JP2609196B2 - Similarity calculator - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to the field of information processing, and more particularly to the field of natural language processing such as machine translation, information retrieval, and question answering, image processing, and voice processing. The present invention relates to a device for calculating the similarity between each of the input information and other information (hereinafter referred to as “comparison information”) to be compared with the input information.

[0002]

2. Description of the Related Art Conventionally, in a similarity calculating apparatus for calculating a similarity between input information and comparison information, the calculation is performed using the input information and the comparison information itself as data.

[0003] For example, image retrieval will be described. The image search is a process of searching for an image similar to the input image from a large number of images prepared in advance as comparison information. In this case, when calculating the similarity between one input image and one comparison image, data of the color and shade of each dot of each image, that is, data of the image itself, has been used.

[0004] In contrast, there are cases in which features not appearing in input information are obtained from external knowledge and similarity calculation is performed. Hereinafter, a case of machine translation will be particularly described.

One type of machine translation system is to compare a large number of examples (a pair of an original sentence and a translated sentence) with an input sentence, find an example most similar to the input sentence, and generate a translated sentence based on the example. There is something. In such a type of machine translation system, the similarity between an input sentence, that is, input information, and an example, that is, comparison information, is determined by a word in the input sentence (hereinafter referred to as an “input word”) and a word in the example (hereinafter, “input word”). Is calculated using information acquired from external knowledge such as a thesaurus. The information obtained from external knowledge such as the thesaurus is information relating to features that do not appear in the input word or the comparison word, although they correspond to the input word or the comparison word, respectively.

[0006] As described above, a device that performs similarity calculation using information on features acquired from external knowledge has been conventionally used.
The calculation of the similarity between the input information and each of the comparison information has been performed sequentially.

FIG. 12 is a flowchart showing a procedure for calculating a similarity between input information and comparison information in such a conventional similarity calculation apparatus.

For example, when performing similarity calculation using an electronic computer, comparison features corresponding to the comparison information and not appearing in the comparison information are stored in advance in the memory of the computer. The similarity calculation is performed according to the following procedure.

Referring to FIG. 12, first, at step S001.
Then, a process of extracting input features corresponding to the input information and not appearing in the input information from the external knowledge is performed.

Subsequently, in steps S002 to S005, the comparison features previously stored in the memory of the computer are fetched one by one into a CPU (Central Processing Unit), and the similarity between the features of the input information is sequentially obtained. I was doing calculations.

[0011]

However, when the similarity is calculated using features that do not appear in the input information or the comparison information, the processing time is extremely long. There is a problem. For example,
In a machine translation system that compares the input sentence with a large number of examples, finds the most similar example, and generates a translation based on it, the number of examples should be increased if more accurate translation is to be performed. is required. However, when the similarity is calculated using tens of thousands of examples, it takes several minutes to calculate the similarity between the input sentence and all the examples even for a short sentence. Therefore, it has not been possible to realize real-time processing with such a type of machine translation system.

Therefore, an object of the present invention is to determine the similarity between the input information and each of the comparison information,
An object of the present invention is to provide a similarity calculation device capable of greatly improving the calculation speed in similarity calculation using an input feature corresponding to input information and a comparison feature corresponding to comparison information acquired from external knowledge.

[0013]

As shown in FIG. 8, the similarity calculating apparatus according to claim 1 calculates information similar to a plurality of pieces of predetermined comparison information. And means for storing external knowledge for deriving features corresponding to the given information but not appearing in the given information itself.
And, between each of the features of the comparison information derived using the external knowledge corresponding to the plurality of pieces of comparison information, and the features derived using the external knowledge corresponding to the input information. A predetermined similarity calculation is performed at each of the comparison information, a means 60 for deriving a similarity between the input information, and a means 38 for outputting the derived similarity. Means 6 for deriving the similarity
0 is a first area for storing a feature corresponding to each of the plurality of pieces of comparison information and a comparison result between the feature and a feature derived corresponding to the input information. And storage means 66 having a storage area 66 which can be referred to by storage contents. The storage means 66 stores, in the first and second areas, storage contents of a specified partial area which can be specified,
All of the comparison information that matches the given comparison target can be detected simultaneously, and the means 60 for deriving the similarity further includes a storage content of a corresponding first area of the plurality of pieces of comparison information. A predetermined portion of the feature derived in accordance with the input information simultaneously detects all the features that match the predetermined portion, and a specific portion corresponding to the predetermined portion of all the corresponding second regions. First match detecting means 62 and 64 for simultaneously storing a predetermined value indicating that a match has been detected in the area;
Means 6 for performing the detection and storage operations by 2, 64 a desired number of times while changing the designation of the predetermined portion.
2, 64 and the second area having a value that matches a given value corresponding to a certain similarity is simultaneously detected in the storage unit 66, and the second area corresponding to the detected second area is detected. By associating the comparison information with the certain similarity,
Second match detecting means 62 and 64 for deriving a similarity to the comparison information.

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

In the similarity calculation device according to the first aspect,
The similarity between each of the input information and each of the comparison information is calculated in parallel using the input features corresponding to the input information and the comparison features corresponding to the comparison information acquired from the external knowledge. Therefore, similarity calculation can be performed faster than in the conventional sequential calculation processing.

In order to realize the parallel operation, the similarity is calculated by using a storage means which can be referred to by the stored contents.
For each of the plurality of pieces of comparison information,
Is stored in the area. Then, all of the predetermined portions of the stored contents of the first area that coincide with the predetermined parts of the features derived corresponding to the input information are simultaneously detected, and the second area of the second area corresponding to the detected comparison information is detected. A predetermined value indicating that a match has been detected is simultaneously stored in a specific partial region corresponding to the predetermined portion. This detection and storage operation is performed a desired number of times while changing the designation of the predetermined portion, so that the degree to which each of the comparison information matches the predetermined portion is written in the corresponding second area. It is.
Then, a second area having a value corresponding to a given value corresponding to a certain similarity is simultaneously detected in the storage means, and comparison information corresponding to the detected second area is converted to the certain similarity. , A similarity to the comparison information is derived. The comparison performed on a plurality of pieces of comparison information depends not only on the number of pieces of comparison information, but only on the number of predetermined portions detected by the first coincidence detecting means. This number is much smaller than the number of comparison information. Therefore, the number of comparisons performed by the means for deriving the similarity described above is one for each piece of comparison information.
It is much less than when performing comparisons one by one.

[0024]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, three embodiments of the present invention will be described in order. Each of the three embodiments is an apparatus for calculating a similarity between one input information and 1000 pieces of comparison information. An example in which a word is used as information and similarity calculation is performed between an input word (hereinafter, referred to as “input word”) and 1000 words to be compared (hereinafter, referred to as “comparison word”) will be described. I do.

In the following description, the present invention will be described by taking the case of calculating the similarity between words as an example. However, the present invention is not limited to only the comparison of words, and the image information as described above is used. Also, the present invention can be applied to similarity calculation in audio information and the like.

Before describing the embodiments, the background where these embodiments are used will be described. A machine translation system of the type described above in which an input sentence is compared with a large number of examples (a pair of an original sentence and a translated sentence) to find the most similar example and generate a translated sentence based on the most similar example. In this type of machine translation system, the most similar example is obtained by calculating the similarity of each word constituting the input sentence and the example. Then, a translated sentence is generated based on the example.

The similarity between words is obtained as follows. First, a synonym dictionary as external knowledge is prepared. A “synonym code” is acquired from the synonym dictionary as information on features corresponding to the input word or the comparison word, and the similarity between words is obtained using the synonym code.

The thesaurus is a dictionary in which words are assigned a synonym code based on a system in which the meanings of words are hierarchically classified. FIG. 2 shows a part of a classification system of a thesaurus of four layers.

Referring to FIG. 2, the synonym code assigned to each word is composed of a three-digit decimal number. Synonym code 100
The 10th, 1st, and 1st places respectively represent the large, middle, and small classifications of the classification system. In the example shown in FIG. 2, the word “interview” belongs to a class whose major category is “transaction”, its middle category is “report”, and its minor category is “edit”, and its synonym code is 457. I have.

In the description of this embodiment, a three-digit decimal synonym code is used as shown in FIG.
However, the present invention can process various types of information, such as general n-digit m-ary numbers and vectors, instead of such synonym codes, as similar to the synonym codes.

(1) First Embodiment Referring to FIG. 3, a similarity calculation apparatus according to a first embodiment of the present invention includes an input unit 32, an input feature extraction unit 34, an external knowledge 24, It includes a similarity calculation unit 36 and an output unit 38.

The input unit 32 includes a keyboard, a character recognition device,
It consists of a voice recognition device and the like. The input unit 32 is for giving an input word to the input feature extraction unit 34.

The external knowledge 24 includes a hard disk and a memory. The external knowledge 24 stores a large number of data in which a heading of each word and its synonym code are paired as shown in FIG.

The input feature extraction unit 34 extracts a synonym code (hereinafter, referred to as “input code”) corresponding to the input word provided from the input unit 32 from the external knowledge 24 and provides the similarity code to the similarity calculation unit 36. It is.

The similarity calculator 36 includes 1000 computers 40-1 to 40-1000. The input of each of the computers 40-1 to 40-1000 is input feature extraction unit 3
4 is connected to the output. Each computer 40-
Outputs of 1 to 40-1000 are connected to inputs of the output unit 38. These computers 40-1 to 40-100
Each of the 0s includes a CPU and a memory. A memory of one computer stores a synonym code (hereinafter, referred to as a “comparison code”) corresponding to one comparison word and a program for calculating similarity. These computers 40-1 to 40-1000 are provided with input codes from the input feature extraction unit 34. Each computer 40-1
40-1000 calculates the similarity between the given input code and the comparison code stored in the memory of each computer according to a method described later, and provides the calculated similarity to the output unit 38. These computers 40-
An example of the comparison code stored in each of the memories 1 to 40-1000 is shown in FIG. In FIG. 5, for example, the comparison code “426” stored in the computer 40-1 corresponds to the comparison word “sales” as shown in FIG.
Is a synonym code corresponding to. Similarly, the other comparison codes are synonym codes corresponding to certain words in the external knowledge 24 shown in FIG.

The output unit 38 includes a display device, a printing device, and the like. The output unit 38 is for outputting the similarity given from the computers 40-1 to 40-1000 of the similarity calculation unit 36.

FIG. 6 shows the computers 40-1 to 40-4 in FIG.
It shows how to determine the similarity between the input code and the comparison code, which is performed in each of 0-1000. In FIG. 6, the leftmost column shows conditions that are satisfied between the input code and the comparison code. The middle column shows the similarity when the condition shown in the leftmost column is met. The rightmost column shows a pair of an input code and a comparison code that meet the conditions shown in the leftmost column.

The symbols "I1,""I2," and "I3" in the condition column for the input code and the comparison code represent the hundreds, tenths, and ones digits of the input code, respectively. The symbols "C1,""C2," and "C3" represent the hundreds, tens, and ones positions of the comparison code, respectively.

Referring to FIG. 6, as shown in the second line, the input code and the comparison code are changed to the hundreds digit and the tenth digit.
If both the first place and the first place match, the similarity is 3. As shown in the third row, when the hundreds digit matches the tenth digit and only the ones digit is different, the similarity is 2. When only the hundreds digit matches and the tens digit differs, the similarity becomes 1 as shown in the fourth row of FIG. As shown in the fifth line of FIG. 6, when the input code and the comparison code are different from each other in the order of 100, the similarity is 0.

Hereinafter, the operation of the similarity calculation device shown in FIG. 3 will be described by taking as an example the case where the input word is "reporting". When the input word “reporting” is input via the input unit 32, the input word “reporting” is given to the input feature extraction unit 34.

The input feature extraction unit 34 receives the input word “interview”
Is extracted from the external knowledge 24. In this case, as shown in the fourth line of FIG. 4, the synonym code corresponding to the input word “reporting” is 457, so that 457 is obtained as the input code.
This input code 457 is given to all of the computers 40-1 to 40-1000 of the similarity calculation unit 36.

Each of the computers 40-1 to 40-1000
Performs the similarity calculation independently of each other when the input code “457” is given. Each computer compares the input code “457” with the comparison code assigned to the computer, and obtains the similarity according to the method of obtaining the similarity shown in FIG.

For example, referring to FIG.
The comparison code stored in the memory 0-1 is “42”.
6 ". Therefore, the similarity is 1 according to the condition (c) in the fourth row of FIG. In the computer 40-2, the comparison code is “149”. Similarity 0 according to condition (d)
Becomes In the computer 40-3, the comparison code is "4
58 ". The similarity becomes 2 depending on the condition (b). In the computer 40-1000, the comparison code is "73
2 ". The similarity becomes 0 according to the condition (d). Similar similarity calculations are performed in the other computers 40-4 to 40-999. Each computer 4 of the similarity calculator 36
0-1 to 40-1000 give the obtained similarity to the output unit 38.

The output unit 38 includes computers 40-1 to 40-4.
When all the similarities from 0 to 1000 are given, the similarities are output to a display device or a printing device.

In the first embodiment, 1000 similarity calculations are performed in parallel by the computers 40-1 to 40-1000. Therefore, the simple calculation is performed about 1000 times faster than the calculation by the conventional sequential processing.

(2) Second Embodiment FIG. 7 is a schematic configuration diagram of a similarity calculation apparatus according to a second embodiment of the present invention. Referring to FIG. 7, this similarity calculation device includes:
An input / output management unit 50, a network 52 to which the input / output management unit 50 is connected, and 10 connected to the network 52.
And 00 similarity calculation units 54-1 to 54-1000.

The input / output management unit 50 includes an input unit 32, external knowledge 24, an input feature extraction unit 34, and an output unit 38. 7 and 3, the same blocks have the same reference numerals and names, and their functions are also the same. Therefore, detailed description thereof will not be repeated here. The input / output management unit 50 is configured by a workstation, a personal computer, or the like.

The external knowledge 24 stores the data shown in FIG. Similarity calculators 54-1 to 54-10
Each of 00 includes a workstation, a personal computer, and the like. Each of these similarity calculation units 54-1 to 5-5
4-1000 includes a CPU and a memory. The memory of one similarity calculation unit stores a synonym code (comparison code) corresponding to one comparison word and a program of similarity calculation.

The operation of the similarity calculation device according to the second embodiment of the present invention will be described below. An input word is provided to an input feature extraction unit 34 by an input unit 32 including a keyboard or the like. The input feature extraction unit 34 extracts an input code corresponding to the input word from the external knowledge 24, and
2 to all of the similarity calculation units 54-1 to 54-1000. Each similarity calculator 54-1 to 54-1000
Calculates the similarity between the transmitted input code and the comparison code stored in the memory of each similarity calculation unit in the same procedure as the method in the first embodiment. All the similarity calculation units 54-1 to 54-1000 send the obtained similarity to the input / output management unit 50 via the network 52.
The output unit 38 of the input / output management unit 50 includes a display device, a printing device, and the like as described above, and includes the similarity calculation units 54-1 to 5-5.
The similarity sent from 4-1000 is output.

Also in the second embodiment, the similarity calculation units 54-1 to 54-1000 independently calculate the similarity. That is, 1000 similarity calculations are performed in parallel. Therefore, similarity calculation is performed about 1000 times faster by simple calculation than the calculation by the conventional sequential processing.

(3) Third Embodiment FIG. 8 is a block diagram showing a schematic configuration of a similarity calculating apparatus according to a third embodiment of the present invention. Referring to FIG. 8, the similarity calculation device includes an input unit 32, external knowledge 24, an input feature extraction unit 34, a similarity calculation unit 60, and an output unit 38. 8 and 3, the same blocks are given the same reference numerals and names. Their functions are the same. Therefore, detailed description thereof will not be repeated here.

The similarity calculator 60 includes a CPU 62, a memory 64, and an associative memory 66 addressable by contents.

The memory 64 stores a similarity calculation program. Further, the memory 64 is provided with an area for storing the similarity corresponding to each piece of comparison information.

In each word of the associative memory 66, a synonym code (comparison code) corresponding to one comparison word is stored in advance. And 1000 for a total of 1000 words
Comparison codes are stored.

Associative memory 66 is assumed to have a matching search function using a mask and a partial parallel writing function. The match search function using a mask refers to a function of performing a match search on a specific portion of a bit string. What is the partial parallel writing function?
A function of writing data in parallel to a specific bit of a plurality of data for which a match is detected.

The data storage format of the associative memory 66 is shown in FIG.
Is shown in Referring to FIG. 9, associative memory 66 is provided with 1000 areas of comparison codes 1 to 1000. The area for each comparison code, for example, the area for comparison code 1, consists of a total of 15 bits. Of these, the first three bits are an area used for obtaining similarity. The latter 12 bits are an area for storing a comparison code.

The first bit of the three bits in the area for calculating the similarity indicates a match / mismatch of the 100th place between the input code and the comparison code, and the second bit indicates the input code and the comparison code. The third bit is a bit indicating a one-digit match / mismatch between the input code and the comparison code.

Each of the 12 bits indicating the comparison code is 4 bits.
It is divided into three areas of bits. As shown in FIG. 9, these three areas are composed of four bits representing the hundreds digit of the comparison code, four bits representing the tenth digit of the comparison code, and four bits representing the ones digit of the comparison code. is there.

The first three of the comparison code areas
It is assumed that the bits are set to an initial value, for example, all 0 prior to the similarity calculation.

Hereinafter, the operation of the similarity calculating apparatus according to the third embodiment will be described. When an input word is input by the input unit 32 including a keyboard, a character recognition device, a voice recognition device, and the like, the input word is given to the input feature extraction unit 34. The input feature extraction unit 34 extracts a synonym code (input code) corresponding to a given input word from the external knowledge 24. The data stored in the external knowledge 24 is the same as that shown in FIG.

The extracted input code is sent to the similarity calculating section 60.
Given to. The similarity calculation unit 60 calculates the similarity between the input code and the 1000 comparison codes as follows, and supplies the result to the output unit 38.

The similarity calculation section 60 performs the following similarity calculation using the associative memory 66. The definition of the similarity is the same as that described with reference to FIG. 6 in the first embodiment.

The operation of the similarity calculator 60 when "reporting" is input as an input word will be described below. A synonym code (input code) corresponding to the input word “interview” is shown in FIG.
"457" as shown in FIG. When the input code “457” is provided to the similarity calculating section 60, the similarity calculating section 60 operates as follows.

First, as shown in FIG. 11, an area having the same data structure as the storage area of the comparison code in the associative memory shown in FIG. 10 is provided in the memory 64 of FIG. This area consists of eight rows of areas 70, 72, 74, 7
6, 78, 80, 82 and 84. In the latter 12 bits of the area 70 on the first line, “45”
7 "is stored. The first three bits of this area 70 are all 0s. The area 70 on the first line is called an input search code.

In the area 72 on the second line, data obtained by masking other than the four bits representing the hundreds digit of the input search code is stored. Area 74 stores data obtained by masking bits other than 4 bits representing the tens place in the input search code. In the area 76, data obtained by masking other than 4 bits representing the first place in the input search code is stored. In the area 78, "111" is stored in the first three bits, and data in which the other bits are masked is stored. Area 80 contains the first three
“110” is stored in the bit, and data in which the other bits are masked is stored. In the area 82, data in which “10” is stored in the first two bits and other bits are masked is stored. In the area 84, data in which “0” is stored in the first bit and all other bits are masked is stored. Hereinafter, similarity calculation is performed according to the following procedure.

An instruction for a match search using data stored in area 72 and masking data other than the four bits representing the hundreds digit of the search code and a match code 100
And an instruction to write 1 to a bit indicating a match / mismatch of the order of order are given to the associative memory 66 in order.

An instruction for a match search based on data stored in the area 74 and masking data other than the 4 bits representing the tens digit of the search code, and adding 1 to the bit indicating the match / mismatch of the tens digit of the matched comparison code The instructions to be written are sequentially given to the associative memory 66.

An instruction for a match search based on data stored in the area 76 and masking data other than the four bits representing the one's place of the search code, and 1 to the bit indicating the match / mismatch of the first place of the matched comparison code The instructions to be written are sequentially given to the associative memory 66.

As a result of performing the above-mentioned processing (1), the data on the associative memory 66 is as shown in FIG.
That is, the comparison code 1 (426) and the input code “45”
Since only the hundreds digit matches "7", only the bit indicating match / mismatch of the hundreds digit becomes "1" as shown in the area of the comparison code 1 in FIG. 10, and the other two bits are "1". 0 ". Since there is no digit that matches the comparison code 2 (149) with the input code, the first 3
The bit becomes “000”. Similarly, the first three bits of the comparison code 3 are “110”.

A match search command based on the data stored in the area 78 of FIG. 11 is given to the associative memory 66, and a matched comparison code is detected. Then, among the areas for storing the similarities corresponding to the respective pieces of comparison information prepared in advance in the memory 64 of FIG. (Fig. 6
Condition (a)).

A match search command based on the data stored in the area 80 of FIG. 11 is given to the associative memory 66, and a matched comparison code is detected. Then, “2” is stored as the degree of similarity in the area on the memory 64 corresponding to the comparison code where the match is detected (condition (b) in FIG. 6).

A match search command based on the data stored in the area 82 of FIG. 11 is given to the associative memory 66, and a matched comparison code is detected. Then, “1” is stored as the degree of similarity in the area on the memory 64 corresponding to the detected comparison code of 1 (condition (c) in FIG. 6).

A match search command based on the data stored in the area 84 of FIG. 11 is given to the associative memory 66, and a matched comparison code is detected. “0” is given as the degree of similarity to the area on the memory 64 corresponding to the detected comparison code of “1”. (Condition (d) in FIG. 6).

The similarity between the input code and all the comparison codes is given to the output unit 38. The match search instruction and the write instruction performed in the above-mentioned steps are performed in parallel for all the comparison codes in the associative memory. Therefore, similar to the first embodiment, the simple calculation is about 1000 times faster than the similarity calculation by the conventional sequential processing.
Similarity calculation can be realized at high speed.

As described above, by using the similarity calculating apparatus according to the present invention, the similarity between the input information and the comparison information can be obtained from the input feature and the comparison information corresponding to the input information, obtained from the external knowledge. It can be obtained by similarity calculation performed in parallel using the corresponding comparison features. In particular, a machine translation system that compares an input sentence with a large number of examples (a pair of an original sentence and a translated sentence) to find the most similar example and generates a translated sentence based on the same, or a large amount of information and input information If it is necessary to find similarity between images using a feature corresponding to the input information in an information search such as image search or sentence search, it is possible to perform such processing at high speed. Become.

[0076]

As described above, according to the similarity calculating apparatus according to the first aspect, each of the features of the comparison information derived using external knowledge corresponding to a plurality of pieces of comparison information and the input information A predetermined similarity calculation is performed in parallel with a feature derived corresponding to the above to derive a similarity. In particular, by using a storage means that can be referred to by the stored contents, the comparison between the features of all the comparison information and the features of the input information can be performed in a single operation. By performing this operation much less than the number of pieces of comparison information, information indicating the degree of similarity between the features of all pieces of comparison information and the features of input information can be obtained. Therefore, it is possible to provide a similarity calculation apparatus capable of realizing a great improvement in calculation speed unlike the conventional sequential processing in the similarity calculation calculated using the comparison features in this way.

[Brief description of the drawings]

FIG. 1 is a diagram showing a schematic configuration of a similarity calculation device according to the present invention.

FIG. 2 is a diagram showing a part of a classification system of a thesaurus.

FIG. 3 is a block diagram illustrating a schematic configuration of a similarity calculation device according to the first embodiment of this invention.

FIG. 4 is a schematic diagram showing contents of external knowledge.

FIG. 5 is a diagram showing a comparison code stored in a computer memory in a similarity calculator of the similarity calculator according to the first embodiment of the present invention.

FIG. 6 is a diagram showing how to determine the similarity between an input code and a comparison code according to the present invention.

FIG. 7 is a block diagram illustrating a schematic configuration of a similarity calculating apparatus according to a second embodiment of this invention.

FIG. 8 is a block diagram illustrating a schematic configuration of a similarity calculation device according to a third embodiment of the present invention.

FIG. 9 is a diagram showing a storage format of a comparison code in an associative memory in the similarity calculating apparatus according to the third embodiment of the present invention.

FIG. 10 is a diagram showing comparison code data on an associative memory according to a third embodiment of the present invention.

FIG. 11 is a diagram showing search data on a memory used in similarity calculation performed by the similarity calculation device according to the third embodiment of the present invention.

FIG. 12 is a diagram showing a procedure of similarity calculation in a conventional similarity calculation device.

[Explanation of symbols]

 24 external knowledge 32 input unit 34 input feature extraction unit 36 similarity calculation unit 38 output unit 50 input / output management unit 52 network 60 similarity calculation unit 66 associative memory

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Jin Iida Kyoto, Soraku-gun, Seika-cho, 5th, Inaniya, 5th, Sanraya, ATIR Automatic Translation Telephone Research Institute, Inc. (56) References JP-A-2-2 235176 (JP, A) JP-A-2-297670 (JP, A) JP-A-58-205282 (JP, A)

Claims (1)

(57) [Claims] 1. A means for inputting information for calculating a similarity between a plurality of pieces of predetermined comparison information, and information corresponding to the given information, which appears in the given information itself. Means for storing external knowledge for deriving a feature that is not included, each of the features of the comparison information derived using the external knowledge corresponding to the plurality of pieces of comparison information, and Means for performing a predetermined similarity calculation between features correspondingly derived using the external knowledge and deriving a similarity between each of the comparison information and the input information; Means for outputting the calculated similarity, wherein the means for deriving the similarity includes, for each of the plurality of pieces of comparison information, a first area storing a corresponding feature; Corresponds to the input information The stored result of comparison between the feature derived on 2
And storage means that can be referred to by stored contents, wherein the storage means includes:
All of the comparison information in which the storage content of the specifiable predetermined partial area matches the given comparison target can be simultaneously detected, and the means for deriving the similarity further includes: , And simultaneously detect all the predetermined portions of the stored content of the corresponding first region that match the predetermined portion of the feature derived corresponding to the input information, and detect all the corresponding second regions. A first coincidence detecting means for simultaneously storing a predetermined value indicating that a coincidence is detected in a specific partial area corresponding to the predetermined part; and a detecting and storing operation by the first coincidence detecting means. ,
Means for performing a desired number of times while changing the designation of the predetermined portion; and simultaneously detecting in the storage means the second area having a value corresponding to a given value corresponding to a certain similarity And a second coincidence detecting means for deriving a similarity to the comparison information by associating the comparison information corresponding to the detected second area with the certain similarity. .