... It may be set whether to search for “W → village” preferentially or to search for “W → * (anything)”.

As described above, the ambiguous search using the character transition information can be extended so that the ambiguous search of one character can be performed.

[7] Learning of Word Dictionaries, Number of Character Transitions, and Number of Appearances FIG. 32 shows the configuration of a learning system for a word dictionary and a transition & appearance number table (character transition probability and appearance probability original data). An outline of the operation will be described. The user enters the word input unit h
1 inputs a new word. Then, the learning system additionally registers the newly input word in the word dictionary with learning function h3. If the newly input word has been registered before, the process proceeds to the next step (counting of transition and number of appearances) without additional registration. Subsequently, the newly input word is decomposed into character transitions, and the number of transitions and the number of appearances of the characters included in the word are recorded in the user transition & appearance number with learning function table h5.
Further, where the input word is recorded in the word dictionary with learning function h3, its address is recorded in the user transition with learning function & appearance count table h5 in association with the character transition. The above may be performed each time the user learns a new word. Then, using the drawings,
Word dictionary h3 with learning function and user transition with learning function &
The configuration of the appearance count table h5 will be described in detail.

FIG. 33 shows the structure of the word dictionary with learning function h3 and the user transition & appearance count table with learning function h5. In the example of this figure, information (phonebook information) of names and telephone numbers can be recorded in the word dictionary with learning function h3. In the telephone directory information table, a pointer pointing to the next telephone directory information is attached so that the information can be increased. In this figure, telephone directory information of three people, "Hiroshi Tanaka", "Keiji Kikuchi", and "Yukari Tadokoro" is registered.

FIG. 33 shows the word dictionary h with the learning function.
9 shows a state of a user transition & appearance count table h5 with a learning function for high-speed retrieval of data No. 3. In the user transition & appearance count table h5 with learning function, information on character transitions appearing in the word (transition source character and its appearance frequency, and transition destination character and character transition frequency of the character transition) and the word including the character transition ( (Relevant word). Here, too, a pointer to the character transition information of Next is attached, so that the information can be increased. Here, in order to realize a high-speed search when performing an ambiguous search for a word or an address later, data of a transition source character and data of a transition destination character connected to each transition source character are written as shown in the figure. It is better to arrange them in code order. Then, when performing a search later, the character transition information can be searched at high speed by a binary search or the like.

FIG. 34 shows that the word dictionary with learning function h3 shown in FIG.
41-5046) is added, and the user transition & appearance count table h5 with the learning function is updated. User transition & appearance count table h5 with learning function
Previously, character transition information such as "Gun → Tsuka" or "Kei → Child" was not registered. Therefore, new character transition information "Gun → Tsuka" and "Kei → Child" has been added. At the time of addition, as shown in the figure, the data of the transition source character and the data of the transition destination character connected to the transition source character are inserted at positions where the data are arranged in the order of the character codes, so that the retrieval can be performed quickly at a later time.

In FIG. 35, in addition to the example of FIG. 34, information “Hiroshi Kikuchi (telephone number 03-453-5587)” is newly added to the word dictionary with learning function h3. The state which updated the transition & appearance frequency table h5 is shown. By the way, the user transition & appearance frequency table h5 with the learning function has previously stored “Kiku → Ike” and “Haku → Nul”
The character transition information "l" was registered. Therefore, here, a word including the transition frequency of the transition source character, the transition frequency of the character transition, and the transition information (the related word) in the character transition information table of “Kiku → Ike” and “Haku → Null” which has been used before.
Can be rewritten. By using the above learning, when the user wants to add new word information, the data for the high-speed fuzzy search can be updated while utilizing the word dictionary and the transition & appearance frequency table which have been conventionally learned.

FIG. 36 shows an example in which the learning function of the word dictionary and the user transition & appearance count table is incorporated in the word fuzzy search device. In the manner described above, the word dictionary with learning function h
3 and learning of the user transition & appearance count table h5 with a learning function. Here, the point is that each time these tables are learned, the fuzzy search system loads the word dictionary with learning function h3 and the user transition & appearance count table with learning function h5, and searches using the latest information. It was made possible.

FIG. 37 shows an example in which the word dictionary and the user transition & appearance count table are divided into a standard dictionary (table) and a user dictionary (table). For example, generally used data is registered in a standard dictionary, and data unique to an individual is registered in a user dictionary. Then, even when a plurality of people share the system, the standard dictionary is shared, and each person may have data unique to the individual. Even when the standard dictionary is updated, by dividing the data unique to the individual into the user dictionary, the personal data is protected and new standard dictionary data can be used.

As described above, when a plurality of user dictionaries and standard dictionaries are used, for example, as shown in FIG.
Optimize the candidate character string and determine the word pointer, i5,
Check if the optimization result is valid i
6. If this is appropriate, the result can be derived by referring to the user's word dictionary h3. However, if it is not appropriate, it is considered that the user has input, as a search key, information that is not in the user transition & appearance frequency table h5 with learning function (information in the standard table). Therefore, a search similar to that described with reference to FIG. 1 may be performed using the standard appearance probability table d8, the standard character transition probability table a4, and the standard word dictionary a6.

If the character string optimization by the user transition & appearance frequency table h5 with the learning function is appropriate, the user transition &
A word search is performed by referring to the word dictionary address including the character transition described in the appearance count table i7.
If the search result is valid i8, display the result i
4. If it is not appropriate, the user uses the user transition with learning function & appearance count table h5 as a search key.
It is probable that you have entered information that is not in the information (information in the standard table). Therefore, the standard appearance probability table d8,
Standard character transition probability table a4, standard word dictionary a6
, A search similar to that described with reference to FIG. 1 may be performed.

[Various Applications of Fuzzy Search] [11] Examples of Application to Various Inputs (Keyboard / Character / Speech) Up to now, the optimal character recognition result has been obtained using a character transition information table and a word dictionary. We have focused on the example of categorization and fuzzy search. However, this method of optimizing a character string and performing an ambiguous search using a character transition information table and a word dictionary can be applied to voice recognition and the like in addition to character recognition. Can be used for characters input by various methods such as a keyboard.

FIG. 38 shows a diagram related to a search module when character string optimization and an ambiguous address search using a character transition information table and a word dictionary are applied to keyboard, character (character recognition input), and voice input. As shown in the figure, the character can be input by various means such as a keyboard, a character (character recognition input), and voice (in the example of the figure,
"Omika").

Characters and voices are converted into character codes by a recognition system such as character recognition or voice recognition. However, if a recognition system is used, the resulting character code (length or large)
Is a candidate for a correct answer, and has not been finalized. In order to make this uncertain candidate character as accurate as possible, the character transition probability information between the characters described in the character transition probability table a4 is used to make the candidate character likely to be a character string (in the example of FIG. The candidate has been corrected). After that, the data address of the candidate word including the character transition described in the character transition probability table used for optimizing the candidate character is referred to, and the word including the character transition is searched (directly referred). Then, from the word dictionary a6, using the hierarchical information of the address, the candidate of the address that the user wants to search is displayed (details are as described above).

If a key character (“Omika”) is input using a keyboard or the like that does not use a recognition system (if there is no input error by the user), the input information can be regarded as determined information. Therefore, for example, it is considered that recognition and optimization of candidate characters have been completed by input using characters or voice.
Therefore, the recognition and the candidate character optimization may be skipped, and the same process may be performed from the word search.

[12] Example of Using Ambiguous Address Search for Map Information Search, etc. FIG. 39 shows an example in which the ambiguous address search is applied to map information search. In this example, for example, when "Omika" is input by voice or the like, a map near "Omika town" is displayed, and can be used for car navigation operations and the like. The point here is that a pointer to the map data of the place represented by each place name word (corresponding map address information) is added to the word dictionary information (place name word and its hierarchical information). For example, when "Omika" is input, if a recognition system is required to encode a voice or character, a candidate character is obtained by recognizing the character and referencing the character transition probability table a4 to optimize the candidate. (If it is a keyboard, skip recognition and candidate optimization.) Furthermore, the address of a word related to the character transition described in the character transition probability table a4 is referred to. In addition to the word “Omika-cho”, the address (MA11) of the map near “Omika-cho” is described therein. Therefore, the user accesses the address MA11 of the map data j2 and displays a map near “Omika-cho”.

In the example of FIG. 39, an example is shown in which a map data pointer is added to a word dictionary representing a place name. The idea of setting pointers to various data in the word dictionary and storing various data together with words in the word dictionary as in this example can be applied in various other ways. For example,
If the postal code is added to the place name word dictionary, a system for inputting a part of the place name and examining the postal code can be performed. Also, if a telephone number is added to a word dictionary in which personal names are hierarchized at the time of a place name, a system in which a part of an address and a full name is ambiguously input and a telephone number is checked can be provided. Also, in the word dictionary,
By setting a pointer to the meaning information (data like a Japanese dictionary) of the word, it is possible to create a Japanese language dictionary system in which the input is vague and the correct notation (word) and its meaning are checked. In addition, a system that searches for books by library title, author, publisher, etc. A system that searches for medicines handled by doctors based on part of the medicine symbol or medicine name. Address / name information (word dictionary) is linked to family register information,
A system that searches for family register information from ambiguous input of names. It can be applied to search for all kinds of information, such as a system for searching for Internet keywords from ambiguous input.

[13] The word dictionary is stored in a plurality of notations to further expand the ambiguity of the fuzzy search. Next, the word dictionary k1 is stored in a plurality of notations (general notation, reading, English, German, Hiragana, Katakana, syllable, etc.) A character transition probability table is provided for every notation, and a method of extending a search ambiguity by setting a pointer from every notation to a word dictionary so that any of these search keys can be used.

FIG. 40 shows an example in which an address word dictionary is provided in a plurality of notations (general notation, reading, and English notation) so that mixed writing of kanji and kana characters and English input can be supported. . Here, the notation (such as kanji) and the reading corresponding to the notation are stored in correspondence with each notation. Character transition information is extracted from any of notation, reading, and English, and a pointer to a word dictionary is attached. Therefore, as shown in the example in the figure, even if the characters are input in a mixture of kanji and kana (“Ishinazaka”), it is possible to specify the word and display it in, for example, a correct notation.

In the example of FIG. 40, the search result is displayed in a general notation. If the word dictionary k1 has a plurality of notations, the result is displayed in a plurality of notations as shown in FIG. May be displayed in the input notation (language).

[14] Using a General Notation Table to Enable Searching from Different Notations Using only the notation as shown in FIG. 38 for the word dictionary, the vague search as shown in FIG. 40 can be used for different notations. As another method, for example, a character string input by mixed writing of kana characters and kana and kanji is converted into kanji, and then only the notation as shown in FIG. 38 is used for the word dictionary. When writing in a foreign language such as English, there is a method of translating and then using only the notation as shown in FIG. 38 for the word dictionary.

[15] Single Markov / Double Markov ...
・ Use N-fold Markovs In general, Kanji has many character types (about 3000 characters or more),
The number of words including a character transition (a combination of two characters) of a certain kanji is limited (in the case of an address, an average of about 7 words / character transition). However, there are few character types in Hiragana, Katakana, English, etc., and the word including the character transition (combination of two characters) is about 40 words / character transition on average even in the example of an address, which is very much compared to the case of Kanji. growing. In such a case, in a combination (transition) of two characters (single Markov), the word including the character cannot be narrowed down, and the amount of candidate word data addresses that must be written in the character transition probability table also increases. Therefore, in Hiragana, Katakana, English, etc., a combination (transition) of three characters (double Markov) or a combination of more characters may be used in the character transition probability table.

At this time, if a combination of three characters (double Markov) or the like is used for characters having many character types such as kanji, the character combination is simply 3,000 characters to the third power and the table is huge. become. Therefore, as shown in FIG. 42, for example, a kanji is stored in the character transition probability table e1 with information of a combination of two characters (single Markov), and a combination of three characters (double Markov) is used for hiragana, katakana, and alphabetical characters. It can be used properly depending on the character type, such as having it in a table.

[16] Method of Retrieving Even from Error Notation In characters, different notations with the same pronunciation ((no / no / no / no)
(Ga / ga / ke) (town / town) (Ota / Ota) etc.). For example, even in the same “Midorioka”, the characters used differ depending on the place name, such as “Midorigaoka, Yamagata City, Miyagi Prefecture” / “Midorigaoka, Sakata City, Yamagata Prefecture”, which is likely to cause confusion. Unless such characters are correctly input, a desired character string cannot be searched, which is inconvenient.

FIG. 43 shows one method for solving such a problem. Such an error-prone character is checked in advance, and an error-prone character set is recorded in the error-prone word table m1. If the search key character input by the user includes a word that is likely to be erroneous (“No” of “Sannomaru” in the example in the figure), the user may refer to the word table m1 that is susceptible to error, and Characters with a gender ("// no / no") are added to the candidate characters. Then, if a candidate character is optimized using the transition probability or an ambiguous search is performed using the transition information, a desired character string can be obtained as a search result. In such a case, the display of the search result is kind to the correct notation.

As another method of retrieving from an error notation, for example, there is a method using a table of a plurality of notations as shown in FIG. For example, in the same example as FIG.
Is correct, but you enter "Sannomaru". Then
Refer to the Kana-Kanji conversion dictionary, and read it and convert it to "Sannomaru". Then, if the word dictionary having the reading information in a plurality of notations is searched from the reading, it is possible to search for the correct notation “Sannomaru” with the same reading, which is different from the input character “Sannomaru”.

In order to search for a similar error notation in a word dictionary of general notations without using a table of a plurality of notations as in FIG. 40, an error notation ("Sannomaru") is obtained by using a kana-kanji conversion dictionary. ) Is converted to Yomi ("Sannomaru"), and from the Yomi conversion, Kana-Kanji conversion is performed to obtain the correct "Minomaru" for searching.

[16] Character Retrieval Using Character Transition Information Table Even in the case of performing an ambiguous character string search as described above, even if the input means is a keyboard or a character or voice is input, recognition is performed. In some cases, word processing or the like is already included in the system, and it is not necessary to optimize candidate character strings using the transition probability between characters. In this case, since a fine value of the transition probability is not used, a character transition information table in which only possible character transitions and data addresses of words including the character transitions are written as shown in FIG. 44 may be used. In this case, the search can be performed in the same manner as the method described above.

[17] Ambiguous Input Search Using Keyboard FIG. 45 shows an outline of an ambiguous input search system using a keyboard (a system that can search even if there is a key input error). The feature of this system is that even if there is an input error with a keyboard, etc., the key characters that are easy to make mistakes are added to the candidate characters, the candidates are optimized by the transition probability, and the fuzzy search is performed, thereby relieving the mistakes. , To obtain desired search results.

In the example shown, the user intends to enter "is", but miss-types the key next to "idi". However, the keys around the input characters "i" and "d" are checked from the key arrangement information of the keyboard, and are added to the candidate characters as candidates that may have been mistyped. Then, by optimizing the candidate based on the subsequent transition probability or performing an ambiguous search, a mistake can be relieved and a desired search result can be obtained.

[18] Ambiguous Search for Information Used Repeatably In the examples described so far, an example has been described in which the search target is only an address or only a name. Even for information having a plurality of items (for example, name list information), if character transitions are extracted from each information (for example, name, affiliation, address, etc.) and a similar transition probability table is created, fuzzy search can be performed.

In information having a plurality of items (for example, list information), the same word is frequently used. For example, considering the example of roster information, since multiple people exist in the same affiliation or one phone number is shared by multiple people, the same characters are displayed in the affiliation of multiple people and the telephone column. Appears repeatedly.

FIG. 46 shows a code for character transitions and words.
This is an example in which list data is provided as code information. In this example, character transition information and information (address) of a word including the character transition are stored in association with each other. The word is coded and stored as a sequence of character transitions, and furthermore, information (address) of name list data including the word.
Are stored in association with each other. The list information is stored as a sequence of the coded word information. When words are stored and coded in the character transition base in this way, the fuzzy search described above can be performed, and data in which the same word repeatedly appears, such as a list, can be compressed and held.

FIG. 47 shows the contents of the list data of FIG.
An example represented by word transition information is shown. Character transition information is stored in association with word information (address) containing the character transition. The words are coded and stored as a sequence of character transitions, and furthermore, information (addresses) (words) (words) of words arranged next to each word so as to have the same form as the word arrangement in the list information. Transition information between them) are stored in association with each other. Since character transition information is based,
In addition to performing the fuzzy search described above,
By using transition information between words, data including words becomes unnecessary.

According to the above-described method, it is possible to reduce the capacity of multi-item information such as a list while maintaining the function of fuzzy search.

[0105]

According to the present invention, a target character string can be searched by using not only the head and end characters of the character string to be searched but also the information of spelling as a key input, thereby greatly improving the operability. Further, as another effect, the operability can be greatly improved because the remaining character string is guessed and the entire character string is output as a recognition result only by inputting the key character by handwriting.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of a handwritten input address recognition device according to an embodiment of the present invention.

FIG. 2 is an operation schematic diagram of a handwritten input address recognition device according to an embodiment of the present invention.

FIG. 3 is a configuration diagram of a conventional character recognition device.

FIG. 4 is an operation schematic diagram when an address is input using a conventional character recognition device.

FIG. 5 is a hardware configuration diagram of a handwritten input address recognition device according to an embodiment of the present invention.

FIG. 6 is a flowchart illustrating a method of creating a character transition probability table.

FIG. 7 is a schematic diagram of a word dictionary.

FIG. 8 is a schematic diagram of a character appearance frequency table.

FIG. 9 is a schematic diagram of a character transition number table.

FIG. 10 is a schematic diagram of a character appearance probability table.

FIG. 11 is a schematic diagram of a character transition probability table.

FIG. 12 is a schematic diagram of a character recognition post-processing method performed using a transition probability table.

FIG. 13 is a schematic diagram for estimating a character string from a post-processing result.

FIG. 14 is a schematic diagram for estimating a character string from a post-processing result.

FIG. 15 is a schematic diagram for estimating a character string from a post-processing result.

FIG. 16 is a schematic diagram for estimating a character string from a post-processing result.

FIG. 17 is a diagram illustrating a display form of an estimated character string.

FIG. 18 is a diagram illustrating a display form of an estimated character string.

FIG. 19 is a diagram illustrating a display form of an estimated character string.

FIG. 20 is a diagram showing a display form of an estimated character string.

FIG. 21 is a schematic diagram for estimating a character string from a post-processing result.

FIG. 22 is a schematic diagram for estimating a character string from a post-processing result.

FIG. 23 is a schematic diagram of a method for determining a priority order of fuzzy search results.

FIG. 24 is a schematic diagram of a method for determining a priority order of ambiguous search results.

FIG. 25 is a schematic diagram of a method for determining a priority order of fuzzy search results.

FIG. 26 is a schematic diagram of a method for determining a priority order of fuzzy search results.

FIG. 27 is a schematic diagram of a method for determining a priority order of fuzzy search results.

FIG. 28 is a schematic diagram of a search result display condition setting panel.

FIG. 29 is a schematic diagram of a search result priority condition setting panel.

FIG. 30 is a schematic diagram of one-character search based on character transition.

FIG. 31 is a schematic diagram of one-character search based on character transition.

FIG. 32 is a schematic diagram of a learning device for a word dictionary and a transition & appearance count table.

FIG. 33: Word dictionary with learning function and transition & with learning function
It is an outline figure of an appearance number table.

FIG. 34: Word dictionary with learning function and transition & with learning function
It is an outline figure of an appearance number table.

FIG. 35: Word dictionary with learning function and transition & with learning function
It is an outline figure of an appearance number table.

FIG. 36 is a schematic diagram of a fuzzy search device for information with a learning function.

FIG. 37 is a schematic diagram of an information fuzzy search device having a user dictionary with a learning function.

FIG. 38 is a schematic diagram of an ambiguous search for character information using a keyboard, handwritten characters, and voice.

FIG. 39 is a schematic diagram of an ambiguous map search device.

FIG. 40 is a schematic diagram of an ambiguous search corresponding to mixed writing using a word dictionary of a plurality of notations.

FIG. 41 is a schematic diagram of an ambiguous search that displays a search result in a plurality of notations using a word dictionary in a plurality of notations.

FIG. 42 is a schematic diagram of a fuzzy search in which a search result is displayed in accordance with an input notation using a word dictionary having a plurality of notations.

FIG. 43 is a schematic diagram of an ambiguous search that displays search results in a plurality of notations using an error-prone table.

FIG. 44 is a schematic diagram of an ambiguous search using a character transition information table.

FIG. 45 is a schematic diagram of an ambiguous search that uses a character transition probability table to rescue a mistake in key input.

FIG. 46 is a schematic diagram of an ambiguous search device that can also compress information using word information based on character transitions.

FIG. 47 is a schematic diagram of an ambiguous search device that can also compress information based on word information based on character transitions and transition probabilities between words.

FIG. 48 is an explanatory diagram of priority order determination in the present invention.

[Explanation of symbols]

a1 tablet, a2 recognition dictionary, a3 character recognition unit, a4 character transition probability table, a5 candidate character optimization and word pointer determination unit, a6 word dictionary, a7 word collation and character string estimation unit, d1 a learning word readout unit, d2 a character appearance count unit, d3 an appearance count table, d4 a character transition count unit, d5 a candidate word storage unit including character transitions, d6 a transition count table, d7: character appearance probability table creation unit; d8: standard appearance probability table; d9: character transition probability table creation unit;
h1: word input unit, h2: word dictionary registration unit, h3: word dictionary with learning function, h4: transition & appearance number learning & word address associating unit, h5: user transition & appearance number table with learning function.

Claims (17)

[Claims] An information retrieval apparatus for retrieving a character group including an input character string, comprising: a dictionary for storing a plurality of character string groups in advance at respective unique addresses; A character transition information table in which transition information and an address of a character string group including a combination of characters in the transition relation among the plurality of character string groups are stored in association with each other; An information search device, wherein the dictionary is accessed by an address in which a corresponding character string group is described in the character transition information table, and a character string group in the dictionary corresponding to the address is displayed as a search result. 2. A dictionary according to claim 1, wherein said plurality of character string groups are stored in advance at respective unique addresses, and said plurality of character groups are stored in at least one or more kinds of notation methods (notation, hiragana, katakana, English characters, , Greek, syllable, different notation, etc.). 3. The information retrieval system according to claim 1, further comprising a language translation unit, wherein when the character string is input in a different language, the translation unit translates the character string into a language stored in the dictionary. An information retrieval device characterized by performing: 4. The kana-kanji conversion unit according to claim 1, further comprising a kana-kanji conversion unit, wherein if a description different from the description stored in the dictionary is input, the kana-kanji conversion unit converts the description into the description stored in the dictionary. An information search device characterized in that an information search is performed by using an information search. 5. An information retrieval apparatus according to claim 2, wherein the retrieval result is displayed in the same notation as the input character string. 6. The method according to claim 1, wherein the transition information from one character to another character is variable in the number of transition information to be used depending on the type of a character string group stored in the dictionary. Information retrieval device. 7. The method according to claim 2, further comprising a kana-kanji conversion unit, wherein the reading of the input character string is obtained by the kana-kanji conversion unit, and a character string whose reading matches the input character string is obtained as a search result. Information retrieval device. 8. A table according to claim 1, further comprising: a table in which a set of notations that are susceptible to errors in the input character string is provided in advance. An information search apparatus characterized by adding a set of notation that is easy to be mistaken to input characters and searching for a character string including the obtained candidate character string. 9. An information retrieval apparatus for retrieving a character group including an input character string, comprising: an information dictionary for storing a plurality of pieces of information in advance at respective unique addresses; a plurality of character string groups; A dictionary that stores addresses of the information related to groups at respective unique addresses; at least a transition information from one character to another character; and a combination of the characters in the transition relationship among the plurality of character string groups And a character transition information table storing the character string information corresponding to the input character strings, and accessing the dictionary by the addresses described in the character transition information table. And accessing the information related to the character string by using the address of the information related to the character string described in the address, and displaying the information. Information retrieval apparatus according to claim. 10. A method according to claim 1, wherein when there are a plurality of character strings including a part of the input characters, the number of characters matching the input characters in the character strings, The number of character transitions that match the character transition of the entered character, the position of the character string that contains the entered character, the frequency of use of the character group, the priority of the character group, the hierarchy of the character string, and the character At least of the alphabetical order of the column group, the character code order of the character string group, the order of strong relation with the information related to the character string group, and the conditions (address, telephone number, etc.) of the user using the search system. An information retrieval apparatus, comprising: means for determining a priority order based on one or more conditions. 11. An information retrieval apparatus according to claim 10, wherein a user can designate a priority condition. 12. An information retrieval apparatus according to claim 11, wherein the retrieval results having a high priority are narrowed down and displayed. 13. The information search device according to claim 1, wherein the number of search results is displayed. 14. The method according to claim 1, wherein at least a probability of transition from one character to another character and an address of a character string group including a combination of the characters in the transition relation among the plurality of character string groups are respectively set. The character transition probability table stored in association with the transition probability information stores the transition probability information separately for the number of appearances of the transition source and the number of transitions. An information search device characterized by learning by updating the number of transitions, the number of transitions, and the address of a word including character transitions. 15. The method according to claim 1, wherein at least a probability of transition from one character to another character and an address of a character string group including a combination of the characters in the transition relation among the plurality of character string groups are respectively defined. When a word is added to the dictionary, the character transition information table stored in association with the
A character transition information learning apparatus characterized in that an address of a word including a character transition is updated and learned. 16. A method according to claim 1, wherein when the character string is one character (w), transition information from one character to another character is regarded as transition information from the character w to the character Null. An information search apparatus characterized in that a character string can be optimized and searched even for characters. 17. An information retrieval apparatus for retrieving a character group including an input character string, comprising: a dictionary for storing a plurality of character string groups in advance at respective unique addresses; A character transition probability table in which a transition probability and an address of a character string group including the combination of characters having the transition relation among the plurality of character string groups are stored in association with each other; Means for selecting a key typed for inputting the character string and a key close to the key as candidates, and optimizing the candidate key using the character transition probability table. The dictionary is accessed by an address of a character string group corresponding to a candidate character string, and a character string group in the dictionary corresponding to the address is displayed as a recognition result. That the information retrieval apparatus.