JP2002334318A - Method and system for similarity judgment using neural network and method and system for similitude formation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and a system for facilitating similarity judgment of a trademark, etc. SOLUTION: While features extracted from two objects to be compared each other are supplied to an input layer, an output layer is made to learn pieces of pattern information about similarity between the two objects to structure a neural network. Input pattern information based upon the features extracted from the two objects whose similarity is to be judged is inputted to the input layer of the structured neural network and the similarity between the two objects is judged from the output of the output layer.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention mainly relates to a novel method / system for directly outputting the similarity between two objects, and a novel method / system for creating an analog of a specific object.

[0002]

BACKGROUND OF THE INVENTION For example, regarding marks, especially trademarks under the Trademark Law, it is very difficult to objectively judge the similarity between two contrasting trademarks. The present invention has been made to reduce the difficulty in determining the similarity.

[0003]

SUMMARY OF THE INVENTION The method and system of the present invention are constructed as described in the above-mentioned claims, which are incorporated herein by reference.

[0004]

DESCRIPTION OF THE PREFERRED EMBODIMENTS For example, the following (1) to (22)
It is expressed as

(1) A feature extracted from two objects to be compared is given to an input layer, and a plurality of learning pattern information for giving a similarity between two objects to be compared is learned to an output layer. A network is constructed, and input layer pattern information based on features extracted from two objects to be compared is input to an input layer of the constructed neural network. A similarity determination method using a neural network, wherein similarity between two objects is determined.

(2) It is constructed by learning a plurality of learning pattern information comprising an input layer learning signal based on two contrasting objects and an output layer teacher signal based on the similarity between the two contrasting objects. An automatic similarity determination unit composed of a neural network, and input layer pattern information created based on two objects to be compared with each other, and the input layer pattern information is provided to the input layer of the automatic similarity determination unit. A similarity using a neural network, comprising: an extraction unit; and an output unit that outputs a similarity between two objects to be compared based on an output of the automatic similarity determination unit based on the input layer pattern information. Judgment system.

(3) A method of creating a similar object of a specific first object, wherein the input layer is provided with features extracted from two objects to be compared, and the output layer is provided with the corresponding two objects. A neural network is constructed by learning a plurality of pieces of learning pattern information that give similarities between two objects, and the first object and a plurality of second objects appropriately created are provided in an input layer of the constructed neural network. The input layer pattern information based on the objects is sequentially input, and the similarity between the first object and each of the second objects is obtained by the output of the output layer. A similar object creation method using a neural network, wherein an analogue object of the first object is created based on the first object object.

(4) A method for determining whether or not two character strings are similar to each other, wherein two character strings to be compared and a result of similarity determination between the two character strings are regarded as one pair. A plurality of pieces of information are collected, and an input layer is provided with an input layer learning signal based on the two character strings, while an output layer is provided with a similarity determination result to perform learning, thereby constructing a neural network. Input the input layer pattern information created based on the two character strings to be judged as similar to the input layer of the neural network,
A method for determining the similarity of character strings using a neural network, wherein the similarity between the two character strings is determined based on the output of the output layer.

(5) A system for determining whether or not two character strings are similar to each other, wherein two character strings to be compared and a result of similarity determination of the two character strings are regarded as one pair. A neural network constructed by collecting a plurality of pieces of information and providing an input layer with an input layer learning signal based on the two character strings to the input layer and giving an output layer a similarity determination result and learning the output layer. A character string similarity determination system using a neural network characterized by the following.

(6) A method of creating a character string similar to a specific first character string, wherein at least each different character of two character strings to be compared is given to an input layer, while an output layer is provided. A neural network is constructed by learning a plurality of pieces of learning pattern information that give similarity determination results of the two character strings to be compared with the first character string and one or more second character strings appropriately created. Based on the above, a plurality of input layer pattern information is created, the input layer pattern information is sequentially input to the input layer of the constructed neural network, the similarity is obtained by the output of the output layer, and a predetermined value or more is obtained. Generating a similar character string of the first character string on the basis of the second character string having the similarity of the first character string and the first character string. .

(7) A method for creating a character string similar to a specific first character string, wherein at least each different character of two similar character strings is given to an input layer and an output layer. A plurality of learning pattern information is learned to construct a neural network, and a plurality of input layer pattern information created based on the first character string are sequentially input to an input layer of the constructed neural network. Of the first character string,
A similar character string creating method using a neural network, wherein a similar character string of the first character string is created based on the output of the output layer.

(8) A method for determining whether two trademarks are similar to each other, wherein a unit for identifying characters in each trademark is provided in an input layer, and a similarity between the two trademarks is provided in an output layer. Or a unit for identifying dissimilarities is created, a plurality of pattern information is created based on the past cases of trademark examination of the JPO, etc., and the plurality of pattern information is learned to construct a neural network, and the constructed neural network is constructed. A trademark similarity determination method using a neural network, comprising inputting input layer pattern information created for a contrasting trademark to an input layer of a network and determining similarity of the trademark based on an output of an output layer.

(9) A method for judging whether two trademarks having one character difference are similar to each other, wherein the neural network includes the following items (a) to (d) in an input layer: The method for judging similarity of a trademark using the neural network according to the above (8), which has at least the item (a). (A) Characters different from each other in the two trademarks. (B) The word length of at least one of the two trademarks. (C) Positions of the different characters in the entire trademark. (D) Each character before and after the different character.

(10) A method for judging whether or not two trademarks having two different characters are similar to each other, wherein the neural network includes one of the following items (a) or (b) in an input layer. (8) The method for judging similarity of a trademark using the neural network according to the above (8). (A) Two different characters of the two trademarks. (B) Two characters in a case where two characters in one trademark are replaced with each other and become the other trademark.

(11) A method for judging whether or not two trademarks having one or more characters are similar to each other, wherein two or more consecutive characters in one trademark can be input,
For each character input unit, regarding the character in one trademark that is input by that unit, there is no corresponding character corresponding to that character in the other trademark, or the corresponding character corresponding to that character is One or more auxiliary units that identify the presence or absence of a difference with the corresponding part of the other trademark, such as being replaced with another character, are provided along with each of the auxiliary units. And the like. A method for judging the similarity of a trademark using a neural network according to the above (8), wherein a unit for inputting a trademark or the like is provided in the input layer.

(12) A method of judging whether or not two trademarks are similar to each other by referring to a similar patent decision example in the past by the JPO. Instead of similarity, a unit for identifying a past trademark decision example is provided, and the above (8) to (1)
A method for judging similarity of a trademark using the neural network according to any one of the above 1).

(13) A method of creating a character string similar to a specific first character string, wherein two character strings that differ by one character to be compared, and the similarity determination result of the two character strings are compared. A pair of pieces of information are collected, and the input layer is provided with the different characters of the two character strings and the characters before and after the different characters, and the output layer is provided with the two types of character strings. A neural network is constructed by learning a plurality of pieces of learning pattern information that give a determination result, and based on one or more character strings created from the first character string and a second character composed of Japanese syllabary characters and the like, A plurality of input layer pattern information is created, the input layer pattern information is sequentially input to an input layer of the constructed neural network, and a similarity is obtained by an output of an output layer.
A similar character string creating method using a neural network, wherein a similar character string of a first character string is created by using constituent characters of a character string and the second character having a similarity degree equal to or greater than a predetermined value.

(14) A method of creating a character string similar to a specific first character string, wherein a plurality of two character strings similar to each other and differing by one character are collected, and one of the two different character strings is input to an input layer. While giving a character and each character before and after the different character, the output layer learns a plurality of pieces of learning pattern information for giving the other different character to construct a neural network, and from the first character string, A plurality of input layer pattern information is created based on the created one or more character strings, the input layer pattern information is sequentially input to the input layer of the constructed neural network, and the output of the output layer is obtained. Generating a character string similar to the first character string by using constituent characters of the first character string and one or more second character strings specified by the output of the output layer. A similar character string creation method using a network.

(15) A trademark similarity determination system using a neural network, comprising an automatic similarity determination unit comprising the neural network according to any one of (8) to (12). .

(16) The above (6), (7), (13)
Or a similar character string creation system using a neural network, comprising a similar character string creation unit comprising the neural network according to any one of (14) and (14).

(17) An input device into which the trademark to be searched is input, and a trademark registered and / or pending in the Patent Office, which is related to the input trademark input from the input device. An automatic similarity determination unit, comprising a trademark database for searching for a trademark, and the neural network according to any one of the above (8) to (11), for determining similarity between the input trademark and the related trademark; An output device for outputting a result of the similarity determination, wherein a similarity determination system for a trademark using a neural network is provided.

(18) A trademark creation unit for creating an unregistered and unfiled candidate trademark based on the search result of the trademark database is provided, and the created candidate trademark and the searched related trademark are classified. No. is determined by the automatic similarity determination unit, and based on the result of the determination, it is possible to find a trademark that has not been registered and has not been filed and has a relatively high possibility of registration. A) a trademark similarity determination system using the neural network described in (1).

(19) An input device for inputting a trademark to be searched, and (6), (7), (13) or (14)
Consisting of the neural network according to any of the above,
A similar character string creating unit for creating a similar trademark similar to the input trademark input from the input device;
Or a trademark storing a trademark pending application, comprising a trademark database for searching for the presence or absence of the input trademark and the similar trademark, and an output device for outputting the search result. A trademark similarity judgment system using a network.

(20) The first device and the second device can be connected to each other so as to be capable of two-way communication with each other.
While comprising the trademark database, the second device comprises:
Comprising the input device and the output device, one or both of the automatic similarity determination unit and the similar character string creation unit,
The trademark similarity determination system using a neural network according to any one of the above (17) to (19), which is provided in either the first device or the second device.

(21) The similarity judgment processing by the automatic similarity judgment unit according to any one of the above (15), (17) and (18), or the similarity judgment processing according to the above (16) or (19) A computer-readable recording medium in which a program for causing a computer to execute a similar character string creation process by a character string creation unit is recorded.

(22) The similarity determination processing by the automatic similarity determination unit according to any one of (15), (17) and (18), or the similarity determination processing according to (16) or (19). A computer-readable recording medium in which a program for causing a computer to execute a process of creating input layer pattern information for a similar character string creating process by a character string creating unit is recorded.

The following is a more specific description. The similarity determination method and system and the similar object creation method and system of the present invention are configured using a neural network.

The network for similarity judgment gives a feature (feature portion, feature amount, etc.) extracted from two objects to be compared as an input layer learning signal, and outputs two compared objects as an output layer teacher signal. Similarity of objects (similar or dissimilar or how similar they are)
Is given and learned. Therefore, if input layer pattern information (equivalent to the input layer learning signal) based on features extracted from two objects to be compared is input to the input layer of the constructed network, the output of the output layer is output. The similarity can be determined based on the value.

As objects to be compared, in addition to characters and character strings represented by character trademarks, symbols and figures, voices, fragrances, and the like,
It can correspond to various things. As a feature in the case of a character string, all or a part (particularly, different character parts) of two character strings to be compared are extracted. In the case of symbols / graphics, for example, symbols read by a scanner or the like are arranged in a certain rectangular area, and the area is decomposed into a large number of areas in a grid pattern, and the amount of type printed in each area is calculated. Then, it is extracted by a character recognition method of obtaining an area ratio between the respective regions. In the case of voice, features such as a scale and a frequency are used, and are extracted by a voice recognition technique of, for example, digitizing a waveform of the voice and obtaining spectrum information at regular intervals. In the case of a scent, features are extracted by analysis of scent components and the like.

When the object to be compared is a character or a character string, specifically, for example, a trademark (mainly a character trademark in the Trademark Law, etc.), various keywords, and the like are the objects. The network for judging the similarity of a character trademark provides a part or all of the constituent characters of the two trademarks to be compared as an input layer learning signal and a similar or non-similarity of the two trademarks to be compared as an output layer teacher signal. It is constructed by giving similarity and learning. Therefore, in the input layer of this constructed network,
If input layer pattern information based on features extracted from two trademarks to be judged similar or not is input, the similarity can be determined based on the output value of the output layer. It should be noted that the learning pattern information for network construction can be collected focusing on past trademark trial cases of the JPO.

In order to input and identify characters on the network, the characters are digitized and input. For example, characters can be input and identified based on character codes. For example, a character code can be input and identified as binary data of binary numbers 0 and 1 in a set of units (unit set).

Alternatively, in order to identify characters in each trademark, a unit set consisting of a set of units such as alphabets and / or Japanese syllabaries can be provided and input and identified. In this case, among the units in the unit set, one of the numerical values such as 1 or 0 is applied to the corresponding unit, and the other numerical value not used is the other of the numerical values such as 1 or 0 for the unit that does not correspond. By inputting, characters can be identified.

Japanese characters are "A", "I",
Each stage of "U", "E", "O", "K", "Sa",
"Ta", "na", "ha", "ma", "ya", "la",
Each line of "wa" and sound-repellent "n", voiced point "゛", semi-voiced point "゜", prompting sound "tsu", characters for "m", "u",
"Yo", long sound "-", foreign words "a", "i",
Combining "e", "o", etc., one of the numerical values, such as 1 or 0, to the applicable one, and the numerical value of the other, such as 1 or 0, to the non-applicable, the other one not used in the above Can also be identified by inputting. Regardless of the cloudy point “゛” or semi-voiced point “゜”, “A”, “I”,
Each stage (vowel) of "U", "E", "O", "Ka",
"Sa", "ta", "na", "ha", "ma", "ya",
"La", "wa", "ga", "the", "da", "ba",
Each line (consonant) of “PA” may be used.

Here, as long as the other corresponding character that is different from each other does not become two characters, the characters with a gong, a character for a consonant sound, a character for a foreign language, and a character with a prolonged sound may be one character with them attached. it can. If the difference between the two trademarks is the presence / absence of a specific character in one trademark but not in the other trademark, the corresponding unit set of the pattern information of the trademark without the specific character includes all the units. The other number may be input to identify that fact, or a dedicated unit may be provided in the unit set and the one numeric value may be entered there to identify that fact.

On the other hand, among English characters, alphabetic characters "A" to "Z" or various phonetic symbols, etc., correspond to one of numerical values such as 1 or 0, and to those not applicable, 1 or 0. Of the numerical values of the above, identification can be performed by inputting the other number that has not been used. If the difference between the two trademarks is the presence or absence of a specific character or the like in one trademark but not in the other trademark, the corresponding unit set of the pattern information of the trademark without the specific character or the like includes:
The other number may be input to all of the units to identify that fact, or a dedicated unit may be provided in the unit set and the one numerical value may be entered there to identify that fact. Good.

The network for creating an analog gives a feature extracted from two objects to be compared as an input layer learning signal, and gives a similarity between the two objects to be compared as an output layer teacher signal. Built by learning. Therefore, the network for similarity judgment can be used. Then, input layer pattern information based on a specific first object that is the basis for creating similar objects and a plurality of second objects that are appropriately created are sequentially input to the input layer of the constructed network. Then, the similarity between the first object and each of the second objects is obtained from the output value of the output layer. Therefore, it is possible to create a similar object of the first object in consideration of the feature of the second object having a similarity equal to or more than the predetermined value and the first object.

In order to create a character string similar to a certain character string represented by a character trademark, at least each of different characters of the two character strings to be compared is given as an input layer learning signal, The signal is constructed by learning by giving similarity judgment results of two character strings to be compared. Then, in the input layer of the constructed network, the specific first character string that is the basis of the similar character string creation and the second character string appropriately created according to a predetermined rule or the like (including only one character) ), And input layer pattern information based on the above. Then, the similarity between the first character string and each of the second character strings is obtained from the output value of the output layer. Therefore,
A similar character string of the first character string can be created in consideration of the second character string having a degree of similarity equal to or greater than a predetermined value and the first character string.

As another method, at least different characters in one character string are given as an input layer learning signal, and different characters in the other character string are given as an output layer teacher signal. The network constructed by learning is used. Then, to the input layer of the constructed network, a plurality of input layer pattern information appropriately created according to a predetermined rule or the like from a specific first character string which is the basis of similar character string creation is sequentially input. Then, by each output of the output layer, a second character string (including the case of only one character) which gives an instruction to replace, insert, or delete the constituent characters of the first character string is obtained. Therefore, a character string similar to the first character string can be created in consideration of the constituent characters of the second character string and the first character string.

The similarity judgment method of a trademark and a similarity judgment system using the same are preferably used for judging whether or not a trademark to be filed in the future is similar to a registered or applied trademark. Is done. In addition, the similar character string (similar trademark) creation method and the similar character string creation system using the same are suitable for searching whether a trademark similar to the trademark to be filed is already registered or filed. Used. In other words, it is preferable to create a trademark similar to the trademark to be filed and to confirm in advance whether the similar trademark or the same trademark as the trademark to be filed has already been registered or filed. Used.

[0040]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The similarity judgment method and system using a neural network and the similar object creation method and system using the neural network according to the present invention will be described in further detail below. Here, for convenience, a character string, particularly a character trademark, will be described, but the present invention can be applied to other characters. First, a method and a system for judging the similarity of trademarks will be described. For this similarity determination method, a neural network is used.

As a network model, for example, there is a three-layer perceptron. As shown in FIG. 1, the input layer (S layer), the intermediate layer (A layer), and the output layer (R layer) are connected in series. Here, since the connection weight between each unit is random, an input signal is given to the S layer one after another,
If the connection weight is updated so that the pattern can be correctly identified from the difference between the output signal derived from the network and the ideal output signal, any pattern can be identified.

Now, let the output signal of the input layer unit i be I _i
Assuming that the connection weight between the input layer unit i and the intermediate layer unit j is W _ji , the output H _j of the intermediate layer unit j is derived as in the following equation. Note that n in the expression is the number of units in the input layer, and θ _j is the threshold value of the intermediate layer unit j.

[0043]

(Equation 1)

If the coupling load between the intermediate layer unit j and the output layer unit k is V _kj , the output O _k of the output layer unit _k is as follows. Note that m in the equation is the number of units in the intermediate layer, and γ _k is the threshold value of the output layer unit k.

[0045]

(Equation 2)

Here, f (x) is 1 when x is plus infinity.
This is a sigmoid function having an S-shaped response characteristic in which x asymptotically approaches x at minus infinity and approaches 0. By changing the slope (u> 0) of the sigmoid function, it is possible to contribute to creation of an optimal learning environment and construction of a model.

[0047]

(Equation 3)

In order for the neural model to derive a correct output signal with respect to the input signal, the connection weights and threshold values shown in [Equation 1] and [Equation 2] must be set to appropriate values. No. Therefore, a method of forming data (pattern information) in which an input / output signal (input layer learning signal and output layer teacher signal) forms one pair and learning the data to form a network is generally performed. . As a supervised learning method, for example, a back propagation method has been proposed. For details, see, for example, "Neuro Information Processing Technology-Fundamentals and Applications-" (by Kazuo Yana and Yoshitake Suzuki, Kaibundo Shuppan), "Introduction and Practice Neurocomputer"
(Nakano, Iinuma, Neuron Net Group, Kiritani, Technical Review).

When the change in the threshold value θ _j of the intermediate layer unit, the threshold value γ _k of the output layer unit, and the gradient u of the sigmoid function are not taken into account, the above [Equation 1] to [Equation 3] are as follows. Is shown in H _j = f (ΣW _ji · I _i ) O _k = f (ΣV _kj · H _j ) f (x) = 1 / (1 + exp (−x))

The network is constructed by learning using a large amount of data on the results of past trademark similarity judgments by the JPO and courts. For example, learning data is collected from a collection of trademark decisions of the JPO.

Input and output to and from the neuro, such as data input to the neuro learning program, are usually performed by normalizing to values of 1 to 0. That is, kana characters other than numerical values and the alphabet itself cannot be input. Therefore, kana characters and the like are usually quantified and input / output by appropriate means and identified.

In this embodiment, a set of units having alphabets, Japanese syllabary, etc. is prepared for each character in both trademarks in order to identify each character in both trademarks. That is, for the identification of each character, at least one set of units (unit sets) relating to individual characters such as Japanese syllabary is prepared for each trademark, and the corresponding character is identified by 0 or 1. Specifically, each of the above unit sets includes “A” to
It is composed of the 50-letter character of "n", voiced and semi-voiced voices, prompting sounds, characters for murmurs, characters for foreign words, long sounds (or characters with voiced dots, etc.). The unit is set to 1 and all other units are set to 0 to identify characters.

FIG. 2 shows an example of a unit set for inputting Japanese characters. Note that in FIG.
Although it is separated from the phonetic characters, it is usually considered to be included in the Japanese syllabary. Here, the character "@" can be omitted because it can be input as "o". Also, "ヂ" and "ヅ"
It can also be entered as "J" or "Z", and can be omitted. In addition, the characters enclosed in parentheses () in the column of resonations and foreign words are the Cabinet Notice No. 1 on December 9, 1954 (see the first edition of the Japanese Dictionary, pages 2159-11) and June 28, 1991. Date Cabinet Notification No. 2 (Kojien Fifth Edition No. 2955-2
(See page 961), which can be represented by characters without parentheses, and may be omitted in some cases. Characters not shown in FIG. 2, for example, "Sui", "Zi", "Guy", "Gu", "Guo", "Ke", "Nye", "Hye", "Fyo", "Vyo", etc. Can also be added. After all, it depends on the characters of the trademark data for pattern creation.

Thus, for example, in order to input the character of "S" in a trademark, only the unit of "S" is set to 1 in the unit set, and the remaining "A" to "S" and "S" to "N" are set. "
In addition, all other units, such as a prompt sound and a long sound, are set to 0, so that the character "S" is identified.

The unit of the corresponding character is set to 0,
The unit that does not correspond may be 1. Also, completely 0
Instead of a value of 1 or 1, a value of 0.1 or less (for example, 0.03) or a value of 0.9 or more (for example, 0.98) may be input. Further, depending on the system, an input in a range other than 0 to 1 is possible.

In the unit set shown in FIG. 2, an example for inputting kana characters has been described. However, in the case of kanji characters, the reading can be converted into kana characters and input. Can be entered. Alternatively, a network dedicated to English trademarks can be created as a unit set consisting of 26 alphabets (or hyphens, commas, periods, etc.) for inputting English characters. Alternatively, a network may be used in which alphabets and the like are added to the unit set in FIG.

Further, when constructing a network exclusively for English character trademarks, a unit set consisting of units of various pronunciation symbols published in English-Japanese dictionaries, English-English dictionaries, etc. is provided, and trademark characters are converted into pronunciation symbols. It is also possible to adopt a configuration in which the input is made after correction. In any case, in order to cope with the case where the difference between the two trademarks is the presence or absence of a specific character, a unit may be provided which is used when there is no space or a corresponding character. However, as described later, in the case of a trademark without a specific character, it can be dealt with by putting all 0s in the unit set.

By the way, if a unit set consisting of units of individual characters such as Japanese syllabary is prepared for each character in the trademark, the number of units will be increased. Therefore, in the case of Japanese syllabary, "A", "A", Each stage of "U", "E", "O" and "K", "Sa", "Ta", "Na", "C",
It is preferable that the input is made separately for each line of "MA", "YA", "LA", and "WA" (see FIGS. 5 to 8). In other words, for example, in the case of “G”, the unit is “K” row “I”, so 1 is inserted only in the units of “K” and “A”, and all the subsequent units can be input as 0. Is good. In addition,
For “A” to “O”, “A” rows may be prepared separately, but here, for “A”, only the unit of “A” is set to 1 and the remaining units are set to 0 (All units in the row are 0). "N" is identified by preparing a separate unit of "N" and putting 1 in only that unit.

Further, a character having a cloud point such as "ga"
In order to input a character with a semi-voiced point such as "P", prepare a unit with a voiced and semi-voiced point, and if it is "Ga", enter "1" only in each unit of "Ka" and "A" and a voiced point. Let them be identified. In the case of "P", "C" and "U" are distinguished by putting 1 only in each unit of the semi-voiced point.
In addition, the prompting sound "tsu", the characters for the sound "ya", "yu",
"Yo", "ヮ", long sound "-", foreign words "a",
In order to input characters with "i", "ゥ", "e", and "o", each of these units is also prepared. For example, in the case of "ti", "te" is replaced by "ta" line "E" stage,
Only 1 is entered in the three units of "ta", "e", and the foreign language character "i". In addition, in the case of "D", it is dealt with by adding 1 to the unit of the cloud point.

Incidentally, in some cases, "場合", "ヱ",
"ヲ", "V", "K", "K", and hyphen "-"
Units such as various symbols may be separately prepared. In addition, voiced and semi-voiced voices, prompting sounds, letters for murmurs, letters for foreign words,
Characters with a long sound or the like may be one character with them attached, or may be one character with them alone. In particular, a character other than a cloud point or a semi-voice point (especially a prompt sound or a long sound) can be converted into a single character by itself. Also, in the above,
In order to input characters having a clouded point such as "ga" and characters having a semi-voiced point such as "p", separately prepare units for a clouded point and a semi-voiced point, and for "ga", "ka" and "a" Although an example has been described in which the identification is made by putting 1 only in each unit of the cloud point, a unit of each character with a cloud point or the like may be provided separately. In other words, it is possible to provide a line unit with a cloud point and a semi-voice point such as "ga" and "pa" lines (row units of "ga", "the", "da", "ba", "pa"). You can also.

Note that units of relatively infrequently used characters, such as the foreign language character "@", may be omitted. That is, the units to be provided depend on the characters in the (learning) pattern information. In addition, it is also possible to change the character to a certain extent by changing “va” to “ba”, “vi” to “bi”, and “ヂ” to “ji”.

In the present embodiment, each character of the 50 sounds "A" to "N" is basically used as one character, and in principle, these characters are used for voiced characters, semi-voiced voices, prompting sounds, characters for murmurs, characters for foreign words, and prolonged sounds. It is determined that the character is one character in a state where the characters are attached. For example,
Of course, characters such as "B" and "Pi" with a cloud point or a semi-voice point,
Characters with a prompting sound, a character for repetitive sounds, a character for a foreign language, a long sound, and the like, such as “R”, “Kyo”, “Wo”, and “Ku”, are assumed to be one character with these characters attached. In addition, a character with two or more prompting sounds, long sounds, and the like, such as "ditch" and "tee", is, in principle, one character with all of them.

However, in the present invention, since two or more trademarks to be compared are considered, there is an exception in the case where these different sounds are attached to different parts. In other words, according to the above rule,
In one trademark, even if one character can be made with a long sound, etc., but in the other trademark, if the character corresponding to the character becomes two characters, only one sound, etc.
I decided to make it into a character.

For example, when comparing the two trademarks “Home Guard” and “Home Guide”, the differences are “−” of “Guard” and “A” of “Guide”. And
Observing only the "guard", it is possible to use "gar" as one character, but in such a case, if the corresponding character corresponding to "gar" is viewed with the other trademark, "guy" It becomes two characters of "Ga" and "I". Therefore, in this case, "gar" is divided into two characters "ga" and "-", and "-" and "i" are compared.

Also, in the comparison between "Earthrat" and "Earthlight", the "Rat" part of one trademark can be one character, but the other corresponding character is "L". In such a case, the character "ra" is divided into two characters "ra" and "tsu", and "tsu" is also regarded as one character. Will be compared with the corresponding letter "i" in the other trademark. In addition, similarly, in the case of the comparison between "Koyu" and "Koyu", "-" and "U" are compared as one character,
In the case of comparison between "Cartier" and "Carty",
"E" and "-" are compared as one character.

In the case of comparison between "corp" and "core", if "co" is one character, the corresponding character of the other trademark will be "core". Therefore, in this case, "ko" is divided into "ko" and "-", and the "-" is compared with the corresponding character "a" in the other trademark. On the other hand, when comparing "U-Star" and "Aster", the difference is "U" and "A". If there is no risk of becoming letters, do not disassemble.

Since the characters other than those different from each other are the same as the characters of the other trademark, there is no possibility that two characters will be added even if a long sound is added. It is one character in the state. That is, in principle, one character is added with a long sound, etc., and the exception rule is applied only to the vicinity of the different character. The judgment is as follows.
Here, for the sake of convenience, comparison of trademarks having a difference of one character is considered.

First, a difference between the two trademarks is found. (1) When the difference is a difference between specific characters (a) 50
In the case of differences between phonetic characters (sometimes accompanied by a prompt, etc.)
If the 50-letter character has a prompt or the like, it is compared as one character including those. For example, in the case of comparison between “thiochol” and “thymocol”, “o” and “mo” are different, so “o” and “mo” are extracted as comparison targets. Also, in the case of comparison between “Royen” and “Roy Fan”, “A” and “Fa” are extracted. Furthermore, in the case of comparing "chacco" and "sacco", "cha" and "sa" are different, but since each has "tsu",
“Chat” and “slash” are extracted as comparison targets.
Similarly, in the case of comparison between "Micro Tech" and "Micro Dick", "T" and "Dit" are extracted as comparison targets, and in the case of comparison between "Sync rate" and "Sinlator", , “G” and “tar” are extracted as comparison targets. Also, in the case of "FetchE" and "FetchB", "E" and "B" are extracted.
Furthermore, in the case of "Java" and "Java",
“Va” and “bar” are extracted.

(B) In the case where there is a difference between the Japanese syllabary character and the prompting sound, etc. The syllabary character and the prompting sound are compared as one character. For example, in the case of a comparison between “Hitter” and “Hatter”, “I” and “tsu” are different, so “i” and “tsu” are extracted as comparison targets. In the case of comparison between "Karoldo" and "Caroud", "Lu" and "-" are different, so "Lu" and "-" are extracted as comparison targets. If “R” is extracted from the other trademark, “Lol” is extracted from the one trademark as a corresponding character, resulting in two characters.

(C) In the case where there is a difference between prompting sounds and the like The comparison is made as one character including the 50-letter character preceding the prompting sound and the like. For example, in the case of a comparison between “Lelead” and “Lerid”, “−” and “T” are different, but “Lee” and “Litt” are extracted as comparison targets. In the case of the comparison between “chakko” and “chaco”, “cha” and “cha” are different, but “cha” and “cha” are extracted as comparison targets. Note that in the case of “chaco” and “chaco”, it may be considered that there is a difference in the presence or absence of the prompting sound “tsu”, and even in such a case, the result is the same by a method described later.

(2) In the case where the difference is in the presence or absence of the specific character (a) In the case of the presence or absence of the Japanese syllabary character From the other trademark without the specific character, a space is extracted as a comparison object in principle. At that time, if the syllabary character is accompanied by a prompt sound or the like in principle, it is one character including the syllabary sound. For example, when comparing "Joana" and "Jona", there is a difference in the presence or absence of the character "A".
“A” and “_” (blank) are extracted as comparison targets.
Also, in the case of comparing "George Sunk" and "George Sunk", there is a difference in the presence or absence of the character "Ju",
“Ju” and “_” (blank) are extracted as comparison targets. Furthermore, in the case of "Ichiko" and "Ichiko", "I"
And "_" (blank) are extracted, and in the case of comparison between "es-eskay" and "es-esque", there is a difference in the presence or absence of the character of "ke", so "ke" and "_" (blank) Is extracted as a comparison target. If there is no corresponding character (“_”), a dedicated unit for identifying the corresponding character may be provided in the unit set, and if there is no corresponding character, 1 may be inserted there.

In the case of “Ichiko” and “Ichiko”,
Of the two "I" in "Ichiko", either "A" can be a different character, but in this embodiment, it was considered that there was a leading "I". In this embodiment, as described above, one character is included as much as possible, such as a prompt sound or a long sound, and similarly, the purpose of extracting the contrast portion in a cohesive form as much as possible is maintained. . Therefore, in the case of "SSK" and "SSK", the common part of both is taken as much as possible,
"K" on the back side was extracted from "SSK". On the other hand, when comparing Sears and Seeds,
Although there is a difference in the presence or absence of "A", since "A" has a long sound "-", "A" is extracted, and "-" corresponding to "A" is extracted from the other trademark. .

(B) In the case of presence or absence of only a gong and a long sound One of the trademarks having such a gong and the like,
One character including the sound is extracted as one character, and from the other trademark without such a prompting sound, the corresponding 50 sounds are extracted as a comparison target. For example, when comparing "Nike" and "Nike",
Although there is a difference in the presence or absence of “tsu”, “ni” is extracted including the character of “ni” before “tsu”, and “ni” is extracted as the corresponding character from the other trademark. Also, "AG"
Also in the case of comparison between “G” and “Age”, although there is a difference in the presence or absence of “−”, “G” and “G” are extracted as comparison targets.

In this way, characters of both trademarks are input. For example, in a network capable of inputting up to three-character trademarks, six unit sets are prepared for inputting the three characters of both trademarks, and each character is input. In this case, it is preferable to input the characters corresponding to both trademarks to the corresponding unit set. That is, for example, when the first trademark is “Aiu” and the second trademark is “Au”, the first trademark is “A” or “I”.
"U" is put in order, but in the second trademark, there is no "I" as compared to the first trademark, so after "A" is inserted, all of the corresponding unit sets that input the next character are set to 0. For example, it is preferable to identify that there is no corresponding character (blank), and then to insert “U”. When considering this correspondence, the above-described difference extraction rule is applied.

By the way, a network capable of inputting up to the three-character trademark can handle not only a three-character trademark but also a trademark having other characters. In other words, the similarity of the trademark is mainly due to the difference character portion occupying a large weight. Therefore, by inputting the vicinity of the difference portion, it is possible to respond by inputting the main part of the trademark having four or more characters. For example, the similarity between “Kakikukeko” and “Kakekukeko” is determined by the above 3
If it is desired to make a determination using a character trademark network, the second character “K” in one trademark is different from the second character “K” in the other trademark. And "Kakeku". Conversely, a trademark having two characters or less can be handled by setting the remaining unit set to 0. Therefore, how many characters of the unit set are prepared is appropriately set.

When all characters are to be inserted, a unit set can be prepared twice as many as the number of characters of the trademark having the longest word length among the trademarks of the collected data. When mainly inputting the difference between the two trademarks, the number of characters constituting a general trademark (for example, 3 to 7 characters (a further number of characters is also possible (for example, 10 characters, etc.)) is 2). It is better to prepare a double unit set.

Next, it is not a matter of how many characters are used for trademarks.
Let us consider a system capable of judging similarity between trademarks having one character difference. That is, consider a network that inputs at least one character different from each other. However, in this case, even if the different characters are the same, the difference between the characters before and after the character or where the difference character is in the trademark, such as the beginning or end of the character, affects the similarity judgment of the trademark. It is thought to give. Therefore, it is preferable to input other items that are likely to affect the similarity determination, in addition to the different characters.

In this embodiment, as shown in FIG.
(1) In addition to the different characters in the two trademarks of trademark I and trademark II, (2) the word length of one trademark (for example, trademark I with the longer word length) in the two trademarks, ( 3) In the one trademark, the position of the different character in the entire trademark is set, and (4) in the one trademark, each character before and after the different character can be input. The items (2) to (4) are not limited to all items, and may be any one or more inputs. When both trademarks have the same number of characters, it is needless to say that either trademark may be a trademark having a longer word length.

By the way, if there is a difference at the beginning of the word, the character before the difference character is blank, if there is a difference at the end, the character after the difference character is blank, and if there is a difference in the presence or absence of the specific character, Difference characters in trademarks without specific characters shall be blank. The blank is identified by putting all 0s in each unit of the unit set. Also in this case, it is needless to say that a dedicated unit for identifying that there is no corresponding character may be handled.

Here, the input of each character of (1) and (4) is dealt with by inputting 1 or 0 to each unit of the unit set for identification of the Japanese syllabary, etc., and (2) is a pattern. Of the information, the number of characters of the trademark with the maximum number of characters is 1 and the number of characters of the trademark with the minimum number of characters is 0, and normalized and input to 0 to 1. (3) is 0 when the difference between the two trademarks is at the beginning of the word and 0 at the end of the word. A certain case was set as 1 and normalized to 0 to 1 and input.

The normalization is performed by the following equation. Q = (P−P _min ) / (P _max −P _min ) Here, the pattern information after normalization is Q, the numerical value of each item in the pattern information is P, the maximum value of each item is P _max ,
The minimum value was indicated by P _min . FIG. 4 shows an example of normalized data of the position of a different character with respect to a trademark having a main number of characters, and this normalization is performed by (number of characters from the head-1) /
(Word length of the trademark-1). The position of the difference character can be dealt with by inputting the number of the character from the beginning of the word.

For example, when comparing “Avon” and “Avon”, the above (1) becomes “I” and “−”,
(2) becomes 4, (3) becomes 0.333 (see FIG. 4) because the difference character is the second character from the beginning of the word, and (4) becomes "D" and "B". If the two trademarks are determined to be similar in a past decision, 1 is put in the output layer in this embodiment. In addition, in the case of comparison between “Suihou” and “Zuihou”, the above (1) becomes “S” and “Z”, (2) becomes 4, and (3) becomes 0 because the difference is the beginning of the word, and (4) Becomes "_" (blank) and "i". If the two trademarks are determined to be similar in the past decision, 1 is put in the output layer.

In the case of comparison between “Oaks” and “Oak”, the above (1) becomes “S” and “_” (blank),
(2) becomes 3, (3) becomes 1 because of the difference in the ending, and (4) becomes "-" and "_" (blank). And
If the two trademarks are determined to be dissimilar in the past decision, 0 is entered in the output layer. Further, in the case of comparison between "dim" and "gym", the above (1) becomes "di" and "ji", (2) becomes 2, and (3) becomes 0 because of the difference in the beginning, and (4) Becomes "_" (blank) and "mu". If the two trademarks have been determined to be dissimilar in the past decision, 0 is entered in the output layer. FIG.
8 to 8 show examples of the pattern information. However,
The word length and the position of the difference character from the beginning of the word are values before normalization.

As shown in FIG. 9, when judging the similarity of two character trademarks having two different characters, two different characters are input from the two trademarks. Again,
Other items that may affect the similarity judgment can be input. For example, (1) of the two trademarks, Trademark I and Trademark II, each having two different characters,
(2) the word length of one of the two trademarks (for example, trademark I having the longer word length); (3) the position of each of the different characters in the one trademark in the entire trademark; (4) In the one trademark, it is each character before and after each of the different characters. Then, one or more of these can be input, and in this embodiment,
A network is constructed by inputting three items (1) to (3).

Note that the one-character difference network is
In this embodiment, four characters can be input for the two character difference network. Therefore, both can be constructed by one network. In other words, for a single character difference, enter a different character and the characters before and after it, and for a two character difference, enter two different characters each. A unit for discriminating is provided separately, and the unit is discriminated by adding 0 when there is a difference between one character and inserting 1 when there is a difference between two characters.

In constructing the network, the results of a large number of past trademark decisions regarding whether or not two character trademarks are called or not are used, and the content of the decision is converted into pattern information according to the above rules, and the result is sent to the neural network learning program. Let them learn. As the specific learning method and learning program, various existing ones (for example, see the above-mentioned literature) can be used.
Here, a three-layer perceptron was considered, and learning was performed by a sequential correction method of back propagation. The number of intermediate layers, the number of units in the intermediate layer, and various parameters necessary for learning are appropriately set.

The learning for constructing the neural network is realized by, for example, a personal computer. That is, the personal computer is provided with a neural network learning means (program), a neural model construction means including a pattern information storage means storing a large number of pattern information necessary for learning, and the like, and a learning result is obtained. The coupling load and the like are stored in the coupling load accumulating means. In order to facilitate the creation of the pattern information, it is preferable to provide a normalization processing means (program) for automatically converting the pattern information only by entering the trademark (character) to be compared.

Each pattern information for learning for constructing a network for judging similarities between trademarks having different one-characters in FIG. 3 will be described. The input layer is composed of a word length of a trademark having a longer word length and a word length. The positions of the different characters in the longer trademark in the entire trademark are respectively normalized to 0 to 1, the different characters in the two trademarks, and the position of the longer trademark in the longer trademark. While placing either 0 or 1 in each unit of a total of four unit sets in which each character before and after the different character can be input,
The output layer indicates 0 if the two trademarks are determined to be dissimilar,
When it was determined that they were similar, one was arranged. By the way, in the present embodiment, the alphabet is not considered, and the alphabet is read in the case of the alphabet.

In this case, as described above, each unit set includes “A”, “A”, “U”, “E”, and “O”, and “K”, “S”, and “T”. ”,“ Na ”,“ c ”,
Each line of "ma", "ya", "la", "wa" and sound-repellent "n", voiced point "゛", semi-voiced point "゜", prompting sound "tsu", letter for relentless sound "ya", "yu" , "Yo", "ヮ", long sound "-",
Foreign language characters "a", "i", "ゥ", "e", "o"
Consists of a total of 28 units. Therefore, the number of units in the entire input layer of the network is 2 + 28 × 4 = 114
Thus, the number of units in the output layer is only one unit for similarity judgment. Note that, as described above, in each unit set, units such as the character for the letter “ヮ” and the character for the foreign language “ゥ” can be omitted or units of other items can be added.

Then, pattern information is created and learned from a large number of past decision cases, and the connection weight between units at the time when the error is appropriately reduced is stored, thereby constructing a neural network. For example, a connection weight or the like when the sum of the square errors between the expected output and the net output is equal to or less than a predetermined value is used. Alternatively, there is a method of using an evaluation pattern (pattern information other than a teacher) in a learning process and using a connection weight or the like at a time when an error with respect to the evaluation pattern is minimized. This is because, as shown in FIG. 10, when the number of times of learning is increased, the recognition of the teacher pattern is improved.
Recognition of other pattern information is effective to prevent over-learning, which is conversely reduced, and employs a connection weight or the like at the number of learning times A in the figure.

When it is desired to judge the similarity of the two trademarks, the input layer pattern information created from the two trademarks is input to the input layer of the network. Then, since the network outputs a value in the range of 0 to 1 to the output layer, the similarity between the two trademarks can be seen from the value. In the case of this embodiment, as described above, the dissimilarity is set to 0 and the similarity is set to 1 in the learning teacher data. Therefore, the closer the output layer value is to 1, the more similar the two trademarks are. I understand. In addition,
The program may be configured to automatically display any of “similar” or “dissimilar” from the value of the output layer. For example, if the value is 0 or more and less than 0.5, "dissimilar" may be displayed, and if the value is 0.5 or more and 1 or less, "similar" may be displayed.

Next, another example of a network for judging the similarity between two-character different trademarks will be considered. In this embodiment, two or more consecutive characters in one trademark can be input, and one or more auxiliary units are provided in each of the character input unit sets. This auxiliary unit is used for the character in one trademark input by the unit set attached to it, that the corresponding character corresponding to the character is not in the other trademark, or that the corresponding character corresponding to the character is in the other trademark. In the middle, it is replaced with other characters,
It is for identifying the presence or absence of a difference from the corresponding part in the other trademark.

In the network shown in FIG. 11, five consecutive characters can be input, and two first and second auxiliary units are provided in each unit set in order to cope with a difference of two characters. In FIG. 11, the characters constituting the trademark are shown as they are. However, as in the above-described embodiment, it is a matter of course that the neuron is constructed by converting the numerical values by appropriate means. That is, each constituent character is identified and input by the unit set. In addition, it is natural that other items such as the word length may be appropriately added (the same applies to each network hereinafter).

The input layer is further provided with a unit set corresponding to each of the auxiliary units, for indicating that there is no corresponding character or for inputting a different character or the like. In this embodiment, since the number of auxiliary units is two, two unit sets for character identification for inputting different characters and the like are provided in the input layer correspondingly. The output layer is provided with a unit for inputting the similarity determination result, as in the above embodiments.

In the case of this network, part or all of the trademark is input to the input layer by inputting five or less consecutive characters in one trademark so as to include one or more different characters. Then, add 1 to the auxiliary unit attached to the unit set containing different characters that are different from each other,
Enter 0 in other auxiliary units. In this case, use the first auxiliary unit for the first difference character,
The second auxiliary unit is used for the second different character.

A different character or the like in the other trademark corresponding to the character of the unit set in which 1 is put in the first auxiliary unit is input to the first unit set for the different character, and
A different character or the like in the other trademark corresponding to the character of the unit set in which 1 is set in the auxiliary unit is input to the second unit set for the different character. If one trademark has no corresponding character in the other trademark, all units in the unit set for different characters are set to 0, or a dedicated unit is included in the unit set. It responds by providing and putting 1 only there.

When there is a difference between two characters, two auxiliary units and a unit set for two different characters corresponding to the two auxiliary units are used. In this case, only one auxiliary unit (the first auxiliary unit in the illustrated pattern information) is used, and the second auxiliary unit and the corresponding unit set for inputting different characters are all set to 0. Corresponding.

In the illustrated example, the network is a network for inputting five consecutive characters. However, a trademark having other characters can be used. In the case of a trademark of 5 characters or less, it is possible to respond by putting all zeros in the remaining unit set. Can respond.

The number of characters that can be input can be appropriately increased or decreased by changing the number of unit sets for inputting characters, not limited to a network for inputting five consecutive characters.
In addition, an example is shown in which it is possible to deal with similarity judgment between trademarks having a difference of two characters or less, but by appropriately increasing or decreasing the number of auxiliary units and the corresponding unit set for inputting different characters, three or more characters can be used. It is also possible to respond to the similarity judgment of the trademark having the difference.

In the network shown in FIG. 11, a different character is input to a unit set at an arbitrary position. However, a unit set to which a different character is input may be aligned with a specific unit. The network of FIG. 12 is configured such that six consecutive characters in one trademark can be input, and different characters are input to the second and fifth unit sets. Then, two unit sets for different characters for inputting characters and the like that replace the second character and the fifth character are provided in the input layer.

In the case of this network, a trademark is input by arranging different characters at the second and fifth positions.
It is possible to cope with similarity judgment of differences up to two characters (up to two different characters). When two common characters are sandwiched between different characters, the respective common characters may be input to the third unit set and the fourth unit set. When only one common character is interposed between the different characters, the common character is put in one of the third and fourth unit sets,
What is necessary is just to input that there is no character in the other unit set. In addition, the fact that there is no character is dealt with by putting 0 in all units in the unit set or providing a dedicated unit and putting 1 in it. If there is no common character between the different characters and the different characters are located adjacently,
The fact that there is no character is input to each of the third and fourth unit sets. If there are three or more common characters between different characters, the common character adjacent to each different character is given priority,
Input to the third unit set and the fourth unit set,
Input of other common characters between adjacent common characters is omitted.

On the other hand, the first unit set receives the common character in front of the different characters input to the second unit set. If there is a different character at the beginning of the word, it is input that there is no preceding character. Similarly, the common character on the rear side of the different character input to the fifth unit set is input to the sixth unit set. If the difference character is at the end of the word, the fact that there is no character behind is input by the same method as described above. If there is a difference between the two letters, the second
And the fifth unit set are used, but it is possible to cope with similarities between trademarks that differ only by one character. In this case, only one of the unit sets (the second unit set in the pattern information in the illustrated example) is used, and the fifth unit set and the like are set to 0 to identify that there is no difference character. I do.

In the illustrated example, a unit set for inputting six consecutive characters is provided, but it is possible to cope with a trademark having a number of characters other than six as in the case of the network shown in FIG. The number of input characters can be appropriately increased or decreased by changing the number of unit sets, not limited to a network for inputting six consecutive characters. further,
An example was shown in which similarity judgments between trademarks having differences of two characters or less can be handled. However, by changing the number of unit sets for different characters, it is possible to judge the similarity of trademarks having three or more characters. Can also be handled. In addition, in this embodiment, the number of input characters between the different characters is two. However, this can be dealt with by increasing or decreasing the number of unit sets arranged between the different characters. For example, by adding one unit set for character input,
It can also be configured to be able to cope with similarity judgment of a trademark in which up to three characters are interposed between different characters.

In this embodiment, an example is shown in which two unit sets are arranged between the second and fifth unit sets for inputting different characters, and up to two common characters between different characters are input. The number of common characters between the different characters may be set to a predetermined number, and a network specialized for the number may be provided.

For example, as described above, in a network in which two characters can be inserted between different characters, only networks having two characters between different characters must be trained to determine whether or not such trademarks are similar. Network. In the same manner, the fourth unit set is omitted in the illustrated example, and a network dedicated to determining similarity between trademarks having one character between different characters is used as a network. By omitting the fourth unit set, a network dedicated to similarity determination between trademarks having different characters adjacent to each other (the number of common characters between the different characters is 0) can be provided. Further, by increasing the number of unit sets for inputting common characters in the illustrated example, a network dedicated to similarity determination between trademarks having three or more common characters between different characters can be provided.

Next, another embodiment of a network for judging the similarity of a two-character trademark is considered. In the case where two characters at an arbitrary position in one trademark are exchanged with each other, the other trademark can be used for similarity determination by constructing a network as shown in FIG.

In this network, by arranging a plurality of unit sets for character input in the input layer, two or more consecutive characters in one trademark can be input to the input layer. Auxiliary unit will be provided in addition. In the illustrated example, five consecutive characters can be input. Auxiliary units provided corresponding to each unit set for character input are filled with 1 only when the character input in the corresponding unit set is one of the replacement characters, and 0 in other cases. Can be

With such a configuration, all or a part of one trademark is input to the input layer in a form including two replacement characters, and 1 is added to each auxiliary unit for the replacement character.
Other characters can be filled with zeros. In this network, the output layer is similar (1) to dissimilar (0).
Another is given. Therefore, when it is desired to judge the similarity between two trademarks in which two characters at arbitrary positions in one trademark are exchanged with each other, the input layer pattern information created for the trademark to be compared is required. By inputting to the input layer, the similarity can be seen by the output of the output layer.

It is also possible to construct a network for professionally judging the similarity between trademarks in which two replacement characters are continuously arranged. In other words, it is also possible to construct a network for judging the similarity of two trademarks in the case where two adjacent characters in one trademark are replaced with each other to become the other trademark. In this case, as shown in FIG. 14, at least two different characters are put in the input layer, and the output layer outputs the similarity determination result. It should be noted that other items such as characters before and after a character string composed of two consecutive replacement characters may be input to the input layer, as in the other embodiments.

In each of the above embodiments, an example in which the result of directly judging the similarity of a trademark has been described. However, the similarity of a trademark can be indirectly determined by referring to a past trademark decision example. You can also build a network. In this case, the output layer is configured to specify the trademark trial decision example instead of the similarity determination result. FIG. 15 shows an example of the network configuration and pattern information in this case.

The neuro is configured to output a trial number to an output layer in order to specify a trial case. In other words, the characteristic part of the two trademarks to be compared is input to the input layer, and the referee number is input to the output layer to construct a network. As the configuration of the input layer, those of the networks of the above embodiments can be used. On the other hand, the judgment of the trial case in the output layer can be identified by, for example, providing a large number of units unique to each trial case and putting 1 in only the relevant unit.

In this case, when the input layer pattern information created for the two trademarks to be compared is input, a value from 0 to 1 is output to the unit of each trial case on the output layer.
By referring to the trial case of the unit that outputs the value closest to 1, it can be used to judge the similarity of the two trademarks. Of course, a plurality of trial cases may be referred to.

The output layer may be configured to output a binary pattern of 0 or 1 peculiar to the trial case. In this case, when the input layer pattern information relating to the trademark to be compared is entered in the input layer. The referee case having the pattern most similar to the output pattern of the output layer may be referred to.

Next, a method of creating a similar character string according to the present invention will be described. Here, a character trademark is considered as a character string. To create a similar trademark, the neural network described in each of the above embodiments can be used.

For example, creation of a similar character string using the network of FIG. 3, that is, a network in which the different characters of the two trademarks to be compared and the characters before and after the two characters are put in the input layer and the similarity is output to the output layer The method is described. Here, for convenience, a network in which the word length and the position of the different character are not considered is considered, but it is needless to say that these may be considered. In this case, a plurality of character strings are developed from a specific first character string that is the basis of the similar character string creation, and the second character (the second character ) In order to output respective similarities between them, and create a similar character string of the first character string using the second character whose similarity is equal to or greater than a predetermined value. . In addition, here, the second character is basically one character consisting of Japanese syllabary, but from the viewpoint of the network,
A character string with the characters before and after the character (the second character)
(Character string). This is because the network is configured to compare three characters of the different character and each of the common characters before and after the different character.

The creation of a similar character will be specifically described. For example, consider a method of creating a similar character string of a three-character trademark “guide”. First, a similar trademark in which each character of the basic trademark "guide" has been replaced with another character,
Consider the creation of a similar trademark with the other letters at the beginning and end of the basic trademark "guide". This is performed by arranging a portion to be replaced or added to a different character in the input layer, and putting characters before and after the part in the input layer.

That is, “__ga” (“_” for the first character, “_” for the different character, and “ga” for the second character) in the input layer.
, And the second character created appropriately is inserted in order. As the second character, 50 sounds of "A" to "N", each character of voiced sound and semi-voiced sound, prompting sound, long sound, relentless sound, characters for foreign words or characters with these are sequentially adopted. For example, the characters shown in FIG. 2 are adopted in order.

Then, each similarity between “__ga” and each second character is output, and among them, the second character whose similarity is equal to or more than a predetermined value is adopted. The predetermined value of the degree of similarity can be appropriately set, but is, for example, a numerical value of 0.5 or more, and a value of 0.7 to 0.8 or more is suitably adopted. As the adopted second character, it is preferable to adopt not only the second character having the maximum similarity but also a plurality of characters whose similarity is equal to or more than a predetermined value. If there is no similarity equal to or greater than the predetermined value, the similarity may not be adopted at all.

In this way, "? Guide" with another word added to the beginning of "guide" is created as a similar character string of "guide". Note that “?” Is the second character adopted by the above method. Also, by inserting “_ guy” and an appropriate second character in the input layer, a similar character string “? Id” in which “ga” in “guide” is replaced with another character is created.

Also, by inserting the “guide” and an appropriate second character into the input layer, a similar character string “gadido” in which the “a” of the “guide” is replaced by another character is created. Also, by inserting “id_” and an appropriate second character in the input layer, a similar character string “Guy?” In which “do” of “guide” is replaced with another character is created. Also, "de__"
And an appropriate second character in the input layer, "guide?" With another word at the end is created as a similar character string of "guide".

Next, consider the creation of a similar character string in which another character is added between the constituent characters of the basic trademark "guide". In this case, an inter-character portion to be added is arranged in a different character in the input layer, and characters before and after the character are entered in the input layer. First, a similar character string in which another character is added between “ga” and “i” in the “guide” by putting “ga_i” and an appropriate second character in the input layer. "Gha?
Id "is created. Also, by inserting “I_D” and an appropriate second character in the input layer, a similar character string “Guid?” In which another character is added between “I” and “D” in “Guide”. Is created.

Further, each character in which one of the constituent characters of the "guide" is deleted is created, and the similarity between each created character string and the "guide" is output to obtain a similarity of a predetermined value or more. It is also possible to leave only the character string. In other words, "guide"
Similarity between the character strings of "id" with "ga" removed, "Gado" with "i" of "guide" removed, and "guy" with "do" removed of "guide" and "guide" I look at you.

In the case of “guide” and “id”, the difference is in the presence or absence of “ga” (“ga” and “_”). What is necessary is just to look at the similarity.
In the case of “guide” and “gad”, the difference is in the presence or absence of “i” (“i” and “_”). You should look at.
Furthermore, in the case of "guide" and "guy", the difference is in the presence or absence of "do"("do" and "_"), so that "id_" and "_" are put in the input layer and similar Just look at the degree.
Then, only those having a predetermined similarity are left.

However, regarding the deletion of the constituent characters,
Since it is sufficient to judge the replacement of the constituent characters, it can be omitted. That is, in this example, "id" should have been adopted if the degree of similarity is high when comparing "_ guy" with the second character "_". "Gado" and "Guy"
Similarly, if the similarity is high by comparing the “guide” or “id_” with the second character “_”, it should be adopted.

In this way, a similar character string of "guide" is created. When searching these from the trademark database, not only the "guide" itself and the similar character string but also the same character string are used. It may be configured to search for a trademark that includes them. When searching for a trademark that includes the word, it is possible to omit the above-described determination of the addition of the beginning or end of the trademark.

In the above description, the creation of a similar trademark having a difference of one character has been described. However, a trademark having a difference of two or more characters can be created using a plurality of second characters adopted as described above. For example, “?

Although the three-letter trademark “guide” has been described here, the similar technique can be applied to the creation of a similar trademark of a trademark having other characters. For example,
If you want to create a similar trademark of the four-character trademark "Segaste", "__se", "_sega", "segas", "gaste", "ste_", "te__", "se_ga", "Ga_
And the similarity between each character string of "su_te" and an appropriate second character, and the second character having the predetermined similarity,
"? Segaste", "? Gaste", "Séste", "Sega? Te", "Segas?", "Segaste?"
Can be created as a similar character string. In addition,
The "?" Part may mean deletion in the same manner as in the example of the above "guide". For example, “Sesute” may include “Sesute”.

Note that a similar character string can be created using not only the network of FIG. 3 but also another network. For example, a similar trademark having two different characters can be created using the network of FIG. In that case, 50
A combination of two second characters such as sounds will be used.
Then, the similar trademark having a difference of two characters and the similar trademark having a difference of one character created as described above can be regarded as similar trademarks.

Another embodiment of a similar trademark (similar character string) creation method will be described. In this embodiment, a network created based only on trial decisions determined to be similar to each other is used. That is, a network is used in which the input layer is provided with the features of one trademark which is considered to be similar to each other, and the output layer is provided with the features of the other trademark which are considered to be similar to each other.

For example, the network shown in FIG. 16 is constructed on the basis of one-character difference trademarks which are considered to be similar to each other. A difference character of one trademark and each character before and after the trademark are put in an input layer and output. The layers are constructed with the different letters of the other trademark given. Therefore, the character string “A
When "BC" is inserted, a similar character string "AB'C" is obtained. Note that the actual output of the output layer is "B '".
Needless to say, only ("A" and "C" are characters common to both trademarks).

Now, consider a case where a similar trademark of, for example, “guide” is created in the same manner as described above. In this case, “__ga” (“_” for the first character, “_” for the different character, and “ga” for the second character) are sequentially entered in the input layer. Then, the second character is adopted from the output of each unit in the output layer. In order to make it easy to specify the second character, rather than putting the Japanese syllabary into columns and lines, the output layer is set to, for example, FIG.
It is good to constitute by the unit of each character of. A plurality of second characters may or may not be adopted as in the above embodiment. In this way,
"? Guide" with another letter at the beginning of "Guide"
It is created as a similar character string of "guide".

Next, by inserting “_ guy” into the input layer, a similar character string “? Id” in which “ga” of “guide” is replaced by another character is created. Also, by inserting the “guide” into the input layer, a similar character string “gadido” in which “a” of the “guide” is replaced with another character is created. Also, by putting "id_" in the input layer,
A similar character string “Guy?” In which “C” of “Guide” is replaced with another character is created.

[0133] By putting "do__" in the input layer, "guide?" With another word at the end is created as a similar character string. Further, by inputting "G_a" into the input layer, a similar character string "Guide" in which another character is added between "G" and "A" of the "Guide" is created. Also, by putting "I_do" in the input layer,
A similar character string "Guide" in which another character is added between "A" and "D" of "Guide" is created. It is to be noted that each “?” Part may mean deletion in the same manner as in the above embodiments.

Next, a system for judging similarity of a trademark using the neural network of the present invention will be described. This system is mainly provided with an automatic similarity determination unit composed of a neural network constructed as described above. FIG. 17 and FIG. 18 are schematic diagrams showing one embodiment of the similarity judgment system of the present invention. The system of this embodiment is a system for assisting in judging the similarity between a trademark registered or applied to the Patent Office and a trademark to be used or filed in the present or future.

The present system attempts to determine the similarity between an automatic similarity determination unit that determines the similarity of a trademark using a neuro, and a trademark database 5 that stores a trademark registered or applied to the JPO. An input device including a keyboard 2 and the like for inputting a trademark, an output device including a display 3 and a printer 4 for outputting a database search result, a similarity determination result by an automatic similarity determination unit, and the like are provided. .

A system without the trademark database 5 can also be used. In this case, the two trademarks input from the input device 2 are converted into input layer pattern information by the normalization processing means of the feature extraction unit. Based on the information, the automatic similarity determination unit performs similarity determination of the trademark and outputs the result to the output device (FIG. 19 shows a processing flow in this case).

In the trademark database, for example, the trademark itself, one or more names thereof, the application number, the registration number, and the like are registered and stored together with the classification and the similar group. The automatic similarity determination unit is configured by the neural network constructed for the similarity determination method of the present invention as described above.

The automatic similarity determination section is realized by, for example, the personal computer 1. That is, the personal computer is provided with a neuro execution means (program) for constructing a neuro using the connection weights and the like obtained by learning and for outputting an output to the output layer based on the input layer pattern information. The personal computer 1 can be connected to a server or the like having the trademark database 5 via a communication line or the like. The trademark database 5 can be realized by a medium such as a CD-ROM. In this case, a CD-R storing a registered trademark or the like is used.
The OM or the like is inserted into a CD-ROM drive or the like of a personal computer and activated to be used.

Then, a predetermined program is started on the personal computer 1 and the trademark to be investigated (input trademark) is input from the input device 2.
Enter Then, the personal computer 1 is connected to the trademark database 5 and, by the data processing means, the same trademark as the input trademark, one character difference from the input trademark (including addition or deletion, and two characters difference). And related trademarks such as those that partially include them. If there is the same trademark, the fact is output to the output device.

[0140] The retrieved related trademark and the input trademark input from the input device 2 are combined with each other by extracting a characteristic portion suitable for a network to be used, such as a different portion between the two trademarks, in a feature extracting section. It is converted into input layer pattern information using normalization processing means. The automatic similarity determination unit performs similarity determination processing from the input layer pattern information (obtains an output of the output layer), and outputs the result to the output device 3. FIG. 20 shows the flow of a series of these processes.

An automatic similarity determination unit and a feature extraction unit for performing automatic similarity determination processing are provided on the trademark database side (server or the like), and transmit a trademark input from an input device via a communication line to search for a related trademark. And the similarity determination process can be performed outside the personal computer. In this case, the personal computer performs input / output control (such as input of a trademark) and communication control. An example of the system configuration in this case is shown in FIG.

By the way, in case the related trademark includes a trademark which is the same as or similar to the input trademark, a candidate trademark creating section for creating an unregistered and unfiled candidate trademark based on the related trademark is provided in the system. May be provided. The candidate trademark creation unit creates an unregistered and unregistered trademark based on the input trademark and / or the registered and pending related trademark retrieved from the trademark database based on the trademark. Create an unfiled candidate trademark. More specifically, a trademark that is different from the input trademark by one character or the like and is not in the trademark database (not a related trademark) is created as a candidate trademark.

For example, by designating one or more characters to be changed from among the constituent characters of the input trademark from the input device, the designated character is replaced with another character and is not in the trademark database. By extracting the trademark, an unregistered and unfiled candidate trademark is created.

The similarity between the created candidate trademark and the related trademark is determined by the automatic similarity determination unit. Then, a trademark having a relatively low similarity is finally left. Therefore, it can be said that the remaining candidate trademarks are unregistered, have not been filed, and are relatively likely to be registered. FIG. 22 shows an example of a series of these flows. In addition, among the candidate trademarks, for the trademark to be adopted, the related trademark is further searched for, and the automatic similarity determination unit determines whether or not each related trademark and the adopted candidate trademark are similar. It is more preferable to confirm the presence or absence.

By the way, in the above-described embodiment, an example has been described in which, in all cases, the similarity is determined using a neuro in order to determine the similarity of two contrasting trademarks.
In the case where there is an identical decision (the same characteristic part) in the past, a result of the decision may be used preferentially to output whether or not the similarity exists. That is, when the neuro is used, the similarity is output as a value of 0 to 1 as a similarity, but if there is the same trial decision example in the past, the result is prioritized.

For example, a network for judging similarity between trademarks having one character difference in FIG. 3 will be described as an example. In this case, based on past trademark trial decisions, etc., a trial decision database that stores a large amount of information that stores at least one character that is different from each other, each character before and after the different character, and a similarity determination result as one pair is used. Can be In addition, it is also possible to accumulate two contrasting trademarks and the similarity determination result, and extract a characteristic portion from the trademark in accordance with a network to be used. In addition, the trial decision database is set in place of or in addition to the trademark database.

Then, as shown in the flowchart of FIG. 23, the two trademarks of the input trademark or the related trademark retrieved from the trademark database based on the trademark inputted from the input device and the input trademark are used. A trademark is input to a feature extraction unit (or data processing means), and a feature portion (each different character and each character before and after it) is extracted from those two trademarks.

Thereafter, regarding the extracted features, the data processing means searches the decision database for the presence or absence of a past decision under the same conditions. If there is a decision under the same conditions in the past, the result of the similarity judgment (result of the decision) is read from the database, and the result is output to the output device. On the other hand, if no decision was made in the past under the same conditions, input layer pattern information is created based on the characteristics, and the input layer pattern information is given to the neuro of the automatic similarity determination unit to output the output layer. As a result, a similarity judgment result of both trademarks is obtained, and the result is output to the output device.

In the case where a plurality of decisions are made under the same conditions and the results are inconsistent, a decision is made based on the majority decision,
It is possible to respond by giving priority to the result of the decision with the most conditions, giving priority to the latest decision, or outputting that there is a decision with a different result, taking into account the character part other than the characteristic part. . If a decision is made under the same conditions,
It is preferable that detailed information (trial decision publication etc.) be obtained based on the trial number and the like.

Neuro for specifying the referee case (FIG. 15)
In the same way, in all cases, it is possible to make similarity judgment using neuro in all cases, and if there is a trial decision under the same conditions in the past, the result is given priority to make similarity judgment. Can be offered. For example, a description will be given of a network for judging the similarity between trademarks having one character difference as an example. In this case, based on a past trademark decision example or the like, at least a mutually different character and each character before and after the different character, In addition, a trial decision database is used which stores a large number of information in which the trial number and, preferably, the result of the similarity determination of the trial are combined as one pair. It is also possible to accumulate the respective names of the two trademarks and extract a characteristic portion corresponding to the network, such as a different character or characters before and after the different names.

Then, as shown in the flowchart of FIG. 24, a feature (in this embodiment, a different character and characters before and after it) is extracted from the two trademarks to be compared by a feature extraction unit (or data processing means). Is done. After that, with respect to the extracted feature, the presence / absence of a decision under the same condition in the past is searched in the decision database. If there is a trial decision under the same conditions in the past, the trial case is specified by the trial number or the like and is output to the output device. On the other hand, if no decision was made in the past under the same conditions, input layer pattern information is created based on the characteristics, and the input layer pattern information is given to the neuro of the automatic similarity determination unit to output the output layer. Thus, a trial case with one or more similar conditions is specified and output.

It should be noted that the processing when a plurality of decisions are made under the same conditions and the results are inconsistent is the same as described above. Incidentally, in addition to specifying the trial case by the trial number or the like, the result of the similarity determination may be read from the trial decision database through the specification of the trial case, and the result of the similarity determination may be output.

Next, a similar character string creation system will be described. This system is configured to include a similar character string creation unit composed of a neural network for creating a similar character string constructed as described above, and creates a similar character string such as a character string input from an input device, It is used as appropriate.

Here, a trademark is considered as a character string, and creation of a similar trademark is considered. For example, a trademark that is the same or similar to a trademark that is to be used or filed in the future or in the future,
It is a system suitable for searching whether or not a patent office has already been registered or filed. In this case, the system is provided with a similar character string creation unit composed of a neural network for creating a similar character string, and this similar character string creation unit is preferably provided in the similarity determination system.

That is, in addition to the similar character string creation unit, the trademark database 5 and an input device and an output device are provided. In addition, a system without the trademark database 5 can also be used. In this case, based on a specific trademark input from the input device 2, input layer pattern information is created by normalization processing means of the feature extraction unit. Based on the information, the similar character string creating unit creates a similar trademark, and outputs the result to an output device.

As described in the similar character string creating method, the similar character string creating unit is constituted by a neural network constructed for a similarity judgment method of a trademark or a neural network dedicated to similar character string creating.

When using a neural network constructed for the similarity judgment method of a trademark, the automatic similarity judgment unit in FIG. 18 functions as a similar character string creation unit. When a neural network dedicated to creating a similar character string is used, a similar character string creating unit composed of a neural network dedicated to creating a similar character string is used instead of the automatic similarity determination unit in FIG. It is provided in. In either system, a character string is developed in the feature extraction unit based on a first trademark (such as an input trademark), which is the basis for creating a similar trademark, and normalized by each character string and an appropriate second character. Input layer pattern information is created by the processing means, and is input to the input layer of the network of the similar character string creating unit.

The similar character string creating section creates a similar character string of the first trademark based on each neuro output. With respect to the input trademark and each similar trademark created by the similar character string creating unit, the trademark database 5 searches for the same trademark as the trademark or a trademark partially including the trademark by the trademark database 5 via the data processing means. You. Then, a search result of the presence or absence of a trademark that is the same or similar to the input trademark is output to the output device 3. FIG. 25 shows a flow of a series of these processes. In this case,
The similar character string creation unit can create the similar trademark on the trademark database side (server), send the trademark input from the input device via a communication line, and create similar trademarks and search for them outside the personal computer.

By the way, the predetermined program, automatic similarity judgment processing, similar character string creation processing, candidate character string (trademark)
Each part of the creation process, or input / output control and communication control,
The part of feature extraction processing (normalization processing) is a CD-ROM
And other various recording media. It is preferable that the connection load of the neuro execution means can be updated by an appropriate means.

It should be noted that the similarity determination method / system of the present invention
The method and system for creating a similar object are not limited to the configuration of each of the above embodiments, and can be appropriately changed. In particular, the configuration of the unit (unit set) for character identification (the method of inputting and identifying characters) and what constitutes one character (one sound) can be changed as appropriate. It is not bound by the description.

Further, in each of the above embodiments, the product category and the like are not considered, but it is natural that the search of the trademark database and the like can be performed in the similar group unit, the classification unit, or all classes. This also applies to the invention described in each claim.

In the similarity judgment network, similarity judgment between dissimilarity and similarity has been described. However, the same trademark is used for learning patterns so as to be able to cope with the case where the contrasting trademark is the same. In addition to the information, it may be configured so that similarity can be determined between dissimilarities and the same. In this case, the similarity of 1 may include the same, or may be 1 and the similarity may be between 0 and less than 1 (not including 0). 5).

In the system for determining similarity or the like, a plurality of networks of different formats are provided so as to correspond to various arrangements of different characters of a character string (trademark), and to correspond to arrangement of different characters of a character string to be compared. Then, the similarity determination may be made in the optimal network. For example, in addition to the one-character trademark shown in FIG. 3, a two-character trademark is used for a trademark in which different characters are adjacent to each other, a trademark having one character between different characters, a trademark having two characters between different characters, and the like. A network can be provided, and the arrangement of different characters can be determined by the feature extraction unit, an optimal network can be selected, and input layer pattern information can be input to the input layer of the selected network.

[0164]

As described in detail above, according to the present invention, for example, it is easy to judge whether or not the names of trademarks are similar.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of a three-layer perceptron.

FIG. 2 is a table showing an example of a unit set for character identification for creating pattern information for neuro construction such as a trademark similarity determination method and system of the present invention.

FIG. 3 is a diagram showing an example of a network for a method and a system for judging similarity of a primary difference trademark of the present invention.

FIG. 4 is a diagram illustrating an example of normalized data of a position of a different character in the network of FIG. 3;

5 is a diagram showing a part of a table showing an example of pattern information for constructing the network of FIG. 3, and one table is constituted up to FIG. 8;

FIG. 6 is a diagram showing a continuation of the table in FIG. 5;

FIG. 7 is a diagram showing a continuation of the table in FIG. 6;

FIG. 8 is a diagram showing a continuation of the table in FIG. 7;

FIG. 9 is a diagram showing an example of a network for a method and a system for judging similarity of a two-character difference trademark according to the present invention.

FIG. 10 is a diagram showing changes in the number of times of learning and errors.

FIG. 11 is a diagram showing an example of a network for judging similarity between trademarks having a difference of two characters or less according to the present invention, showing a configuration thereof and a part of pattern information.

FIG. 12 is a diagram showing another example of a network for judging the similarity between trademarks having a difference of two characters or less according to the present invention, showing a part of the configuration and part of pattern information.

FIG. 13 is a diagram showing an example of a network for judging similarity between two-character replacement trademarks according to the present invention, showing a configuration thereof and a part of pattern information.

FIG. 14 is a diagram showing another example of a network for judging similarity between two-character replacement trademarks according to the present invention, showing a configuration thereof and a part of pattern information.

FIG. 15 is a diagram showing an example of a similar trial decision example search network for judging similarity of a trademark of the present invention, showing a configuration thereof and a part of pattern information.

FIG. 16 is a diagram showing an example of a similar character string creation network according to the present invention, showing its configuration and a part of pattern information.

FIG. 17 is a schematic diagram showing an example of a system for judging similarity of a trademark according to the present invention.

FIG. 18 is a diagram showing an example of a system for determining similarity of a trademark according to the present invention.

FIG. 19 is a diagram showing an example of a processing flow of a trademark similarity determination method of the present invention.

FIG. 20 is a diagram showing another example of the processing flow of the trademark similarity determination method of the present invention.

FIG. 21 is a diagram showing another example of a system for judging similarity of a trademark according to the present invention.

FIG. 22 is a diagram showing still another example of the processing flow of the method for judging similarity of a trademark of the present invention.

FIG. 23 is a diagram showing still another example of the processing flow of the trademark similarity determination method of the present invention.

FIG. 24 is a diagram showing still another example of the processing flow of the method for judging similarity of a trademark of the present invention.

FIG. 25 is a diagram showing an example of a processing flow of a similar character string creating method for a trademark of the present invention.

[Explanation of symbols]

 1 PC 2 Keyboard 3 Display 4 Printer 5 Trademark Database

Claims (3)

[Claims] 1. An input layer is provided with each feature extracted from each of two contrasting objects, while an output layer is provided with:
By learning a plurality of pieces of learning pattern information including learning pattern information that gives a similarity between two objects to be compared and each of the pieces of learning pattern information, each of the objects includes a different one. The input layer pattern information based on the features extracted from each of the two objects to be compared is input to the input layer of the constructed neural network. A similarity / dissimilarity determination method using a neural network, characterized by determining the similarity between objects. 2. Learning pattern information comprising an input layer learning signal based on each of two objects to be compared and an output layer teacher signal based on the similarity between the two objects to be compared. An automatic similarity determination unit composed of a neural network constructed by learning a plurality of learning pattern information including ones in which the respective objects are different from each other with pattern information; and two objects to be compared with each other. Based on the output of the automatic similarity determination unit based on the input layer pattern information, the feature extraction unit that creates the input layer pattern information based on the input layer pattern information and provides the input layer pattern information to the input layer of the automatic similarity determination unit A similarity determination system using a neural network, comprising: an output unit that outputs a degree of similarity between two objects. 3. A method for creating an analog of a specific first object, comprising: learning a plurality of learning pattern information for providing an input layer and an output layer with features extracted from two objects similar to each other. Then, an input layer pattern information created based on the first object is inputted to an input layer of the constructed neural network to obtain an output of an output layer, and an output of the output layer is obtained. A method for creating an analog using a neural network, wherein the analog is created based on the first object.