JP5470620B2 - Annotation acquisition device, annotation acquisition method, and program - Google Patents

Description

  The present invention relates to a different notation acquisition device and the like for extracting term pairs of different notations.

  As a conventional technique for extracting different notations, there is research by Aramaki et al. (See Non-Patent Document 1). This study was used to extract terminology for medical terminology. Note that the different notation is a term that is a synonym but a different expression, such as “spaghetti” for “spaghetti”. Two terms with different notation are called different notation pairs.

  The first concept of the variant notation is as follows. For example, for example 1 (inquiry email, inquiry email) and example 2 (gakuen struggle, school dispute) of term pairs, example 1 is a different notation pair, and example 2 is a Japanese synonym pair instead of a different notation pair. To do. In other words, in the first concept, if the same word is a variant, it is a different notation pair, and if it is not the same word, it is not a different notation pair even if the meaning is the same. Although struggle and conflict have approximately the same meaning, but are not the same word, Example 2 is not a different pair. On the other hand, “inquiry” and “inquiry” in Example 1 are different from each other, but can be determined to be variants of the same word, and are therefore different pairs.

  In addition, the second concept of the different notation pair is a concept in which synonyms are also differently described. In the second way of thinking, not only Example 1 above, but also Example 2 (Gakuen Struggle, Gakuen Conflict) are different pairs.

  Furthermore, the different notation and different notation pairs may be similar in concept to the above, and the different notation and different notation pairs will be widely understood.

  Further, as a conventional technique, there is a technique about a machine learning method (see, for example, Non-Patent Document 2 to Non-Patent Document 4).

Eiji Aramaki, Takeshi Imai, Kengo Miyo, Kazuhiko Ohe: Orthographic Disambiguation Incorporating Transliterated Probability, International Joint Conference on Natural Language Processing (IJCNLP2008), pp.48-55, 2008. Masaki Murata, Language processing based on machine learning, Faculty of Science and Engineering, Ryukoku University. Invited lecture. 2004.http: //www2.nict.go.jp/x/x161/member/murata/ps/kougi-ml-siryou-new2 .pdf Japanese-English translation of tense aspect modality using support vector machine, Maki Murata, Ma Aoi, Kiyotaka Uchimoto, Hitoshi Isahara, IEICE Language Understanding and Communication Study Group NLC2000-78, 2001. CRL in the SENSEVAL2J dictionary task, Masaki Murata, Masao Uchiyama, Kiyotaka Uchimoto, Mao Ai, Hitoshi Isahara, IEICE NLC2001-40, 2001.

  However, the conventional technology does not deal with general Japanese variants, and the application of the conventional technology to general variants of Japanese did not provide sufficient accuracy for extracting different variants. .

  The different notation acquisition device according to the first aspect of the present invention is a character type that is a feature of a character pair storage unit that can store one or more term pairs having an edit distance of one or more, and a character type of an edited portion that is a character having a different term pair. A plurality of features including one or more features of a related feature, a dictionary-related feature that is a feature acquired using a term dictionary, and a similarity feature that is a feature that indicates the similarity of two terms constituting a term pair; , A learning data storage unit that can store two or more learning data in which positive / negative information that is information indicating whether a term pair is a different pair of term pairs, and a character type for each term pair in the term pair storage unit A feature acquisition unit that acquires a plurality of features including one or more of related features, dictionary-related features, and similarity features, two or more learning data in a learning data storage unit, and a feature acquisition unit for each term pair Using a plurality of features acquired by The term pair storage unit includes a machine learning unit that determines whether each term pair is a term pair having a different notation, and an output unit that outputs a determination result in the machine learning unit. .

  With this configuration, it is possible to extract differently-notated term pairs with high accuracy regardless of the field of term pairs.

  In addition, the different notation acquisition device of the second invention is different from the first invention in that the character type-related feature is different in the character type of the edited part of the two terms included in the term pair, and the two terms This is information indicating whether or not the edited part is a number, and the feature acquisition unit is different for each term pair in the term pair storage unit, and the character type of the edited part of the two terms in the term pair is different, and two This is a different notation acquisition device that determines whether or not the edited part of a term is a number having the same value, and acquires the determination result as a character type-related feature.

  With this configuration, it is possible to extract differently-notated term pairs with high accuracy regardless of the field of term pairs.

  In addition, the different notation acquisition device of the third invention is different from the first invention in that the character type-related feature is that the two types of terms that the term pair has are Roman characters, and the editing of the two terms This is information indicating whether or not the location is a difference between uppercase and lowercase. The feature acquisition unit has a Roman character for the editing location of the two terms in the term pair for each term pair in the term pair storage unit. And it is a different notation acquisition apparatus which determines whether it meets the conditions that the edit location of two terms is a difference between a capital letter and a small letter, and acquires a judgment result as a character type related feature.

  With this configuration, it is possible to extract differently-notated term pairs with high accuracy regardless of the field of term pairs.

  The different notation acquisition device of the fourth aspect of the invention further includes a term dictionary that can store one or more term information having terms and a representative notation of terms, as compared with the first aspect. The feature is information indicating whether or not the representative notation of the two terms included in the term pair is the same, and the feature acquisition unit represents the representative of the two terms included in the term pair for each term pair in the term pair storage unit. This is a different notation acquisition device that acquires a notation from a term dictionary, determines whether or not two acquired representative notations are the same, and acquires the determination result as a dictionary-related feature.

  With this configuration, it is possible to extract differently-notated term pairs with high accuracy regardless of the field of term pairs.

  Further, the different notation acquisition device of the fifth invention is different from the supervised machine learning method in that the dictionary-related features are different from the supervised machine learning method or the same classification method as in the first invention. However, it is the result of determining whether a term pair is a term pair with different notation by a classification method with different learning data, and the feature acquisition unit performs supervised machine learning for each term pair in the term pair storage unit. The classification method is different from the law, or the classification method is the same classification method but the learning data is different, and it is judged whether the term pair is a different pair of terminology, and the judgment result is acquired as a dictionary-related feature. It is a notation acquisition device.

  With this configuration, it is possible to extract differently-notated term pairs with high accuracy regardless of the field of term pairs.

  Further, the different notation acquisition device of the sixth invention further comprises a term dictionary capable of storing one or more term information having terms and readings of terms, as compared with the first invention. Is information indicating whether or not the readings of the two terms included in the term pair match, and the feature acquisition unit reads the two terms included in the term pair from the term dictionary for each term pair in the term pair storage unit. This is a different notation acquisition device that determines whether the readings of two terms match, and acquires the determination result as a dictionary-related feature.

  With this configuration, it is possible to extract differently-notated term pairs with high accuracy regardless of the field of term pairs.

  Also, with the different notation acquisition device of the seventh invention, in contrast to any of the first to sixth inventions, the machine learning unit may determine whether each term pair in the term pair storage unit is a term pair having a different notation. The output unit is a different notation acquisition device that outputs the score acquired by the machine learning unit.

  With this configuration, it is possible to extract term pairs with different notations with higher accuracy regardless of the field of term pairs.

  Further, in the different notation acquisition device of the eighth invention, in contrast to the seventh invention, the output unit includes threshold storage means for storing a threshold of the score, and the score acquired by the machine learning unit is equal to or greater than the threshold. A threshold judgment means for judging whether or not the threshold value is greater than a threshold value, and a term pair corresponding to the score judged by the threshold judgment means to be equal to or greater than the threshold value or greater than the threshold value is a judgment result that the term pair is a different notation, and the judgment result Or it is an different notation acquisition apparatus provided with the output means which outputs any one or more of a term pair of different notation or a term pair which is not different notation.

  With this configuration, it is possible to extract term pairs with different notations with higher accuracy regardless of the field of term pairs.

  Further, the different notation acquisition device of the ninth aspect of the present invention is a character type related feature that is a feature related to a character type of an edited portion that is a different character of a term pair, a dictionary related feature that is a feature acquired using a term dictionary, A plurality of features including one or more features of similarity features that indicate the similarity of two terms constituting a term pair, and positive / negative information that indicates whether the term pair is a different pair of terms A learning data storage unit that can store two or more learning data in association with information, and an allo notation that can store one or more different notation patterns having a pair of a first character string and a second character string indicating a notation pattern A pattern storage unit, a reception unit that receives one or more terms, and one or more terms in the different notation pattern storage unit applied to each of the one or more terms received by the reception unit Generated with one or more terms A term pair generation unit that generates an alternate notation candidate term pair that is a term pair of one or more different notation candidates having a word, and a character type for each of the one or more alternate notation candidate term pairs generated by the term pair generation unit Stores learning data for a feature acquisition unit that acquires a plurality of features including one or more features of related features, dictionary related features, and similarity features, and for each different notation candidate term pair generated by the term pair generation unit Whether each of the different notation candidate term pairs in the term pair storage unit is a different notation term pair by supervised machine learning using two or more learning data of the part and a plurality of features acquired by the feature acquisition unit It is a different notation acquisition apparatus which comprises the machine learning part which determines whether or not, and the output part which outputs the determination result in a machine learning part.

  With this configuration, it is possible to automatically generate a candidate for a different pair of terminology.

  Further, the different notation acquisition device of the tenth invention is different from the ninth invention in that an different notation term pair storage unit capable of storing one or more different notation term pairs having an edit distance of 1 and an different notation term pair. The editing part acquisition unit for acquiring the editing part of one or more different pairs of terms stored in the storage part, and the first character string and the second character string are paired from the editing part acquired by the editing part acquisition unit. A different notation acquisition device further comprising an different notation pattern acquisition unit for acquiring different notation patterns, and an different notation pattern storage unit for storing the different notation patterns acquired by the different notation pattern acquisition unit in the different notation pattern storage unit is there.

  With such a configuration, it is possible to automatically acquire a different notation pattern for automatically generating a candidate for a different pair of terminology.

  In addition, the different notation acquisition device of the eleventh aspect of the invention provides that the edit distance of the term pair is 2 with respect to any one of the first to tenth inventions, and the feature acquisition unit has two difference characters of the term pair. Including at least one of a character type-related feature, a dictionary-related feature, and a similarity feature, with the difference character acquisition unit that acquires each of the pairs and the two difference characters acquired by the difference character acquisition unit as targets independently A feature acquisition unit for acquiring two sets of a plurality of features, and the machine learning unit learns a plurality of features of each set for each set of the plurality of features acquired by the feature acquisition unit; Whether or not a plurality of features of each pair of term pair storage units is a pair of features corresponding to a term pair of different notation by using supervised machine learning method using two or more learning data of the data storage unit Judgment and using two judgment results, the term pair whose edit distance is 2 is a different term A different notation acquisition device that determines whether a.

  With such a configuration, it is possible to determine whether or not a pair of terms with an editing distance of 2 is a term pair with different notation with high accuracy.

  According to the different notation acquisition apparatus of the present invention, it is possible to extract a differently-notated term pair with high accuracy regardless of the field of term pairs.

The block diagram of the different notation acquisition apparatus in Embodiment 1 of this invention The flowchart explaining operation | movement of the same notation acquisition apparatus Flow chart explaining operation of homology acquisition processing The figure which shows the example of the same term dictionary The figure which shows the concept of margin maximization of the support vector machine method The figure which shows the breakdown which judged whether it is a Japanese variant notation pair or a Japanese variant notation pair by majority vote in the Japanese term pair whose edit distance used in the experiment is 1 The figure which shows the coincidence evaluation method by Landis et al. Figure showing the result of applying the baseline method to the closed data and open data Figure showing the result of examining whether the feature is effective using the bootstrap method Diagram showing examples of equivalence The figure which shows the examination result of the ratio which the term pair classified into Japanese different notation pair whose edit distance is 1 from the large-scale similar word list is included in various dictionaries using the proposed method The figure which shows the classification accuracy of the same SVM In the same various dictionaries and term pair DB, five pairs of Japanese term pairs that were classified as Japanese variant pairs with an edit distance of 1 and Japanese term pairs that were not classified were randomly selected. Figure showing the results The figure which shows the evaluation standard of the same threshold Figure showing the ratio of recall and precision The figure which shows the experimental result when the same baseline method is used The figure which shows the experimental result when the same baseline method is used The figure which shows the experimental result when the same baseline method is used The block diagram of the different notation acquisition apparatus in Embodiment 2 of this invention The flowchart explaining operation | movement of the same notation acquisition apparatus The figure which shows the example of the same notation pattern Overview of the computer system in the above embodiment Block diagram of the computer system

  Hereinafter, an embodiment of a different notation acquisition apparatus and the like will be described with reference to the drawings. In addition, since the component which attached | subjected the same code | symbol in embodiment performs the same operation | movement, description may be abbreviate | omitted again.

  (Embodiment 1)

  In the present embodiment, a feature relating to a character type of a different character, a feature obtained using a term dictionary from a pair of terms having an edit distance of 1 or 2 or more, and a feature of one or more of the similarities of two terms. A description will be given of a different notation acquisition apparatus that takes out a plurality of features to be included and uses the plurality of features to determine whether a term pair is a differently expressed term pair by a supervised machine learning method.

FIG. 1 is a block diagram of the different notation acquisition apparatus 1 in the present embodiment.
The different notation acquisition apparatus 1 includes a term pair storage unit 11, a learning data storage unit 12, a term dictionary 13, a reception unit 14, a feature acquisition unit 15, a machine learning unit 16, and an output unit 17. The feature acquisition unit 15 includes a difference character acquisition unit 151 and a feature acquisition unit 152. The output unit 17 includes a threshold storage unit 171, a threshold determination unit 172, and an output unit 173.

  The term pair storage unit 11 can store one or more term pairs having an edit distance of 1 or 2 or more. The edit distance is the number of different characters. A term pair is two terms. A term pair with an edit distance of 2 is a term with two different numbers of characters. The term is usually a noun or a noun phrase, but may be a term of another part of speech such as an adjective.

  The learning data storage unit 12 can store two or more learning data. The learning data has a plurality of features of the term pair and positive / negative information. The learning data may have a term pair. Here, the plurality of features of the term pair includes one or more of a character type related feature, a dictionary related feature, and a similarity feature. The feature is information used as a clue when the different notation acquisition apparatus 1 learns.

  The character type-related feature is a feature related to a character type of an edited portion that is a character having a different term pair. The character type-related feature is information indicating, for example, whether or not the character types of the editing positions of the two terms included in the term pair are different and whether the editing positions of the two terms have the same value. The character type-related features are, for example, that the number of characters of two terms included in a term pair is the same, the character types of the edited portions of the two terms are different, and the edited portions of the two terms have the same value. It is information indicating whether it is a number. The character type-related feature is, for example, information indicating whether or not the character type of the edited part of the two terms of the term pair is Roman and the edited part of the two terms is different between uppercase and lowercase letters. It is. The character type-related features are, for example, that the number of characters of two terms included in a term pair is the same, the character type of the edited part of the two terms is Roman, and the edited part of the two terms is capitalized. This is information indicating whether or not the difference is lowercase.

  The dictionary-related feature is a feature acquired using the term dictionary 13. For example, the dictionary-related feature is obtained by using a stacking algorithm to classify a term pair by a classification method different from the supervised machine learning method used by the machine learning unit 16 or a classification method that is the same classification method but has different learning data. Is a result of determining whether or not is a pair of terms with different notations. Here, “learning data is different” includes a case where the sets of term pairs from which the learning data is based are different, a case where the features of the learning data are different, and the like. The dictionary-related feature is information indicating whether, for example, the representative notation of two terms included in a term pair is the same. The dictionary-related feature is information indicating whether readings of two terms included in a term pair match, for example. The dictionary-related feature is information indicating, for example, whether the number of characters of two terms included in the term pair is the same and whether the readings of the two terms match. The classification method is a classification method for determining whether or not a pair of terms is notated. Further, the classification method different from the supervised machine learning method is that the classification method and algorithm are different from the supervised machine learning method.

  The similarity feature is a feature that indicates the similarity of two terms constituting a term pair. The degree of similarity between two terms is obtained by using information on whether or not the terms appear in a similar context on the Web. The technique for obtaining the similarity of terms is “Shinichi Kazama, De Saeger, Stijn, Kentaro Torisawa, Maki Murata“ Creating a large-scale similar word list using dependency stochastic clustering ”The 15th Language Processing Society of Japan Annual convention (NLP2009) ”etc. That is, the method for obtaining the similarity between two terms is a known technique. There is no limitation on the method of calculating the similarity between the two terms.

  The positive / negative information is information indicating whether or not the term pair is a differently expressed term pair. The positive / negative information is a positive example (for example, “1”) if it is a term pair with different notation, and a negative example (for example, “0”) if it is not a term pair with different notation.

  The term dictionary is a group of information including information on terms in different notations. Examples of term dictionaries and data structures will be described later.

  As another feature, there is an edited portion character feature which is information on the character at the edited portion or the characters around the character at the edited portion.

  The term dictionary 13 can store one or more term information. The term dictionary 13 is an information group including, for example, information on terms in different notations. The term dictionary 13 does not need to explicitly associate two different terms. The term information includes, for example, a term and a representative representation of the term. The term information includes, for example, a term and a term reading.

  The reception unit 14 receives input from the user. This input is, for example, an operation instruction for operating the different notation acquisition device 1. The accepting unit 14 may accept a target term pair for determining whether or not the notation is used. The input means such as operation instructions may be anything such as a numeric keypad, a keyboard, a mouse, or a menu screen. The accepting unit 14 can be realized by a device driver for input means such as a numeric keypad or a keyboard, control software for a menu screen, and the like.

  The feature acquisition unit 15 acquires a plurality of features including at least one of a character type related feature, a dictionary related feature, and a similarity feature for each term pair in the term pair storage unit 11. The plurality of features are, for example, 68 features described later.

  The feature acquisition unit 15 has a condition that, for each term pair in the term pair storage unit 11, the character types of the edited portions of the two terms included in the term pair are different, and the edited portions of the two terms have the same value. Is determined as a character type-related feature. For example, the feature acquisition unit 15 determines that the above condition is met because the edited portions are “3” and “three” for the term pairs “three-party talk” and “three-party talk”. The feature acquisition unit 15 has, for example, “125” and “125” for the term pairs “125” and “125”, so the number of characters in the term is not the same. Is determined to satisfy the above condition.

  For example, for each term pair in the term pair storage unit 11, the feature acquisition unit 15 has two types of editing positions of the two terms in the term pair in Roman letters, and the editing positions of the two terms in uppercase and lowercase letters. It is determined whether or not the condition of difference is met, and the determination result is acquired as a character type related feature. For example, the feature acquisition unit 15 determines that the edited portion matches the term pair “A” and “a”, and if the edited portion does not match the term pair “A” and “b”. to decide.

  The feature acquisition unit 15 is, for example, a classification method different from the supervised machine learning method used by the machine learning unit 16 or the same classification method for each term pair in the term pair storage unit 11 but with different learning data. According to the method, it is determined whether or not the term pair is a different notation term pair, and the determination result is acquired as a dictionary-related feature. A machine learning method using such dictionary-related features is called a method using a stacking algorithm. The classification method different from the supervised machine learning method used by the machine learning unit 16 means that, when the above machine learning method is SVM, a machine learning method such as a decision tree different from SVM, a classification method based on a rule described later Etc.

  Specifically, the stacking algorithm is a classification method according to the following procedure, for example. First, teacher data is created using a JUMAN dictionary. That is, a word pair whose edit distance is one character is extracted from a set of words in the JUMAN dictionary. Here, 904612 pairs of words with an edit distance of one character can be extracted. Among them, word pairs (25934 sets) having the same representative notation are taken out. Next, in the JUMAN dictionary, word pairs having the same representative notation are set as positive examples, and word pairs other than those are set as negative examples. Thus, teacher data is created.

  Next, machine learning using the teacher data as learning data is performed. The teacher data is not limited to the teacher data described above, and other teacher data may be used. In addition, as a feature used in machine learning, a feature obtained by removing the feature of S54 from all the features of the present invention (the features of S1 to S68) is used. Note that other features may be used as the features used in machine learning.

  In practice, the data to be given the feature of S54 is classified using the learning result. Information indicating whether the classification result is a positive example or a negative example is added to the data as the feature of S54.

  Then, by using the learning data to which the feature of S54 is given (having 68 features), by performing machine learning on the problematic term pair, whether the problematic term pair is a different notation pair Judge whether or not.

  In the method based on the stacking algorithm, the result of learning whether or not the representative notation matches in the JUMAN dictionary can be given as a feature. Therefore, the term pairs that are not actually described in the JUMAN dictionary can be represented by the JUMAN dictionary. Information indicating whether or not the term pair has a tendency to match the notation can be given.

  For example, the feature acquisition unit 15 acquires, for each term pair in the term pair storage unit 11, representative representations of two terms included in the term pair from the term dictionary 13, and whether or not the two representative representations acquired are the same. And the determination result is acquired as a dictionary-related feature.

  The feature acquisition unit 15 acquires, for example, readings of two terms from the term dictionary 13 for each term pair in the term pair storage unit 11, determines whether the readings of the two terms match, and determines the determination result. Obtained as dictionary-related features. Further, the feature acquisition unit 15 acquires, for each term pair in the term pair storage unit 11, the number of characters of the two terms included in the term pair is the same, and acquires readings of the two terms from the term dictionary 13, It may be determined whether the readings of the two terms match, and the determination result may be acquired as a dictionary-related feature.

  Note that the above-described determination result is, for example, “1” when the condition is met, and “0” in other cases.

  The difference character acquisition means 151 acquires two sets of difference characters, respectively, for the term pairs whose edit distance is two. For example, consider a case where a pair of terms having two edit distances are (1) “can” “can” (2) “can understand” “can” (3) “IX (roman numeral 9)” “9”. . (1) is the case where both term pairs have the same number of characters. (2) is a case where the number of characters of either term is two more or less than the number of characters of the other term. (3) is a case where the number of characters of one term is one more or less than the number of characters of the other term. In the case of (1), the difference character acquisition unit 151 numbers characters “1, 2, 3,...” From the front to the rear for the terms “can” and “can”, and the same for each term. A set of two difference characters, “de”, “de”, “ki”, and “coming”, having character numbers and different characters “de”, “out”, “ki”, and “coming” being differential characters. To get. In the case of (2), the difference character acquisition unit 151 acquires two sets of difference characters, “reason” “” and “solution” “” (“” is NULL). In the case of (3), the difference character acquisition unit 151 uses two sets of difference characters “I” “9” and “X” “” or two sets of differences “I” “” and “X” “9”. Get a character.

  Further, the difference character acquisition unit 151 acquires a set of difference characters for a term pair whose edit distance is 1. For example, a pair of terms whose edit distance is 1 is (1) “difficulty” “difficulty” (2) “Firefox” “FireFox” (3) “hepatitis virus” “hepatitis virus” (4) “in literature history” “literature” “History” (5) “Flower” “Saki” (6) “Closet” “Closet” (7) “Adult / Child” “Adult Child” (8) “First” and “First” , The difference character acquisition means 151 respectively (1) “go” “go” (2) “f” “F” (3) “i” “i” (4) “up” “” (5) “ki” "" (6) "" "-" (7) "·" "" (8) "1" "One" is acquired.

  Further, the difference character acquisition unit 151 acquires three or more sets of difference characters for a term pair having an edit distance of 3 or more. For example, if the term pair whose edit distance is 4 is “1025” and “thousand twenty-fifth”, the difference character acquisition unit 151 sets “1” “1000”, “0” “two”, “2”. Four sets of difference characters “10”, “5”, and “5” are acquired. Here, the difference character is a character different in two terms.

  The feature acquisition unit 152 targets two difference characters acquired by the difference character acquisition unit 151 as independent targets, and acquires a plurality of features including one or more of character type related features, dictionary related features, and similarity features 2 Get a pair. For example, consider a case where a pair of terms having two edit distances is (1) “can”, “can” (2) “can understand” “can” (3) “IX” “9”. For the term pair of (1), the feature acquisition unit 152 extracts features for each of the two sets of difference characters “de”, “out”, “ki”, and “coming”, and extracts each feature from the difference character. The extracted features are considered different, and two types of test data are created. The feature acquisition unit 152, for example, regarding character type-related features indicating whether or not the character types of the editing positions of the two terms of the term pair are different and whether the editing positions of the two terms are numbers having the same value, It is determined that the edited portions of “de” and “out” are not numbers having the same value, and the character type related feature “0” is acquired. For example, the feature acquisition unit 152 acquires readings of two terms from the term dictionary 13 and acquires a dictionary-related feature “1” indicating whether the readings of the two terms match. The feature acquisition unit 152 acquires “de” reading “de” from the term dictionary 13 and determines that “de” and “depart” read match. Also, the feature acquisition unit 152, for example, with respect to the difference character “de” “out”, regarding the feature of the character before and after the difference character (edited part), the feature “” (none) of the previous character, the subsequent character Acquire the features “Ki” and “Ki”. Also, the feature acquisition unit 152, for example, with respect to the difference characters “ki” and “coming”, the features “out” and “de” of the previous character and the feature “ Get ". With this process, another difference character is included in the feature.

  In addition, for the term pair (2), the feature acquisition unit 152 extracts features for each of the two sets of difference characters of “reason” “” and “solution” “” as in (1). The features extracted from the difference characters are considered to be different, and two types of test data are created.

  Further, for the term pair (3), the feature acquisition unit 152, for example, each of two sets of difference characters “I”, “9”, “X”, and “”, as in (1) and (2). The features are extracted from the target, and the features extracted from the difference characters are considered to be different, and two types of test data are created.

  The feature acquisition unit 152 acquires a plurality of features including one or more of a character type related feature, a dictionary related feature, and a similarity feature using one or more sets of difference characters acquired by the difference character acquiring unit 151. To do. A specific method for acquiring features such as character type-related features, dictionary-related features, and similarity features will be described later.

  For each term pair, the machine learning unit 16 uses two or more learning data in the learning data storage unit 12 and a plurality of features acquired by the feature acquisition unit 15 to perform the term pairing by the supervised machine learning method. It is determined whether or not each term pair in the storage unit 11 is a term pair having a different notation.

  The machine learning unit 16 may determine whether each term pair in the term pair storage unit 11 is a differently expressed term pair, and may also acquire a score indicating the accuracy of the differently expressed term pair.

  The machine learning unit 16 uses a plurality of features of each set and two or more learning data of the learning data storage unit 12 for each of the two sets of features acquired by the feature acquisition unit 152. Using the supervised machine learning method, it is determined whether or not a plurality of features of each set in the term pair storage unit 11 is a set of features corresponding to different pairs of terminology, and the edit distance is calculated using the two determination results. It is determined whether or not the term pair with 2 is a term pair with different notation.

  The algorithm of the supervised machine learning method does not matter. The supervised machine learning method is, for example, a support vector machine (SVM). SVM is "http://chasen.org/~taku/software/TinySVM/" "http://en.wikipedia.org/wiki/%E3%82%B5%E3%83%9D%E3%83% BC% E3% 83% 88% E3% 83% 99% E3% 82% AF% E3% 82% BF% E3% 83% BC% E3% 83% 9E% E3% 82% B7% E3% 83% B3 '' (Search on December 12, 2009). Details of the supervised machine learning method will be described later.

  In addition, using the above two determination results, it is possible to use different terminology pairs when both are positive examples and when both are negative examples, It may not be, and the score that is closer to 0 of the two scores is adopted, and when the adopted score is positive, it is a positive example (another term), and when negative, it is a negative example ( It may be judged that it is not a term of different notation), and the score with the larger absolute value of the score is adopted, and when the adopted score is positive, it is a positive example (term of different notation), and when it is negative, it is negative It may be judged as an example (not an idiom). Alternatively, the smaller score of the two scores may be acquired, and the positive score may be determined as a positive example when the lower score is positive, and the negative score may be determined as negative. That is, there is no limitation on how to use the two determination results. In the case of (2) above (when the number of characters of one term is two more or less than the number of characters of the other term), about 15,000 tags from the large-scale similar word list As a result, the two-character difference data of this pattern did not have a term pair that was determined to be a different notation pair.

  Furthermore, two sets of difference characters (for example, “I” “9” and “X” “” or “I” “” and “X” “9”), that is, two problems (problem 1, problem 2) ), The features are extracted for each difference character, and the features extracted from the difference characters are considered to be different, and four types of test data are created. Then, for each of the two questions, the one with the calculated score closer to 0 is acquired, and the score with the highest absolute value of the scores for each question is set as the score of the question. In this case, it may be determined as a negative example. For example, if the term pair whose edit distance is 2 is (3) “IX” “9”, the problems “I” “9” and “X” “” and “I” “” and “X” can be generated. Then, the machine learning unit 16 acquires the smaller score of “I” “9” and “X” “”, and acquires the smaller score of “I” “” and “X” “9”. Of the two acquired scores, the one with the larger value is set as the score in the term pair “IX” “9”. The machine learning unit 16 may determine that the score is a positive example when the score is positive and a negative example when the score is negative. For example, the machine learning unit 16 obtains the score of “I” “9” and “X” “” that is close to 0, and the scores of “I” “” and “X” “9” The one closer to 0 may be acquired, and the larger of the two acquired scores may be used as the score for the term pair “IX” “9”. That is, it does not matter how the score is calculated using the judgment results of the four types of test data.

  The output unit 17 outputs the determination result in the machine learning unit 16. The output unit 17 may output the score acquired by the machine learning unit 16. The determination result includes information indicating whether each term pair is a different notation, or one or more term pairs having different notations, or one or more term pairs not having different notations. Moreover, the output part 17 may output both a judgment result and a score, and may output one.

  Also, output means display on a display, projection using a projector, printing on a printer, sound output, transmission to an external device, storage in a recording medium, processing to another processing device or other program, etc. It is a concept that includes delivery of results.

  The threshold storage unit 171 stores a score threshold.

  The threshold determination unit 172 determines whether the score acquired by the machine learning unit 16 is equal to or greater than the threshold or greater than the threshold.

  The output unit 173 sets the term pair corresponding to the score determined by the threshold judgment unit 172 to be equal to or greater than the threshold value or greater than the threshold value as a judgment result that is a different term term pair. Output one or more of non-notation term pairs.

  The term pair storage unit 11, the learning data storage unit 12, the term dictionary 13, and the threshold storage unit 171 are preferably non-volatile recording media, but can also be realized by volatile recording media.

  The process in which the information stored in the term pair storage unit 11, the learning data storage unit 12, and the term dictionary 13 is stored is not limited.

  The feature acquisition unit 15, the machine learning unit 16, and the threshold determination unit 172 can usually be realized by an MPU, a memory, or the like. The processing procedure of the feature acquisition unit 15 and the like is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The output unit 17 may or may not include an output device such as a display or a speaker. The output unit 17 can be realized by output device driver software, or output device driver software and an output device.

  Next, operation | movement of the different notation acquisition apparatus 1 is demonstrated using the flowchart of FIG. Here, it is assumed that the different notation acquisition device 1 determines whether a term pair with an editing distance of 1 is a differently expressed term pair.

  (Step S <b> 201) The receiving unit 14 determines whether an operation start instruction has been received. If an instruction is accepted, the process goes to step S202. If no instruction is accepted, the process returns to step S201.

  (Step S202) The feature acquisition unit 15 substitutes 1 for a counter i.

  (Step S203) The feature acquisition unit 15 determines whether or not the i-th term pair exists in the term pair storage unit 11. If the i-th term pair exists, the process goes to step S204, and if it does not exist, the process ends.

  (Step S204) The feature acquisition unit 15 reads the i-th term pair from the term pair storage unit 11.

  (Step S205) The feature acquisition unit 15 performs a process of acquiring the feature of the i-th term pair. The feature acquisition process will be described with reference to the flowchart of FIG.

  (Step S206) The machine learning unit 16 performs supervised machine learning using the plurality of features acquired in step S205 and two or more learning data of the learning data storage unit 12, and acquires a score.

  (Step S207) The output unit 17 determines whether the i-th term pair is a term pair of different notation using the score acquired in Step S206. For example, the threshold value determination unit 172 configuring the output unit 17 reads the threshold value from the threshold value storage unit 171, and if the score acquired in step S206 is greater than or equal to the threshold value, the i-th term pair is notationally different. When it is determined that the term pair is a term pair and the score is equal to or less than the threshold value or smaller than the threshold value, it is determined that the i-th term pair is not a term pair of different notation.

  (Step S208) The output unit 17 goes to Step S209 if the judgment result in Step S208 is a differently-notated term pair, and goes to Step S210 if it is not an unnotated term pair.

  (Step S209) The output unit 17 outputs the i-th term pair as a differently expressed term pair.

  (Step S210) The feature acquisition unit 15 increments the counter i by one. The process returns to step S203.

  Next, the feature acquisition process in step S205 will be described with reference to the flowchart of FIG.

  (Step S301) The difference character acquisition means 151 which comprises the feature acquisition part 15 acquires the edit location of two terms.

  (Step S302) The feature acquisition unit 152 of the feature acquisition unit 15 acquires a character type-related feature using the edited portion acquired in step S301. A specific method for acquiring the character type-related features will be described later.

  (Step S303) The feature acquisition unit 152 acquires a dictionary-related feature using the term dictionary 13. A specific method for acquiring dictionary-related features will be described later.

  (Step S304) The feature acquisition unit 152 acquires the similarity between two terms. This similarity is a similarity feature.

  (Step S305) The feature acquisition unit 152 acquires other predetermined features. Examples of other predetermined features will be described later.

  (Step S306) The feature acquisition unit 152 uses a plurality of features acquired in steps S302 to S305 by using a stacking algorithm, and uses a classification method different from the supervised machine learning method in step S206 to generate a term pair. It is determined whether or not a pair of terms is notated, and the determination result is acquired.

  Hereinafter, a specific operation of the different notation acquisition apparatus 1 in the present embodiment will be described.

  Now, it is assumed that the term dictionary 13 has a structure as shown in FIG. In FIG. 4, one term of information is one record. Each record has attribute values of “term”, “reading”, “part of speech”, “representative notation”, “category”, and “domain”. The term dictionary 13 is, for example, a JUMAN dictionary (see “http://nlp.kuee.kyoto-u.ac.jp/nl-resource/juman.html” [searched on December 13, 2009]). The term dictionary 13 is, for example, a Japanese word net dictionary (see http://nlpwww.nict.go.jp/wn-ja/index.ja.html [searched on December 13, 2009]), It is a variant character dictionary or an EDR electronic dictionary (see http://www2.nict.go.jp/r/r312/EDR/J_index.html [searched on December 13, 2009]). The variant character dictionary is a dictionary having pairs of variant characters. Variant characters are characters that have the same reading and usage, but have different shapes. It is often found in kanji with old and new fonts. For example, “Sawa” and “Sawa” are related to different characters. Note that the variant dictionary is not a variant, but may have a pair of Chinese characters that can be substituted like a variant. Furthermore, the term dictionary 13 may include a dictionary having pairs of kanji characters that can be substituted like variant characters, apart from the variant character dictionary.

  Here, it is assumed that the plurality of features included in the learning data stored in the learning data storage unit 12 and the plurality of features acquired by the feature acquisition unit 15 are 68 types. The 68 features (S1 to S68) will be described below. Hereinafter, as specific examples of the term pairs, the features are illustrated using the term pairs “show window” and “show window”.

  S <b> 1 is “the first notation editing portion”, and “c” in the above specific example. When acquiring the feature S1, the difference character acquisition unit 151 compares the characters while shifting the two terms constituting the term pair one character at a time to obtain an edited portion. For example, the difference character acquisition unit 151 compares the first character “SHI” of “Show window” with the first character “SHI” of “Show window”, determines that they are the same, and determines that the second character is also the same. Then, it is determined that the third character “U” is different from “—”, and the first notation editing portion “U” and the second notation editing portion “—” are acquired.

  S2 is "the second notation editing location", and is "-" in the above specific example. The first notation is the first term that makes up the term pair (eg, “show window”), and the second notation is the second term that makes up the term pair (eg, “show” Window ").

  S3 is “one character in front of the edited part”, and in the above specific example, “o”.

  S4 is “one character behind the edited portion”, and in the above specific example, “U”.

  S5 is “two consecutive characters in front of the edited part”, and in the above specific example, “Sho”.

  S6 is “three consecutive characters in front of the edited part”, and in the above specific example, “Sho”.

  S7 is “the second character in front of the edited part”, and in the above specific example, “S”.

  S8 is “the third character in front of the edited part”, and in the above specific example, “(del)”. (del) indicates that there is no character.

  S9 is “two characters behind the edited part”, and in the above specific example, “Wi”.

  S10 is “three characters behind the edited part”, and in the above specific example, “WIN”.

  S11 is “the second character after the edited part”, and in the above specific example, “I”.

  S12 is “the third character behind the edited portion”, and in the above specific example, “N”.

  S13 is “a character string“ information of S1−information of S2 ””, and in the above specific example, “Woo”.

  S14 is “a character string that is“ information of S3—information of S13 ””, and in the above-described specific example, “Woo-o”.

  S15 is “a character string that is“ information of S5—information of S13 ””, and is “show” in the above specific example.

  S16 is “a character string“ information of S6—information of S13 ””, and in the above specific example, is “show”.

  S17 is “'information of S13-information of S4'”, and in the above specific example, “W-W”.

  S18 is "'information of S3-information of S13-information of S4'", and in the above-mentioned specific example, is "show-woo".

  S19 is “a character string that is“ information of S5—information of S13—information of S4 ””, and in the above specific example, is “show-woo”.

  S20 is “a character string that is“ information of S6—information of S13—information of S4 ””, and is “show-woo” in the above specific example.

  S21 is “a character string that is“ information of S13—information of S7 ””, and in the above specific example, “Woo-shi”.

  S22 is “a character string that is“ information of S3—information of S13—information of S7 ””, and in the above-described specific example, is “shower”.

  S23 is “a character string that is“ information of S5−information of S13−information of S7 ””, and in the above specific example, is “show-city”.

  S24 is “a character string that is“ information of S6—information of S13—information of S7 ””, and in the above specific example, is “show-city”.

  S25 is “a character string that is“ information of S13—information of S8 ””, and is “woo- (del)” in the above specific example.

  S26 is “a character string that is“ information of S3—information of S13—information of S8 ””, and in the above-described specific example, is “show- (del)”.

  S27 is “a character string that is“ information of S5—information of S13—information of S8 ””, and is “show-(del)” in the above specific example.

  S28 is “a character string that is“ information of S6—information of S13—information of S8 ””, and is “show-(del)” in the above specific example. If two terms are given and the edit location can be determined, the feature acquisition unit 152 can acquire the features of S3 to S28 by simple character string processing (operation).

  S29 is “character type of S1”, and in the specific example, “Katakana”. A technique for obtaining a character type (kanji, hiragana, kana, alphabet, etc.) of a character when a character is given is a known technology.

  S30 is “character type of S2”, and in the above specific example, “Katakana”.

  S31 is “S3 character type”, and in the above specific example, “Katakana”.

  S32 is “character type of S4”, and in the specific example, “Katakana”.

  S33 is “character type of S13”, and in the above specific example, “Katakana”.

  S34 is “character type of S14”, and in the above specific example, “Katakana”.

  S35 is “character type of S17”, and in the specific example, “Katakana”.

  S36 is “character type of S18”, and in the above specific example, “Katakana”.

  S37 is “part of speech of S1”, and in the above specific example is “noun”. Here, the part of speech of a character is the part of speech of the term to which the character (here, S1) belongs. For example, morphological analysis is applied to the term, the term is divided into words, and the part of speech information is acquired. The part of speech of the character is the part of speech indicated by the acquired part of speech information.

  S38 is “part of speech of S2”, and in the above specific example is “noun”.

  S39 is “part of speech of S3”, and in the above specific example is “noun”.

  S40 is “part of speech of S4”, and in the above specific example, “noun”.

  S41 is “part of speech of S13”, and in the above specific example, “noun”.

  S42 is “part of speech of S14”, and in the above specific example is “noun”.

  S43 is “part of speech of S17”, and in the above specific example is “noun”.

  S44 is “part of speech of S18”, and in the above specific example is “noun”.

  S45 is “part of speech and position information of S1”, and in the above specific example, “noun, 3”. Here, “3” indicates the third character.

  S46 is “part of speech and position information of S2,” which is “noun, 3” in the above specific example.

  S47 is “part of speech and position information of S3”, and “noun, 2” in the above specific example.

  S48 is “part of speech and position information of S4”, and “noun, 4” in the above specific example.

  S49 is “part of speech and position information of S13”, and “noun, 3” in the above specific example.

  S50 is “part of speech and position information of S14”, and “noun, 2” in the above specific example.

  S51 is “part of speech and position information of S17”, and “noun, 6” in the above specific example.

  S52 is “part of speech and position information of S18”, and “noun, 3” in the above specific example.

  S53 is “similarity of Japanese term pairs”, and in the above specific example, for example, 0.9.

  S54 is information indicating whether or not the representative notation of the JUMAN dictionary of the Japanese term pair matches using the stacking algorithm, and is “1” in the above specific example. That is, here, the classification method different from the supervised machine learning method used by the machine learning unit 16 is to classify the two terms constituting the term pair according to whether or not the representative notations in the JUMAN dictionary match. Is the method. First, a word pair whose edit distance is one character is extracted from a set of words in the JUMAN dictionary. Here, 904612 pairs of words with an edit distance of one character can be extracted. Among them, word pairs (25934 sets) having the same representative notation are taken out. Next, in the JUMAN dictionary, word pairs having the same representative notation are set as positive examples, and word pairs other than those are set as negative examples. Thus, teacher data is created. Next, machine learning using the teacher data as learning data is performed. The teacher data is not limited to the teacher data described above, and other teacher data may be used. In addition, as a feature used in machine learning, a feature obtained by removing the feature of S54 from all the features of the present invention (the features of S1 to S68) is used. Note that other features may be used as the features used in machine learning. In practice, the data to be given the feature of S54 is classified using the learning result. Information indicating whether the classification result is a positive example or a negative example is added to the data as the feature of S54. In the method based on the stacking algorithm, the result of learning whether or not the representative notation matches in the JUMAN dictionary can be given as a feature. Therefore, the term pairs that are not actually described in the JUMAN dictionary can be represented by the JUMAN dictionary. Information indicating whether or not the term pair has a tendency to match the notation can be given.

  S55 is “when the number of characters in the pair of Japanese terms is the same and both editing locations are numbers, whether they are the same value or different values”. In the above specific example, it is “0”. In the case of the term pair “secondary cache” and “secondary cache”, the feature of S55 is “1” in the case of the term pair “per week” and “per week”. It should be noted that the condition that the number of characters is the same is removed, and S55 is preferably “whether the edited portions of the Japanese term pair are both numbers, and are the same value or different values”.

  S56 is “when the number of characters in the Japanese term pair is the same and both of the edited portions are hiragana, whether they are the same voice or different voices”. In the above specific example, it is “0”. In the case of the term pairs “okachan” and “okachan”, the feature of S56 is “1”. It should be noted that the condition that the number of characters is the same is removed, and S56 is preferably “whether the edited portion of the Japanese term pair is both hiragana and is the same speech or different speech”.

  S57 is “when the number of characters in the Japanese term pair is the same and both editing portions are in katakana, whether they are the same voice or different voices”. In the above specific example, “1”. In the case of the term pair of “olive oil” and “olive oil”, the feature of S57 is “1” also in the case of the term pair of “window” and “window”. It should be noted that the condition that the number of characters is the same is removed, and S57 is preferably “whether the edited portion of the Japanese term pair is both katakana and whether the speech is the same or different”.

  S58 is “whether the number of characters in the pair of Japanese terms is the same and both edited portions are Roman characters, whether only uppercase and lowercase characters are different”. In the above specific example, “0”. In the case of the term pair of “300 kbps” and “300 kbps”, the feature of S58 is “1” in the case of the term pair “on Windows” and “on windows”. It should be noted that the condition that the number of characters is the same is removed, and S58 is preferably “whether the edited portions of the Japanese term pairs are both in Roman letters and only the difference between uppercase and lowercase letters”. “Windows” is a registered trademark.

  S59 is “whether the number of characters in the Japanese term pair is the same, and if one edit location is marked with a cloud point, it becomes another edit location”. In the above specific example, it is “0”. In the case of a term pair of “about touch” and “about touch”, in the case of a term pair of “about drink” and “about drink”, the feature of S59 is “1”. In S59, it is preferable to remove the condition that the number of characters is the same, and “whether or not the editing point of one of the Japanese term pairs becomes the other editing point”.

  S60 is “whether the number of characters in the Japanese term pair is the same and if one editing place is given a semi-turbid point, it becomes another editing place”, and in the above specific example, “0”. In S60, it is preferable to remove the condition that the number of characters is the same, and "whether a semi-turbid point is added to one editing portion of a Japanese term pair to become another editing portion".

  S61 is “The editing part is only one of the pair of Japanese terms, and the editing part is' Correct ',' System ',' Class', 'Type', 'Form', 'Mr.', '-', ' It is “whether” or “0” in the above specific example. In the case of the term pairs “server” and “server”, in the case of the term pairs “Hansen's disease patient” and “Hansen's disease patient”, in the case of the term pairs “Japanese patch” and “Japanese patch”, “30 types” In the case of a term pair of “so” and “about 30”, the feature of S61 is “1”.

  S62 is “whether the edited portion is only in one of the Japanese term pairs and the term of the edited portion matches the last character of the Japanese term pair”, and is “0” in the above specific example. In the case of “pregnant / nursing” and “pregnant / nursing” term pairs, in the case of “domestic / imported vehicles” and “domestic vehicles / imported vehicles”, the feature of S62 is “1”. .

  S63 is “whether the edited part is only in one of the Japanese term pairs and the edited part is a term indicating the number of digits (for example,“ thousand ”“ ten thousand ”, etc.). It is. In the case of the term pairs “25,000” and “25000”, the feature of S63 is “1” in the case of the term pairs of “16,500 yen” and “16,500 yen”.

  S64 is “whether or not the representative notation defined in the JUMAN dictionary of the Japanese term pair matches.” In the above specific example (term pair “show window” “show window”), for example, “1”. . The feature acquisition unit 152 acquires the feature by acquiring the representative notation of each term constituting the term pair from the term dictionary 13 and comparing it.

  S65 is “whether the Japanese term pair is defined as a synonym pair in the Japanese word net dictionary”. In the above specific example (the term pair “show window” “show window”), for example, “1”. . The feature acquisition unit 152 searches the Japanese word net dictionary using each term constituting the term pair as a key, and determines whether the term pair is defined as a synonym pair. This process is a normal search process.

  S66 is “whether the edited part of the Japanese term pair is defined as a variant character in the variant dictionary.” In the above specific example (term pair “show window” “show window”), for example, “0”. is there. The variant dictionary has information on two variant pairs.

  S67 is “whether the number of characters in the Japanese term pair is the same and the edited part is kanji and hiragana and the readings in the JUMAN dictionary match”. ) Is “0”.

  S68 is “whether the number of characters in the pair of Japanese terms is the same and both of the editing parts are kanji and whether the readings of the JUMAN dictionary match.” The above specific example (term vs. “show window” and “show window”) ) Is “0”.

  Further, the above 68 features include a character type related feature, a dictionary related feature, a similarity feature, an edited portion character feature, and the like as described above.

  Further, when the above 68 features are grouped, for example, it can be considered that they are divided into the following groups G1 to G7.

  G1 is a feature from S1 to S52, and is an edited portion character feature that is information about the edited portion and the surrounding character string.

  G2 is a feature of S53 and a similarity feature.

  G3 is a feature of S54 and is a feature that is information using a stacking algorithm.

  G4 is the feature of S55 to S60, and is an edit location related feature that is information related to the edit location.

  G5 is a feature of S61 to S63, and is a term pair pattern feature which is information on a term pair pattern.

  G6 is a feature from S64 to S66, and is a dictionary-related feature that is information by various dictionaries.

  G7 is a feature from S67 to S68, and is a reading-related feature that is information related to reading.

  First, in the different notation acquisition apparatus 1, data of correct correct notation term pairs (positive examples) is manually constructed in advance, and positive and negative information (for example, positive examples) indicating positive term pairs and positive examples (for example, , “1”) are stored in the learning data storage unit 12 in association with each other. Further, in the different notation acquisition device 1, data of a term pair that is not different notation (negative example) is manually constructed in advance, and the negative term pair and positive / negative information (for example, “0” indicating that it is a negative example) ]) In association with each other and stored in the learning data storage unit 12.

  Next, the above-described 68 features of each term pair are acquired by the feature acquisition unit 15 of the different notation acquisition device 1, and the 68 features are stored in the learning data storage unit 12 in association with the positive / negative information or the term pair. To do.

  Through the above processing, learning data in the learning data storage unit 12 is constructed.

  Next, the term pair storage unit 11 stores one or more term pairs for which it is desired to determine whether the term pairs are different notations.

  Then, the user inputs an operation start instruction to the different notation acquisition apparatus 1. Then, the reception unit 14 receives an operation start instruction.

  Next, the term pairs stored in the term pair storage unit 11 are sequentially processed as follows. That is, the feature acquisition unit 15 performs a process of acquiring the feature of each term pair. Such a feature acquisition process has already been described.

  Next, the machine learning unit 16 performs supervised machine learning using the acquired 68 features and the learning data stored in the learning data storage unit 12 to obtain a score.

  Next, the output unit 17 determines whether each term pair is a term pair of different notation using the acquired score of each term pair.

Next, the output unit 17 outputs only the term pairs that are determined to be the differently expressed term pairs. Here, the output is, for example, accumulation in a predetermined storage medium.
(About machine learning)

  Hereinafter, the machine learning performed by the machine learning unit 16 and the classification method (used in the stacking algorithm) different from the supervised machine learning method performed by the machine learning unit 16 will be described.

  First, the machine learning method prepares many sets of problem-solution pairs, learns them, learns what kind of solution the problem becomes, and uses the learning results to create a new problem. In some cases, the solution can be estimated (see, for example, Non-Patent Document 2 to Non-Patent Document 4).

  In order to convey the problem situation to the machine, what kind of solution is to be solved, it is necessary to have a feature (elements constituting the problem with information used for analysis). The problem is expressed by the feature.

  In other words, the machine learning method prepares many sets of feature set-solution pairs, learns with it, learns what kind of feature set the solution will be, and uses the learning results. This is a method of extracting a set of features from a new problem and inferring a solution in the case of the feature.

  As a machine learning method, for example, a k neighborhood method, a simple Bayes method, a decision list method, a maximum entropy method, a support vector machine method, or the like can be used.

  The k-nearest neighbor method is a method for obtaining a classification destination (solution) by using the k most similar cases instead of the most similar case, and by majority decision of the k cases. k is a predetermined integer number, and generally an odd number between 1 and 9 is used.

  The Simple Bayes method is a method of estimating the probability of each classification based on Bayes' theorem and determining the classification having the highest probability value as a classification destination.

In the simple Bayes method, the probability of outputting the classification a in the context b is given by the following formula 1.

Here, the context b is a set of features f _j (εF, 1 ≦ j ≦ k) set in advance. p (b) is the appearance probability of the context b. Here, since it is independent of the classification a and is a constant, it is not calculated. P (a) (where P is a tilde at the top of p) and P (f _i | a) are probabilities estimated from teacher data (same meaning as judgment information), respectively, It means the probability of having a feature f _i for classification a. When the value obtained by performing maximum likelihood estimation as P (f _i | a) is used, the value is often zero, and it is difficult to determine the classification destination because the value of the expression in the second row of Expression 2 is zero. Occurs. Therefore, smoothing is performed. Here, smoothed data using the following formula 2 is used.

Here, freq (f _i , a) means the number of cases having the feature f _i and the classification a, and freq (a) means the number of cases having the classification a.

  The decision list method uses features and combinations of classification destinations as rules, stores them in the list in a predetermined priority order, and when input to be detected is given, from the highest priority in the list This is a method in which input data is compared with the feature of the rule, and the classification destination of the rule having the same feature is set as the classification destination of the input.

In the decision list method, the probability value of each classification is obtained using only one of the features f _j (εF, 1 ≦ j ≦ k) set in advance as a context. The probability of outputting classification a in a context b is given by Equation 3 below.

However, f _max is given by Equation 4 below.

P (a _i | f _j ) (where P is a tilde at the top of p) is the rate of appearance of the classification a _i when the feature f _j is in the context.

In the maximum entropy method, when a set of preset features f _j (1 ≦ j ≦ k) is F, an expression (expression) that represents entropy while satisfying the following conditional expression (expression 5): 6) This is a method of obtaining a probability distribution p (a, b) when maximizing 6) and determining a classification having the largest probability value among the classification probabilities obtained according to the probability distribution.

However, A and B mean a set of classifications and contexts, and g _j (a, b) is a function that is 1 if the context b has a feature f _j and the classification is a, and is 0 otherwise. means. Further, P (a _i | f _j ) (where P is a tilde at the top of p) means the rate of appearance of (a, b) in the known data.

  Formula 5 is to obtain the expected value of the frequency of the pair of output and feature by multiplying the probability p and the function g meaning the appearance of the pair of output and feature. With the constraint that the expected values calculated based on the probability distribution obtained on the left side are equal, entropy maximization (smoothing of the probability distribution) is performed to obtain the probability distribution of the output and the context. Details of the maximum entropy method are described in the following references (1) and (2).

  Reference (1): Eric Sven Ristad, Maximum Entropy Modeling for Natural Language, (ACL / EACL Tutorial Program, Madrid, 1997

  Reference (2): Eric Sven Ristad, Maximum Entropy Modeling Toolkit, Release 1.6beta, (http://www.mnemonic.com/software/memt,1998))

  The support vector machine method is a method of classifying data composed of two classifications by dividing a space by a hyperplane.

  FIG. 5 shows the concept of margin maximization in the support vector machine method. In FIG. 5, a white circle means a positive example, a black circle means a negative example, a solid line means a hyperplane that divides the space, and a broken line means a surface that represents the boundary of the margin area. 5A is a conceptual diagram when the interval between the positive example and the negative example is narrow (small margin), and FIG. 5B is a conceptual diagram when the interval between the positive example and the negative example is wide (large margin). is there.

  At this time, if the two classifications consist of positive examples and negative examples, the larger the interval (margin) between the positive examples and the negative examples in the learning data, the lower the possibility of incorrect classification with open data. As shown in FIG. 5B, a hyperplane that maximizes this margin is obtained, and classification is performed using it.

  Basically, it is as described above. Usually, an extension of the method that the training data may contain a small number of cases in the inner area of the margin, or an extension that makes the linear part of the hyperplane nonlinear ( The one with kernel function introduced) is used.

This extended method is equivalent to classification using the following discriminant function (f (x)), and the two classes can be discriminated depending on whether the output value of the discriminant function is positive or negative.

Where x is the context (set of features) to be identified, and x _i and y _j (i = 1,..., L, yj∈ {1, −1}) mean the context and classification destination of the learning data. The function sgn is
sgn (x) = 1 (x ≧ 0)
-1 (otherwise)
In addition, each α _i is for maximizing Expression (8-1) under the constraints of Expressions (8-2) and (8-3) of Expression 8.

The function K is called a kernel function, and various functions are used. In this embodiment, for example, the following polynomial (formula 9) is used.

In Expressions 8 and 9, C and d are constants set experimentally. For example, C was fixed at 1 throughout all treatments. Moreover, two types of 1 and 2 are tried for d. Here, x _i satisfying α _i > 0 is called a support vector, and the portion taking the sum of Expression 7 is usually calculated using only this case. That is, only actual cases called support vectors are used for actual analysis.

  Details of the extended support vector machine method are described in the following references (3) and (4).

  Reference (3): Nello Cristianini and John Shawe-Taylor, An Introduction to Support Vector Machines and other kernel-based learning methods, (Cambridge University Press, 2000)

  Reference (4): Taku Kudoh, Tinysvm: Support Vector machines, (http://cl.aistnara.ac.jp/taku-ku//software/Tiny SVM / index.html, 2000)

  The support vector machine method handles data with two classifications. Therefore, when handling cases with three or more classifications, a pair-wise method or a one-VS rest method is usually used in combination with this.

  In the pairwise method, in the case of data having n classifications, every pair (n (n-1) / 2) of two different classification destinations is generated, and a binary classifier indicates which is better for each pair. That is, it is obtained by the support vector machine method processing module and finally obtains the classification destination by majority decision of the classification destination by n (n−1) / 2 binary classification.

  For example, when there are three classification destinations a, b, and c, the one VS rest method generates three sets of classification destination a and other, classification destination b and other, classification destination c and other, The learning process is performed on the set of the support vector machine method. Then, in the estimation process based on the learning result, the learning results of the three sets of support vector machines are used. See how the problem to be estimated is estimated in the three support vector machines, and among the three support vector machines, which is the other classification destination, and the separation plane of the support vector machine This is a method for obtaining a classification destination of a thing farthest from the object. For example, when a problem to be solved is farthest from the separation plane in the support vector machine created by the learning process of “classification destination a and others”, the classification destination of the problem to be solved is estimated as a. .

  The method of obtaining the degree of the solution (classification destination) that the machine learning unit 16 estimates about the problem to be solved is different depending on various methods used by the machine learning unit 16 as a machine learning method.

  For example, in the embodiment of the present invention, when the machine learning unit 16 uses the k-nearest neighbor method as the machine learning method, the machine learning unit 16 sets the feature data extracted from the examples of the cases of the teacher data. The similarity between cases based on the ratio of overlapping features (the number of the same features) is defined, and the defined similarity and the case are stored in the learning data storage unit 12 as learning result information Keep it.

  Then, the machine learning unit 16 refers to the stored similarity and case when the feature of the problem to be solved is extracted by the feature acquisition unit 15, and the feature of the problem to be solved extracted by the feature acquisition unit 15 For k, the k cases are selected in descending order of the similarity of the features of the problem to be solved, and the classification destination determined by the majority vote in the selected k cases is estimated as the classification destination (solution) of the problem to be solved. . That is, in the machine learning unit 16, the degree of the kind of solution (classification destination) that is likely to be the problem to be solved is set as the number of votes of majority vote in the selected k cases.

  When the simple Bayes method is used as a machine learning method, a set of a solution of the case and a set of features is stored as learning data in the learning data storage unit 12 for the case of the teacher data. Then, when the feature acquisition unit 15 extracts the feature of the problem to be solved, the machine learning unit 16 uses the Bayes' theorem based on the combination of the judgment information solution and the feature set in the learning data storage unit 12. Based on the above, the probability of each classification in the case of a set of the features of the problem to be solved acquired by the feature acquisition unit 15 is calculated, and the classification having the highest probability value is classified into the classification of the features of the problem to be solved (solution ). That is, in the machine learning unit 16, the degree of ease of becoming a solution in the case of a set of features of the problem to be solved is set as the probability of being classified.

  Further, when the decision list method is used as the machine learning method, a list in which rules of features and classification destinations are arranged in a predetermined priority order with respect to the example of the teacher data is previously stored by the learning data storage unit 12 by some means. Remember me. When the features of the problem to be solved are extracted by the feature acquisition unit 15, the machine learning unit 16 extracts the features and rules of the extracted problem to be solved in descending order of the priority of the list in the learning data storage unit 12. The features are compared with each other, and the classification destination of the rule having the same feature is estimated as the classification destination (solution) of the problem to be solved.

  In addition, when using the maximum entropy method as a machine learning method, the classification that can be a solution is identified from the example of the teacher data, the predetermined conditional expression is satisfied, and the feature when maximizing the expression showing entropy is determined. A probability distribution composed of two terms of a set and a class that can be a solution is obtained and stored in the learning data storage unit 12. When the features of the problem to be solved are extracted by the feature acquisition unit 15, the machine learning unit 16 uses the probability distribution of the learning data storage unit 12 to extract the set of features of the extracted problem to be solved. A probability of a class that can be a solution is obtained, a class that can be a solution having the largest probability value is specified, and the specified class is estimated as a solution of the problem to be solved. That is, in the machine learning unit 16, the degree of ease of becoming a solution in the case of a set of features of the problem to be solved is set as the probability of being classified.

In addition, when using the support vector machine method as a machine learning method, the classification that can be a solution is identified from the example of the teacher data, the classification is divided into positive examples and negative examples, and predetermined execution using a kernel function is performed. Stores learning data by finding the hyperplane that maximizes the interval between the positive and negative examples of the case and divides the positive and negative examples by the hyperplane in a space whose dimension is a set of case features according to the function Store in unit 12. When the features of the problem to be solved are extracted by the feature acquisition unit 15, the machine learning unit 16 uses the hyperplane of the learning data storage unit 12 to divide the set of features of the problem to be solved by the hyperplane. In the specified space, it is specified whether it is on the positive example side or the negative example side, and the classification determined based on the specified result is estimated as the solution of the problem to be solved. That is, in the machine learning unit 16, the degree of ease of becoming a solution in the case of a set of features of the problem to be solved is set as the magnitude of the distance from the separation plane to the case of the problem to be solved.
(Experimental result 1)

  Hereinafter, the experimental result of the different notation acquisition apparatus 1 will be described. First, data used for the experiment will be described.

  The data used in the experiment is a large-scale similar word list. The large-scale similar word list is about 100 million pages of the search engine research base TSUBAKI (see http://tsubaki.ixnlp.nii.ac.jp/se/index.cgi [searched on December 13, 2009]) • One million words are extracted from 6 billion sentence data, and a maximum of 500 synonyms are generated with similarities for each of the one million words. One million Japanese terms included in this large-scale similar word list and Japanese terms corresponding to the respective Japanese terms are taken as Japanese term pairs.

  Then, 14185 Japanese word pairs with an edit distance of 1 were randomly extracted from the large-scale similar word list. Tagging whether the extracted Japanese term pairs are Japanese variant pairs or not was performed by a majority vote of three evaluators. It is determined using the Kappa statistic K how much the tagging of the three evaluators matches.

  There is a feature of different information depending on whether a combination of a Japanese term pair having two terms of the right term and the left term is on the left or the right. In order to increase the amount of information corresponding to the feature, data obtained by changing the combination of Japanese term pairs is used. That is, in this experiment, a total of 28370 sets of experimental data are used in addition to 14185 sets extracted from the large-scale similarity list.

  Also, assuming that 28370 sets of experimental data are a single set of data, the fairness of the experiment may be lost, and half of the 28470 sets of experimental data are used as data for devising features. The remaining half of the experimental data is used as data for examining whether or not the feature devised by the closed data is valid for other data. Data that devise features is called closed data. The data to be considered is called open data. The accuracy of the closed data is measured by learning by 10 division cross validation. In the 10-part cross validation, the test target data is divided into the first to tenth, and the following learning (1) to (10) is performed. (1) Learning is performed using the first as test data and the second through tenth as learning data. (2) Learning is performed using the second as test data and the first, third to tenth as learning data. (3) Learning is performed with the third as test data and the first, second, fourth to tenth as learning data. (4) Learning is performed with the fourth as test data and the other as learning data. (5) Learning is performed with the fifth as test data and the other as learning data. (6) Learning is performed using the sixth as test data and the other as learning data. (7) Learning is performed using the seventh as test data and the other as learning data. (8) Learning is performed using the eighth as test data and the other as learning data. (9) Learning is performed using the ninth as test data and the other as learning data. (10) Learning is performed with the tenth as test data and the other as learning data. In addition, 10 division | segmentation cross validation is a well-known technique.

  In addition, closed data is used as learning data (data stored in the learning data storage unit 12), and open data is used as test data (data for determining whether or not a term pair is different). Measure.

  FIG. 6 shows a breakdown of the Japanese term pairs having an editing distance of 1 used in the experiment, which are determined by majority voting as a Japanese different notation pair or not a Japanese different notation pair.

  The kappa statistic K is a numerical value representing the degree of coincidence in the category rating of K person raters. Since the method for calculating the kappa statistic K is known, the description thereof is omitted.

  If there is a perfect match between the graders, the value of K will be 1. If the chance level matches, the value of K is 0. If the degree of coincidence increases, the value of K approaches 0 to 1. FIG. 7 shows a method for evaluating the degree of coincidence according to Landis et al. In this experiment, Kappa statistic K was calculated for two categories of 14185 pairs of Japanese terminology, which were either Japanese variant pairs or not Japanese variant pairs. .84. Since this is a degree of coincidence of 0.8 or more, it can be evaluated as almost perfect coincidence.

  Next, in order to show that the method for determining whether or not a different pair of terminology is used in the different notation acquisition device 1 is described, the determination method of the different notation acquisition device 1 and the baseline method as a comparison are explained. To do.

The baseline method applies the following rules to the Japanese terminology pairs that are subject to the extraction of Japanese variant pairs with a short edit distance (for example, the edit distance is 1), and mechanically uses the Japanese variant. A determination is made as to whether it is a pair or not a Japanese variant pair.
(Rule 1) If the edited portion of a Japanese term pair having the same number of characters represents a number having the same value, it is determined that the pair is a Japanese variant notation.
(Rule 2) If the edited portion of a Japanese term pair having the same number of characters represents the same alphabet, it is determined that the pair is a Japanese variant.
(Rule 3) Japanese term pairs having the same number of characters can be read using JUMAN, and if the readings match, it is determined to be a Japanese variant notation pair.
(Rule 4) If the rule does not match Rule 1, Rule 2, and Rule 3, it is determined that the pair is not a Japanese variant.

  In the baseline method applying rule 1 to rule 4 above, the Japanese term pair “second edition” and “second edition” are determined as follows. The edited parts in this Japanese term pair are “2” and “2”, which represent numbers. Since Rule 1 is applied to represent the same value, the baseline method determines that the Japanese term pair is a Japanese variant notation pair.

  The Japanese term pair “Tea” and “tea” are determined as follows. The edited parts in this Japanese term pair are “T” and “t”, which represent the alphabet. Then, rule 2 is applied, and since this Japanese term pair is a word having the same meaning, it is determined that this Japanese term pair is a Japanese notation pair.

  Japanese terms vs. “praise” and “praise” are judged as follows. This Japanese terminology pair can be read using JUMAN for “praise” and “praise”. Then, rule 3 is applied, the readings are the same, and it is determined that this Japanese term pair is a Japanese variant notation pair.

  The Japanese terms “silky white” and “milky white” are judged as follows. The edited parts in this Japanese term pair are “shi” and “mi”, representing katakana. Since Silky White and Milky White are undefined in the JUMAN dictionary, how to read them cannot be checked. Therefore, rule 4 is applied and it is determined that the pair is not a Japanese variant.

  Next, the machine learning method of the machine learning unit 16 used in this experiment will be described in detail. This machine learning method is a support vector machine method. As described above, the support vector machine method is a method of classifying data composed of two classifications by dividing a space by a hyperplane. At this time, if the two classifications are a positive example and a negative example, it is determined that the larger the interval between the two in the learning data, the lower the possibility of erroneous classification. It is fundamental to obtain a hyperplane that maximizes this interval and classify it by obtaining it. However, here, a method in which a small number of cases may be included in the inner region of the interval in the learning data or an extension such as making the linear portion of the hyperplane nonlinear is used. These extended methods are equivalent to classification using a discriminant function, and can classify into two depending on whether the output value of the discriminant function is positive or negative. Further, when handling data consisting of three or more, the pair-wise method is used in parallel. In the case of data with N classifications, the pairwise method creates every pair of two different classification destinations, finds which one is better for each pair by a binary classifier (support vector machine method), and finally determines the classification destination It is a method to seek by majority vote. Hereinafter, the support vector machine method is appropriately expressed as SVM.

  The features used to extract Japanese variant pairs with a short edit distance (for example, the edit distance is 1) by the SVM are the features of S1 to S68 described above. These features are extracted from pairs of Japanese terms with a small editing distance that are randomly extracted from the large-scale similar word list. Each machine learning determines whether a Japanese term pair is a Japanese variant pair or a Japanese variant pair depending on the feature. Various features were used to obtain as much information as possible from the Japanese term pairs. Each feature can be classified into G1 and G7 as described above. G1, G2, and G3 are features that can correspond to all pairs of Japanese terms having a small editing distance. The character type is information indicating whether the target character represents hiragana, katakana, numbers, alphabet, or any other type. Part of speech is subjected to morphological analysis using JUMAN as a term, and the term is divided into words to obtain part of speech information. This is information about which part of speech the target character belongs to. The position information is information indicating which position of the part of speech is the first, last, or other part of the part of speech when the target character belongs to the part of speech. The similarity is information on the similarity used when generating the large-scale similar word list.

  As described above, the stacking algorithm is to learn, by machine learning, data obtained by classifying experimental data by a classification method different from the original purpose, and to add classification information of a learning result to a feature. In this experiment, the data used for the stacking algorithm uses 904612 pairs of Japanese terms that can discriminate the representative representation of JUMAN obtained from the large-scale similar word list other than 28370 pairs of Japanese terms used in the experiment. . Among the 904612, the positive example is 25934, and the negative example is 878678. Thereby, although it is approximate, the information of the feature of S64 can be given also to a Japanese term pair undefined in the JUMAN dictionary. G4 and G5 are features specialized for a Japanese term pair having a characteristic and a small editing distance. G4 is an equal character string by substitution, and G5 is a Japanese term pair that becomes an equal character string by deletion. G6 and G7 are features using a JUMAN dictionary, a Japanese word net dictionary, and an EDR dictionary. Regarding the JUMAN dictionary, feature information is not given to Japanese term pairs in which undefined terms appear.

The effectiveness of the above-described features S1 to S68 will be examined by analyzing a significant difference. Significant consideration is determined using the bootstrap method. The bootstrap method uses data classified into two by a classification method. Each of the two classified data is extracted at random without changing the number of data (for example, the number of questions is 1400) (for example, the number of extracted data is 1400). Each F value is obtained from the extracted data, and each F value is compared. The process of taking out and comparing the F value is repeated 10,000 times. The F value is defined by Equation 2 below. That is, the harmonic average of recall and precision.

  If the number of times that the F value of the other method is higher than the F value of the other method is 9500 times (95%) or more than the F value of the one method, the F value of the significance level is 5%. It can be said that the method with the higher is significant. If neither of them is 9500 times (95%) or more, it cannot be determined whether the significance level is 5%. Here, a significance level of 5% is applied, but for example, a significance level of 10% may be applied.

  In this experiment, all features and data obtained by removing only one type of feature from all features are compared by the learning result of SVM. All the 68 features described above are set as the total features, and the features to be removed are performed in all the features in S1 to S68. This significant difference is examined based on the learning result by SVM with 10-division cross validation (10 CV) for closed data and the learning result by SVM with open data (OC) using closed data and open data. Note that “open / close” refers to an experiment in which closed data is used as learning data and open data is used as test data. Hereinafter, the SVM experiment by 10-fold cross validation is described as 10 CV, and the SVM experiment by open close is expressed as OC.

  Next, we report the results of experiments on the baseline method and the method using machine learning.

  In this experiment, it was determined whether the 28370 pairs of Japanese terms with a short editing distance included in the large-scale similar word list described above are Japanese variant pairs or not. FIG. 8 shows the result of applying the baseline method to the prepared closed data and open data. FIG. 8 also shows the results of 10 CV and OC. TinySVM was adopted as the SVM implementation used in the experiment, and the soft margin parameter C was set to 1 with a first-order polynomial kernel. “Total features” in each table indicates an experiment using all the features of S1 to S68, and “Feature selection” indicates an experiment using all features other than the omitted features.

  The correct answer rate in FIG. 8 is a ratio of correctly determining whether each of the experimental data is a Japanese variant pair with an edit distance of 1 or a Japanese variant pair with an edit distance of 1. The F value in FIG. 8 is an F value in the case of extracting a Japanese variant notation pair whose edit distance is 1 for each experimental data. It can be seen that, for 10 CV and OC, the accuracy rate and F value of SVM using all features are higher than those of the baseline method. The F-value when extracting a Japanese variant pair from a Japanese term pair with an edit distance of 1 is 0.433 higher for 10 CV and 0 for open data in the SVM using the total feature compared to the baseline. .460 was expensive. From the results of the baseline, it can be said that it is difficult to extract Japanese grammar pairs from Japanese term pairs with an edit distance of 1. However, the methods using various features and machine learning proposed in this report are: It can be seen that more Japanese variant pairs can be extracted than the baseline method.

  Next, FIG. 9 shows the result of examining whether the feature is effective using the bootstrap method described above. The features from S1 to S68 are shown in FIG. In FIG. 9, the omitted features are features that are determined to be significant by the bootstrap method. The total feature is a technique based on the feature handled in this experiment, and the feature selection is a technique based on the feature that omits the omitted feature from the total feature. Each value is the number of times that the total feature has an F value higher than the feature selection, or the number of times that the feature selection has an F value higher than the total feature. In this experiment, if the number of times that the total feature is higher than the feature selection is 9500 times (95%) or more, the omitted feature is useful for improving accuracy, and the omitted feature is effective. . The features whose total feature is 9500 times (95%) or more when the F value is higher than the feature selection are S47, S55, S58, S67 in the case of 10 CV, and S52, S54, S55, S58 in the case of OC. , S67. In both 10CV and OC, the features whose total feature number was higher than the feature selection by 9500 times (95%) were S55, S58, and S67. From this result, it can be said that the features of S55, S58, and S67 are effective features for data that extracts pairs of Japanese idioms having a small edit distance. Although the features of S47 and S52 can be said to be effective for the data used in each experiment, it can be said that if the data of a Japanese term pair with a short editing distance changes, it may become ineffective. Therefore, it cannot be said that S47 and S52 are effective for data for extracting a pair of Japanese notation having a small edit distance.

  Next, a comparison is made using various synonym dictionaries to determine whether the proposed method of the different notation acquisition device 1 (hereinafter also simply referred to as the proposed method) can extract a Japanese different notation pair whose edit distance is 1. The results will be described. Various synonym dictionaries are an EDR dictionary, a Japanese word net dictionary, and a JUMAN dictionary. It was found that there are 21224779 sets of Japanese terms in the EDR dictionary, 890616 sets in the Japanese word net dictionary, and 23348 sets in the JUMAN dictionary, with the edit distance of 1. The EDR dictionary has words related to personal names. In this experiment, it was judged that the person's name was not a synonym and was removed. As a result, the number of Japanese term pairs whose edit distance is 1 included in the EDR dictionary is 933037 sets. The JUMAN dictionary treated word pairs with the same representative notation as Japanese term pairs.

  FIG. 11 shows the examination results of the ratios of the term pairs classified as Japanese variant pairs whose editing distance is 1 from the large-scale similar word list using the proposed method in various dictionaries. . Hereinafter, all Japanese term pairs having an editing distance of 1 in the large-scale similar term list will be referred to as a Japanese term pair database, which will be referred to as term pair DB. Further, in the term pair DB, all the Japanese term pairs classified by the proposed method as Japanese different notation pairs are referred to as a Japanese different notation pair database and denoted as different notation DB. In the term pair DB, all of the non-Japanese term pairs classified by the proposed method as not being Japanese non-differential pairs are referred to as a non-Japanese different notation pair database and referred to as a non-different notation DB. The EDR dictionary contained 20.45% of the Japanese word net, 1.71% of the Japanese word net, and 6.52% of the JUMAN dictionary. In any dictionary, the proportion of Japanese term pairs in the different notation DB is not high. From these results, it is understood that the different notation obtained by the different notation acquisition device described in this specification and the existing dictionary have a small overlap, so that the different notation acquisition device can add many different notations to the existing dictionary. . Further, for example, in the EDR dictionary, the coverage rate is about 20%, but it can be said that the coverage rate is considerably high.

  In addition, Japanese term pairs with an edit distance of 1 included in various dictionaries are classified according to the proposed method as either Japanese variant pairs with an edit distance of 1 or Japanese variant pairs with an edit distance of 1. . The accuracy rate in the SVM classification is shown in FIG.

  Also, in various dictionaries and term pair DBs, five pairs of Japanese term pairs that are classified as Japanese variant pairs with an edit distance of 1 and Japanese term pairs that are not classified are randomly extracted. The results are shown in FIG. The learning data is data combining open data and closed data. The test data is classified into the term pair DB and various dictionaries, and the term pair DB and various dictionaries are classified by OC.

  In FIG. 12, the reason why the correct rate of classification is low in the Japanese word net is that, as shown in FIG. As shown in FIG. 12, the JUMAN dictionary can be classified with a high accuracy rate of 80%. In addition, the JUMAN dictionary has few problems that include non-Japanese grammar pairs, and because the proposed method can properly extract different grammars, it can achieve a high accuracy rate of 80%. Conceivable.

Next, we will describe the evaluation of extracting Japanese variant pairs with an edit distance of 1. The SVM can also classify data depending on whether the output value of the discrimination function (score output by the machine learning unit 16) is positive or negative. Here, depending on whether the output value of the discrimination function is positive or negative, Rather than classifying data, positive or negative data is classified by an arbitrary value (threshold value), and an evaluation of Japanese variant notation pair extraction with an edit distance of 1 is performed. That is, the threshold determination unit 172 determines whether the score acquired by the machine learning unit 16 is equal to or greater than or greater than the threshold stored in the threshold storage unit 171. The method that classifies positive data as Japanese variant pairs with an edit distance of 1 and classifies negative data as non-Japanese pairs with an edit distance of 1 is not 100% accurate. Thus, a Japanese term pair that is not a Japanese variant pair with an edit distance of 1 may be determined as a Japanese variant pair and extracted. For this reason, if it is desired to reliably extract even a small amount, it is possible to reliably extract Japanese variant notation pairs by setting a high threshold value. In addition, even if erroneous data is included, if it is desired to perform exhaustive extraction, it is possible to set a low threshold value. FIG. 14 shows threshold evaluation criteria. As shown in FIG. 14, it was found that the highest value of F value 0.9323 can be obtained by setting the threshold value to -0.2. Further, the ratio between the recall rate and the matching rate is shown in FIG. According to FIG. 15, if the reproduction rate is to be increased, the coverage rate must be increased and the matching rate is lowered. If the precision ratio is to be increased, the coverage ratio must be lowered, and the recall ratio is lowered.
(Experimental result 2)

  In the second experiment, 725 positive examples from learning data having 745 positive examples (for example, “spaghetti” and “spaghetti” pairs) and 13,440 negative examples (non-positive examples) 13,460 negative test data were extracted, and the F value was 0.93. In the experimental result 2, the features and learning data used in the experimental result 1 do not completely match, but there is a duplication of features and learning data that is sufficient to show the effectiveness of the proposed method.

  The F value was about 0.0972 in the case of a baseline method in which all were judged to be positive examples, that is, whatever was a positive example. Even in the conventional method using only the character of the character at the edit location or the character around the character at the edit location, the F value is 0.85 in the pair of terms to be judged whether they are different notations. That is, it can be seen that the effect of the method using a large number of features (here, 68) as in the proposed method is remarkable.

  In addition, features based on existing different notation dictionaries, and those using these methods may have significantly higher F values than methods based on features based on stacking techniques (the above dictionary-related features). We are confirming the effectiveness of these methods.

  In addition, as a rule-based method, it is possible to determine that the character type of the edited portion is a Chinese numeral or an Arabic numeral, or the same alphabet, or that an existing different notation dictionary is used. I also tried a method to estimate things definitively. The F value in this case is 0.4202, indicating that it is better to use supervised machine learning instead of the rule-based method.

A difference pattern that tends to be different notation is learned from the difference data of the correct different notation, and a candidate B for different notation is generated from the above difference pattern for a certain term A, and the terms A and B are different notations. By using the operation of determining whether it is a pair or not, there is an effect that the number of different expressions that can be acquired increases significantly. Such an operation will be described in the second embodiment.
(Experimental result 3)

In the third experiment, 100,000 words and 100 similar words were used. Of all pairs of 100,000 words and 100 similar words, there were 1.7 million term pairs that only changed one letter. In the experimental result 3, the features and learning data used in the experimental result 1 do not completely match, but there is a duplication of features and learning data that is sufficient to show the effectiveness of the proposed method. Then, using the technique of the different notation acquisition device 1, 70,000 pairs of different notations can be extracted therefrom. An example of different notation that can be constructed is shown below.
Business Week Business Week
JAVA Script JAVA Script
When I was writing When I was writing
? Irish Trad Irish Trad
Own server Own server
If not? If not?
? On WWW server On WWW server
Nikko carving Nikko carving
Next to each other next to each other

It should be noted that the difference notation of one character in the difference included in an EDR (Electronic Dictionary Research) electronic dictionary is 24,185 words. In addition, there are 82,270 words in which the difference included in the Japanese WordNet is a one-character different notation. However, Japanese WordNet contains many things that are not different notations (synonyms), and it is difficult to appropriately extract different notations. Further, the difference notation of one character included in the JUMAN dictionary is 23,348 words. Compared with these, the effectiveness of the proposed technique can be seen. “JAVA” is a registered trademark.
(Experimental result 4)

  In the fourth experiment, the accuracy is calculated by the above-described baseline method (method in which rules 1 to 4 are applied). In the baseline method, an F value is obtained by setting a positive feature (S55, S58, S67) having a high significant difference as a positive example and a negative example as a negative example.

  FIG. 16 shows the results when an experiment by 10-division cross-validation is performed using all of open data and closed data by the baseline method. In FIG. 16, “0” is a negative example, and “1” is a positive example. In FIG. 16, “0” and “1” in the leftmost column indicate correct classification. “0” and “1” in the uppermost first row indicate the output results of the experiment target technique (baseline technique in FIG. 16). That is, the number of data whose correct classification is “0” and the experimental result is “0” is 26892, the number of data whose correct classification is “0” and the experimental result is “1” is 1018, and the correct classification Is “1” and the experimental result is “0”, the number of data is 8, and the correct classification is “1” and the experimental result is “1”, and the number of data is 452. The reproducibility of the negative example (“0”) was 99.97%, and the precision was 96.35%. In addition, the recall rate of the positive example (“1”) was 30.75%, and the precision rate was 98.26%. The reproducibility of all data was 96.38%, and the precision was 96.38%. Furthermore, the “total number” is the number of experimental data. By substituting the above recall rate and precision rate into Equation 10, the F value of the negative example calculated was 0.9813, and the F value of the positive example was 0.4684. The meaning of each data in FIGS. 17 to 21 is the same as that in FIG.

  FIG. 17 shows the results when an experiment based on 10-division cross-validation is performed using only closed data by the baseline method. In FIG. 17, the F value of the negative example was 0.9814, and the F value of the positive example was 0.4833.

  FIG. 18 shows the results when the open-close method is used in the baseline method. In FIG. 18, the F value of the negative example was 0.9812, and the F value of the positive example was 0.4529.

In the experimental result 4, the baseline method shows that the F value of the positive example is extremely small compared to the F value in the proposed method (for example, 0.912 of the experimental result 1), and the effectiveness of the proposed method is extremely high. I can say that.
(Experimental result 5)

In the fifth experiment, the accuracy in the case of the baseline method in which all were positive examples was calculated. In the case of all baseline examples, the recall was “100%” and the precision was “0.0525%”. Then, the F value of the positive example (“1”) calculated by substituting the recall rate and the matching rate into Equation 10 was “0.0998”. It can be said that the F value of the positive example of the baseline method in which all are positive examples is extremely small compared with the F value in the proposed method, and the effectiveness of the proposed method is extremely high. In the proposed method, the correct answer rate “99.12%”, the recall rate “99.07%”, the precision rate “92.29%”, and the F value “0.912” are obtained. Note that the features and learning data used in this experiment do not completely match the features and learning data used in the evaluation of the proposed method, but the features and learning data are sufficient to show the effectiveness of the proposed method. There are duplicates.
(Application examples)

  Hereinafter, application examples of the different notation acquisition apparatus 1 will be described. An application example is an information retrieval device incorporating the different notation acquisition device 1. The information search device includes an alternate notation acquisition device 1 and a search unit. That is, the reception unit 14 receives a keyword (KW1). And the different notation acquisition apparatus 1 acquires the different notation term (KW2) of KW1 received. Then, the search unit performs information search using a search formula of KW1 + KW2 (+ is OR). Needless to say, the search target of the information search is not limited. Further, the search unit may be a process that only activates a so-called Web search engine.

  When a user inputs “spaghetti” when performing a keyword search for information using the information search device, the information search device acquires “spaghetti”, which is a different notation of “spaghetti”. Then, the information search device searches for information using both of these “spaghetti” and “spaghetti” as keywords. As a result, information that represents either “spaghetti” or “spaghetti” is hit, so that information retrieval with few omissions can be realized.

  In particular, the information retrieval apparatus has a great effect on retrieval of patent information that cannot be overlooked. For example, consider that an information search apparatus is used in a patent search system. For example, “computer” and “computer” exist in patent documents such as patent specifications, claims, abstracts, etc., so if you do not enter “computer + computer” as a keyword, the search will be omitted. Arise. Therefore, the searcher needs to consider the different notation of the keyword to be searched with great care when searching. Despite being synonymous, such as “digital” and “digital”, there are many different expressions, especially in patent gazettes. However, by adopting this information retrieval apparatus, it is possible to retrieve patent information without omissions without taking such consideration into consideration.

  As described above, according to the present embodiment, it is possible to extract differently-notated term pairs with high accuracy regardless of the field of term pairs.

  Note that, according to the present embodiment, the method for determining whether or not a term pair having an editing distance of 1 is a term pair having different notation has been described. However, as described above, the different notation acquisition device 1 can determine whether or not a term pair with an edit distance of 2 is also a different notation term pair.

  That is, the difference character acquisition unit 151 of the feature acquisition unit 15 acquires a pair of two difference characters for each term pair having two edit distances. For example, consider the following three specific term pairs. (1) “Can” “Can” (2) “Can understand” “Can” (3) “IX (Roman numeral 9)” “9” In such a case, the difference character acquisition unit 151 acquires two sets of difference characters, “de”, “out”, “ki”, and “coming”, for the term pair (1). Further, the difference character acquisition unit 151 acquires two sets of difference characters of “reason” “” and “solution” “” (“” is NULL) for the term pair in (2). Also, the difference character acquisition unit 151 sets two pairs of difference characters “I” “9” and “X” “” or “I” “” and “X” “9” for the term pair in (3). 2 sets of difference characters are acquired.

  Then, the feature acquisition unit 152 sets a plurality of features including one or more of a character type related feature, a dictionary related feature, and a similarity feature for two difference characters acquired by the difference character acquiring unit 151 independently. Get two sets. That is, for the term pair (1), the feature acquisition unit 152 extracts features for each of the two sets of difference characters “de”, “out”, “ki”, and “to”, The features extracted from the characters are considered different, and two types of test data are created. The feature acquisition unit 152, for example, regarding character type-related features indicating whether or not the character types of the editing positions of the two terms of the term pair are different and whether the editing positions of the two terms are numbers having the same value, It is determined that the edited portions of “de” and “out” are not numbers having the same value, and the character type related feature “0” is acquired. Also, the feature acquisition unit 152 acquires, for example, a dictionary-related feature “1” indicating whether the readings of the two terms match for “de” and “out”. The feature acquisition unit 152 acquires “de” reading “de” from the term dictionary 13 and determines that “de” and “depart” read match. Also, the feature acquisition unit 152, for example, with respect to the difference character “de” “out”, regarding the feature of the character before and after the difference character (edited part), the feature “” (none) of the previous character, the subsequent character Acquire the features “Ki” and “Ki”. Also, the feature acquisition unit 152, for example, with respect to the difference characters “ki” and “coming”, the features “out” and “de” of the previous character and the feature “ Get ". With this process, another difference character is included in the feature.

  In addition, for the term pair (2), the feature acquisition unit 152 extracts features for each of the two sets of difference characters of “reason” “” and “solution” “” as in (1). The features extracted from the difference characters are considered to be different, and two types of test data are created. Further, for the term pair (3), the feature acquisition unit 152, for example, each of two sets of difference characters “I”, “9”, “X”, and “”, as in (1) and (2). The features are extracted from the target, and the features extracted from the difference characters are considered to be different, and two types of test data are created.

  Next, the machine learning unit 16 determines whether or not the two types of test data for (1), (2), and (3) are term pairs having different notations. As a result of the determination, for example, when it is determined that the two types of test data are term pairs having different notations, the original term pairs (for example, “can” and “can”) are term pairs having different notations. As a result, the output unit 17 outputs the determination result. Note that, as described above, the output unit 17 adopts the score of the two scores for the two types of test data that has a score closer to 0, and if the adopted score is positive, a positive example (another notation) Term), if negative, it may be judged as a negative example (not a term in different notation), or the score with the larger absolute value of the score is adopted, and if the adopted score is positive, the positive example ( If it is negative, it may be judged as a negative example (not a term of different notation), and the smaller score of the two scores is acquired, and the smaller score is positive May be determined as a positive example (a term in different notation), and in a negative case as a negative example (not a term in different notation).

  In addition, two sets of difference characters (3) (for example, “I” “9” and “X” “” or “I” “” and “X” “9”), that is, two problems (problems) 1. If problem 2) is possible, the features are extracted for each difference character, and the features extracted from the difference characters are considered to be different, and four types of test data are created. Then, for each of the two questions, the one with the calculated score closer to 0 is acquired, and the score with the highest absolute value of the scores for each question is set as the score of the question. In this case, it may be determined as a negative example. Further, for example, when the term pair is (3) “IX” “9”, problems “I” “9” and “X” “” and “I” “” and “X” can be generated. Then, the machine learning unit 16 acquires the smaller score of “I” “9” and “X” “”, and acquires the smaller score of “I” “” and “X” “9”. Of the two acquired scores, the one with the larger value is set as the score in the term pair “IX” “9”. The machine learning unit 16 may determine that the score is a positive example when the score is positive and a negative example when the score is negative. For example, the machine learning unit 16 obtains the score of “I” “9” and “X” “” that is close to 0, and the scores of “I” “” and “X” “9” The one closer to 0 may be acquired, and the larger of the two acquired scores may be used as the score for the term pair “IX” “9”. The machine learning unit 16 may determine that the score is a positive example when the score is positive and a negative example when the score is negative.

  Further, in the present embodiment, for a term pair with an edit distance of 3 or more, similar to a term pair with an edit distance of 2, the test data of 3 or more is created and the determination result of the test data of 3 or more is used. It may be determined whether or not the original term pair is a different term term pair. In such a case, for example, a new feature may be introduced into the machine learning method, such as using one or two of three or more difference characters as a feature.

  Further, in the present embodiment, the different notation acquisition apparatus 1 can determine that the Japanese term pairs “you” and “anta” are different term term pairs, for example. , English) term pairs (for example, “colour” and “color”) can also be determined as terminology pairs.

  Further, in the present embodiment, the different notation acquisition apparatus 1 compiles edited portions without corresponding each character when the characters of the edited portions of two terms constituting a term pair are two or more characters. The collected character strings may be machine-learned as they are. That is, for the term pairs “123 sets” and “1233 sets”, the edited portions may be combined and processed as “123” and “1233”. For the term pairs “123 set” and “1233 set”, for example, S1 “Edited part of the first notation” “123”, S2 “Edited part of the second notation” “1233” ", S3" one character in front of the editing part "" "(none), S4" one character after the editing part "" group ", S55" if the editing part is both numbers, the same or different values "1" (same value), S56 "when the edited part of the Japanese term pair is both hiragana, whether the same voice or different voice" "0", etc. are obtained. Then, learning data for the term pairs “123 sets” and “1233 sets” may be configured, accumulated in the learning data storage unit 12, and used. Further, the feature acquisition unit 15 of the different notation acquisition device 1 processes the edited portions collectively for the test data of the term pairs having two or more characters in the edited portions, and acquires, for example, the above-described 68 features. The machine learning unit 16 may determine whether or not the test data is a different pair.

Furthermore, the processing in the present embodiment may be realized by software. Then, this software may be distributed by software download or the like. Further, this software may be recorded and distributed on a recording medium such as a CD-ROM. This also applies to other embodiments in this specification. Note that the software that implements the information processing apparatus according to the present embodiment is the following program. In other words, this program was obtained using a term pair having a feature distance related to a character pair whose edit distance is one or more and a character type of an edited portion which is a different character of the term pair, and a term dictionary on the storage medium. A dictionary-related feature that is a feature, a plurality of features that include one or more features of similarity features that are features that indicate the similarity of two terms that constitute the term pair, and a term pair in which the term pair is notated 2 or more learning data in association with positive / negative information, which is information indicating whether or not, and for each term pair of the storage medium, the computer stores character type related features, dictionary related features, similarity similarity features Using a feature acquisition unit that acquires a plurality of features including one or more of them, two or more learning data of the storage medium for each term pair, and a plurality of features acquired by the feature acquisition unit , With supervised machine learning, And machine learning unit where each term pair of the serial storage medium to determine whether a term pair different notation, a program, to function as an output unit for outputting the decision result in the machine learning unit.
(Embodiment 2)

  In the present embodiment, a different character acquisition device 2 that holds replacement character pairs, generates a term pair from a term using the replacement character pair, and generates a different notation term pair by machine learning for the term pair. Will be described. In addition to the function of the different notation acquisition device 1, the different notation acquisition device 2 has a function of having a function of generating different notation term pairs using patterns.

FIG. 19 is a block diagram of the different notation acquisition device 2 according to the present embodiment.
The different notation acquisition device 2 includes a term pair storage unit 11, a different notation term pair storage unit 21, a learning data storage unit 12, a term dictionary 13, an different notation pattern storage unit 22, a reception unit 23, an edited part acquisition unit 24, an different notation. The pattern acquisition unit 25, the different notation pattern storage unit 26, the term pair generation unit 27, the feature acquisition unit 15, the machine learning unit 16, and the output unit 17 are provided.

  The different notation term pair storage unit 21 can store one or more different notation term pairs having an edit distance of one.

  The different notation pattern storage unit 22 can store one or more different notation patterns having a pair of a first character string and a second character string indicating different notation patterns.

  The receiving unit 23 receives input from the user. The receiving unit 23 receives one or more terms. This term is a term from which a term pair is generated. The accepting unit 23 accepts the term from an input from a user, reading from a storage medium, or receiving using a communication means.

  The edit location acquisition unit 24 acquires the edit location of one or more different notation term pairs stored in the different notation term pair storage unit 21.

  The different notation pattern acquisition unit 25 acquires the different notation pattern having the first character string and the second character string as a pair from the edit location acquired by the edit location acquisition unit 24. The different notation pattern acquisition unit 25, for example, displays different notation patterns “2” and “two” having a pair of the first character string “2” and the second character string “2” from the term pairs “second” and “second”. get. The different notation pattern acquisition unit 25 acquires the different notation patterns “del” and “−” from the term pairs “own server” and “own server”, for example.

  The different notation pattern storage unit 26 stores the different notation pattern acquired by the different notation pattern acquisition unit 25 in the different notation pattern storage unit 22.

  The term pair generation unit 27 applies one or more different notation patterns of the different notation pattern storage unit 22 to one or more terms accepted by the accepting unit 23, generates one or more terms, and generates one or more terms. An alternate notation candidate term pair that is a term pair of one or more different notation candidates having each term and the generated term is generated.

  The different notation term pair storage unit 21 and the different notation pattern storage unit 22 are preferably non-volatile recording media, but can also be realized by volatile recording media.

  The process in which the different notation term pair is stored in the different notation term pair storage unit 21 does not matter.

  The edit location acquisition unit 24, the different notation pattern acquisition unit 25, the different notation pattern storage unit 26, and the term pair generation unit 27 can be usually realized by an MPU, a memory, or the like. The processing procedure of the editing location acquisition unit 24 and the like is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  Next, operation | movement of the different notation acquisition apparatus 2 is demonstrated using the flowchart of FIG. In the flowchart of FIG. 20, a process of accumulating different notation patterns and a process of generating term pairs will be described. Regarding the operation of the different notation acquisition device 2, the different notation term pair determination processing and the determination result output processing that are the same as the operation of the different notation acquisition device 1 will not be described.

  (Step S2001) The accepting unit 23 determines whether or not an instruction for generating a different notation pattern has been accepted. If an instruction for generating a different notation pattern is accepted, the process goes to step S2002, and if not, the process goes to step S2009.

  (Step S2002) The edit location acquisition unit 24 substitutes 1 for a counter i.

  (Step S2003) The edit location acquisition unit 24 determines whether or not the i-th term pair exists in the different notation pattern storage unit 22. If the i-th term pair exists, the process goes to step S2004. If the i-th term pair does not exist, the process returns to step S2001.

  (Step S2004) The edit location acquisition part 24 acquires the difference character (edit location) of the i-th term pair.

  (Step S2005) The different notation pattern acquisition unit 25 configures the different notation pattern from the difference character (edited portion) acquired in step S2004.

  (Step S2006) The different notation pattern storage unit 26 determines whether or not the different notation pattern acquired in step S2005 exists in the different notation pattern storage unit 22. If it exists, go to step S2007, otherwise go to step S2008.

  (Step S2007) The different notation pattern accumulating unit 26 accumulates the different notation pattern acquired in step S2005 in the different notation pattern storage unit 22.

  (Step S2008) The edit location acquisition unit 24 increments the counter i by 1. The process returns to step S2003.

  (Step S2009) The reception unit 23 determines whether a term has been received. If the term is accepted, the procedure goes to step S2010, and if not, the procedure returns to step S2001.

  (Step S2010) The term pair generation unit 27 substitutes 1 for a counter i.

  (Step S2011) The term pair generation unit 27 determines whether or not the i-th different notation pattern exists in the different notation pattern storage unit 22. If it exists, the process goes to step S2012, and if it does not exist, the process ends.

  (Step S2012) The term pair generation unit 27 determines whether or not the term accepted in step S2009 matches the i-th notation pattern. If they match, go to step S2013, otherwise go to step S2016. Note that the different notation patterns “2” and “2” do not match the term “WWW server”. This is because both of the different notation patterns are characters, and neither of the characters includes the term “WWW server”. Also, the different notation patterns “del” and “−” match the term “WWW server”. When “del” is included in the different notation pattern, all terms match the different notation pattern.

  (Step S2013) The term pair generation unit 27 applies the i-th different notation pattern to the terms accepted in step S2009, and acquires one or more different notation terms. When the term is “eye tracking” and the i-th different notation pattern is “del” “•”, the term pair generation unit 27 applies the different notation pattern “del” “•” to the term “eye tracking”. Insert "・" between each character, and use the seven different terms "a tracking", "eye tracking", "eye racking", "eye trucking", "eye tracking king", "eye tracking king" Creates “Itracking Gu”. When the term is “first” and the i-th different notation pattern is “one” “1”, the term pair generation unit 27 adds the different notation patterns “one” “1” to the term “first”. Is applied to generate “first”.

  (Step S2014) The term pair generation unit 27 generates one or more term pairs using the terms accepted in Step S2009 and one or more different notation terms generated in Step S2013. For example, when the term is “eye tracking” and the i-th notation pattern is “del” “•”, the term pair generation unit 27 sets the term pairs “eye tracking”, “a-itracking”, “eye”. Tracking, Eye Tracking, Eye Tracking, Eye Tracking, Eye Racking, Eye Tracking, Eye Traking, Eye Tracking, Eye Tracking King, Eye Tracking, Eye Tracking 7 "," eye tracking "and" eye tracking "are generated. When the term is “first” and the i-th different notation pattern is “one” “1”, the term pair generation unit 27 generates the term pair “first” “first”.

  (Step S2015) The term pair generation unit 27 accumulates one or more term pairs generated in step S2013 in the term pair storage unit 11.

  (Step S2016) The term pair generation unit 27 increments the counter i by 1. The process returns to step S2011.

  Hereinafter, a specific operation of the different notation acquisition device 2 in the present embodiment will be described.

  An example of the different notation pattern acquired by the different notation pattern acquisition unit 25 and stored in the different notation pattern storage unit 22 by the different notation pattern storage unit 26 is shown in FIG. In FIG. 21, “del” indicates that the other pattern character is to be deleted. That is, the different notation pattern acquisition unit 25 generates a term pair for the other pattern characters of del with respect to the terms used in all large-scale similar word lists.

  In this situation, as described above, the term pair generation unit 27 determines that the accepted term “eye tracking” matches the first different notation pattern “del” “•”. When the term “eye tracking” is input, the first different notation pattern “del” “•” is applied, and the term pair generation unit 27 sets the term “eye tracking” to “•” for each character. Inserted in between to generate seven different terms "a tracking", "eye tracking", "eye racking", "eye tracking", "eye tracking king", "eye tracking king", "eye tracking" To do. Next, the term pair generation unit 27 includes the term pairs “eye tracking”, “a itracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye racking”, “eye tracking”. Seven term pairs of “tracking”, “eye tracking”, “eye tracking” “eye tracking king”, “eye tracking” “eye tracking”, “eye tracking” “eye tracking” are generated. Then, the different notation pattern storage unit 26 stores seven term pairs in the different notation pattern storage unit 22.

  Next, the term pair generation unit 27 determines that the accepted term “eye tracking” matches the second different notation pattern “del” and “−”. Next, the second different notation pattern “del” “−” is applied to the term “eye tracking”, and the term pair generation unit 27 performs “eye tracking” “eye tracking” “eye tracking”. "Eye tracking", "eye tracking", "eye tracking", and "eye tracking". Next, the term pair generation unit 27 includes the term pairs “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”. Seven term pairs of “king”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking” are generated. Then, the different notation pattern storage unit 26 stores seven term pairs in the different notation pattern storage unit 22.

  Next, the term pair generation unit 27 determines that the accepted term “eye tracking” matches the third different notation pattern “del” “yes”. Then, the third different notation pattern “del” “I” is applied to the term “eye tracking”, and the term pair generation unit 27 performs “ai i tracking” “eye tracking” “eye tracking”. "I racking", "eye tracking", "eye tracking king", "eye tracking", and "eye tracking" are generated. Next, the term pair generation unit 27 includes the term pairs “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”, “eye tracking”. Seven term pairs of “King”, “Eye tracking”, “Eye tracking”, “Eye tracking”, “Eye tracking”, “Eye tracking”, “Eye tracking” are generated. Then, the different notation pattern storage unit 26 stores seven term pairs in the different notation pattern storage unit 22.

  Next, since the accepted term “eye tracking” does not include the characters constituting the fourth different notation pattern “−” “1”, the term pair generation unit 27 does not match this different notation pattern. to decide.

  Next, since the received term “eye tracking” includes the character “I” that constitutes the fifth different notation pattern “I” and “I”, the term pair generation unit 27 matches this different notation pattern. Judge that. Then, the fifth different notation pattern “A” and “I” are applied to the term “eye tracking”, and the term pair generation unit 27 generates “A tracking”. Next, the term pair generation unit 27 generates one term pair of the term pair “eye tracking” and “a tracking”. Then, the different notation pattern storage unit 26 stores one term pair in the different notation pattern storage unit 22.

  Next, similarly, the term pair generation unit 27 applies and processes the sixth and subsequent different notation patterns.

  Then, the term pair generation unit 27 accumulates new term pairs in the term pair storage unit 11.

  As described above, according to the present embodiment, it is possible to automatically generate a candidate for a different pair of terminology. Further, according to the present embodiment, it is possible to automatically acquire a different notation pattern for automatically generating a candidate for a different pair of terminology.

Note that the software that implements the information processing apparatus according to the present embodiment is the following program. That is, this program stores, on the storage medium, a character type-related feature that is a feature related to a character type of an edited portion that is a character having a different term pair, a dictionary-related feature that is a feature acquired using a term dictionary, and the term pair. A plurality of features including one or more features of similarity features that are the features indicating the similarity of two terms constituting; and positive / negative information that is information indicating whether the term pair is a different pair of terms 2 or more learning data in association with each other is stored, and one or more different notation patterns having a pair of a first character string and a second character string indicating a different notation pattern are stored in a storage medium. An accepting unit that accepts one or more terms, and applies one or more different notation patterns of the storage medium to one or more terms accepted by the accepting unit to generate one or more terms, One or more terms and the generation For each one or more different notation candidate term pairs generated by the term pair generation unit, the term pair generation unit generates a different notation candidate term pair that is a term pair of one or more different notation candidates having A feature acquisition unit that acquires a plurality of features including at least one of a character type-related feature, a dictionary-related feature, and a similarity feature, and each non-notation candidate term pair generated by the term pair generation unit, A term in which each different notation candidate term pair in the term pair storage unit is differently expressed by supervised machine learning using two or more learning data of the storage medium and a plurality of features acquired by the feature acquisition unit. It is a program for functioning as a machine learning part which judges whether it is a pair, and an output part which outputs the judgment result in the machine learning part.
FIG. 22 shows the external appearance of a computer that executes the program described in this specification and realizes the different notation acquisition device of the above-described embodiment. The above-described embodiments can be realized by computer hardware and a computer program executed thereon. FIG. 22 is a schematic view of the computer system 340, and FIG. 23 is a block diagram of the computer system 340.

  In FIG. 22, a computer system 340 includes a computer 341 including an FD drive and a CD-ROM drive, a keyboard 342, a mouse 343, and a monitor 344.

  23, in addition to the FD drive 3411 and the CD-ROM drive 3412, the computer 341 stores an MPU 3413, a bus 3414 connected to the CD-ROM drive 3412 and the FD drive 3411, and a program such as a bootup program. A RAM 3416 for temporarily storing application program instructions and providing a temporary storage space; and a hard disk 3417 for storing application programs, system programs, and data. Although not shown here, the computer 341 may further include a network card that provides connection to the LAN.

  A program that causes the computer system 340 to execute the functions of the different notation acquisition apparatus and the like of the above-described embodiment is stored in the CD-ROM 3501 or FD 3502, inserted into the CD-ROM drive 3412 or FD drive 3411, and further a hard disk 3417 may be transferred. Alternatively, the program may be transmitted to the computer 341 via a network (not shown) and stored in the hard disk 3417. The program is loaded into the RAM 3416 at the time of execution. The program may be loaded directly from the CD-ROM 3501, the FD 3502, or the network.

  The program does not necessarily include an operating system (OS), a third party program, or the like that causes the computer 341 to execute the functions of the different notation acquisition device of the above-described embodiment. The program only needs to include an instruction portion that calls an appropriate function (module) in a controlled manner and obtains a desired result. How the computer system 340 operates is well known and will not be described in detail.

  Further, the computer that executes the program may be singular or plural. That is, centralized processing may be performed, or distributed processing may be performed.

  Further, in each of the above embodiments, it goes without saying that two or more communication units existing in one apparatus may be physically realized by one medium.

  In each of the above embodiments, each process (each function) may be realized by centralized processing by a single device (system), or by distributed processing by a plurality of devices. May be.

  The present invention is not limited to the above-described embodiments, and various modifications are possible, and it goes without saying that these are also included in the scope of the present invention.

  As described above, the different notation acquisition apparatus according to the present invention has the effect that it is possible to extract a highly accurate different notation term pair regardless of the field of term pairs, Useful.

DESCRIPTION OF SYMBOLS 1, 2 Different notation acquisition apparatus 11 Term pair storage part 12 Learning data storage part 13 Term dictionary 14, 23 Reception part 15 Feature acquisition part 16 Machine learning part 17 Output part 21 Different notation term pair storage part 22 Different notation pattern storage part 24 Edit location acquisition unit 25 Different notation pattern acquisition unit 26 Different notation pattern storage unit 27 Term pair generation unit 151 Difference character acquisition unit 152 Feature acquisition unit 171 Threshold storage unit 172 Threshold determination unit 173 Output unit

Claims (10)

A term pair storage unit capable of storing one or more term pairs having an edit distance of one or more;
Indicates the character type-related feature that is a feature related to the character type of the edited part that is a different character of the term pair, the dictionary-related feature that is a feature acquired by using the term dictionary, and the similarity of two terms that constitute the term pair Stores two or more learning data in which a plurality of features including one or more features of similarity features that are features and positive / negative information that is information indicating whether the term pair is a different pair of terms Possible learning data storage,
A feature acquisition unit that acquires a plurality of features including one or more of a character type related feature, a dictionary related feature, and a similarity feature for each term pair in the term pair storage unit;
For each term pair, each of the term pair storage units is obtained by supervised machine learning using two or more learning data in the learning data storage unit and a plurality of features acquired by the feature acquisition unit. A machine learning unit that determines whether or not the term pair is a different term term pair;
A different notation acquisition device comprising: an output unit that outputs a determination result in the machine learning unit ;
A term dictionary capable of storing one or more term information having terms and representative representations of the terms;
The dictionary-related features are:
Information indicating whether or not the representative notation of two terms in a term pair is the same,
The feature acquisition unit
For each term pair in the term pair storage unit, representative representations of two terms included in the term pair are acquired from the term dictionary, and it is determined whether or not the two representative representations acquired are the same. A different notation acquisition device that acquires a result as a dictionary-related feature . The character type-related features are:
Information indicating whether or not the character types of the edited portions of the two terms in the term pair are different and whether the edited portions of the two terms are numbers having the same value ,
The feature acquisition unit
Whether the term pairs in the term pair storage unit have different character types in the editing positions of the two terms in the term pair and whether the editing positions of the two terms have the same value or not The different notation acquisition apparatus according to claim 1, wherein the determination result is acquired as a character type-related feature. The character type-related features are:
Information indicating whether or not the character type of the edited part of the two terms in the term pair is Roman, and whether the edited part of the two terms is a difference between uppercase and lowercase letters,
The feature acquisition unit
For each term pair in the term pair storage section, it matches the condition that the character type of the edited part of the two terms in the term pair is Roman, and the edited part of the two terms is different between uppercase and lowercase letters The different notation acquisition apparatus according to claim 1, wherein it is determined whether or not to perform determination, and the determination result is acquired as a character type related feature. The term information is
Corresponding to the term, also has a reading of the term,
The dictionary-related features are:
Information indicating whether or not the representative notation of two terms in a term pair is the same,
The feature acquisition unit
For each term pair in the term pair storage unit, representative representations of two terms included in the term pair are acquired from the term dictionary, and it is determined whether or not the two representative representations acquired are the same. The different notation acquisition apparatus according to claim 1, wherein the result is acquired as a dictionary-related feature. The machine learning unit
It is determined whether each term pair in the term pair storage unit is a term pair with different notation, and also obtains a score indicating the accuracy of being a term pair with different notation,
The output unit is
The different notation acquisition apparatus in any one of Claims 1-4 which outputs the score which the said machine learning part acquired. The output unit is
Threshold storage means for storing score thresholds;
Threshold determination means for determining whether the score acquired by the machine learning unit is equal to or greater than the threshold or greater than the threshold;
The term pair corresponding to the score determined by the threshold judging means to be greater than or equal to the threshold or greater than the threshold is taken as a judgment result that is a different term term pair, and is not the judgment result or the different term term pair or different notation. The different notation acquisition apparatus according to claim 5 , further comprising an output unit that outputs at least one of the term pairs. Character type-related features that are features related to the character type of the edited part that are different characters of the term pairs, dictionary-related features that are information indicating whether or not the representative notation of two terms of the term pairs are the same, the term pairs A plurality of features including one or more features of similarity features that are the features indicating the similarity of two terms that constitute, and positive / negative information that is information indicating whether the term pair is a term pair of different notations A learning data storage unit capable of storing two or more learning data associated with
A different notation pattern storage unit capable of storing one or more different notation patterns having a pair of a first character string and a second character string indicating different notation patterns;
A reception unit that accepts one or more terms;
When it is determined whether or not any one or more of the different notation patterns of the different notation pattern storage unit includes any one of the character strings received by the accepting unit In addition, the portion of the character string included in the term is a character string of the different notation pattern, and the one or more terms are generated by replacing the other character string of the included character string, A term pair generation unit that generates a different notation candidate term pair that is a term pair of one or more different notation candidates having each term and the generated term;
A feature acquisition unit that acquires a plurality of features including at least one of a character type-related feature, a dictionary-related feature, and a similarity feature for each one or more different notation candidate term pairs generated by the term pair generation unit; ,
For each different notation candidate term pair generated by the term pair generation unit, supervised machine learning using two or more learning data in the learning data storage unit and a plurality of features acquired by the feature acquisition unit A machine learning unit that determines whether or not each of the different notation candidate term pairs in the term pair storage unit is a differently expressed term pair by law,
A different notation acquisition device comprising: an output unit that outputs a determination result in the machine learning unit;
A different notation term pair storage unit capable of storing one or more different notation term pairs having an edit distance of 1,
An edit location acquisition unit for acquiring an edit location of one or more different notation term pairs stored in the different notation term pair storage unit;
From the edited part acquired by the edited part acquisition unit, for each term pair having different notation, the edited part included in the character string of one term constituting the term pair is set as the first character string, and the term pair is configured. get the edits contained in the string of terms as a second string, and different notation pattern acquisition unit that acquires different notation pattern having a said first string and the second string in pairs,
The different notation acquisition apparatus which further comprises the different notation pattern storage part which accumulate | stores the different notation pattern which the said different notation pattern acquisition part acquired in the said different notation pattern storage part. The edit distance of the term pair is 2,
The feature acquisition unit
Difference character acquisition means for acquiring a pair of two difference characters of the term pair, respectively;
Feature acquisition means for acquiring two sets of plural features including one or more of character type-related features, dictionary-related features, and similarity features, targeting two difference characters acquired by the difference character acquisition means independently And
The machine learning unit
A supervised machine learning method using a plurality of features of each set and two or more learning data of the learning data storage unit for each set of the two sets of features acquired by the feature acquisition means Thus, it is determined whether or not a plurality of features of each set of the term pair storage unit is a set of features corresponding to a term pair of different notation, and the edit distance is 2 using the two determination results the term pairs different notation acquisition device according to any one of claims 1 to 7 for determining whether a term pair different notation. On the storage medium,
A term pair whose edit distance is 1 or more,
A term dictionary capable of storing one or more term information having a term and a representative representation of the term; and
Indicates the character type-related feature that is a feature related to the character type of the edited part that is a different character of the term pair, the dictionary-related feature that is a feature acquired by using the term dictionary, and the similarity of two terms that constitute the term pair Stores two or more learning data in which a plurality of features including one or more features of similarity features that are features and positive / negative information that is information indicating whether the term pair is a different pair of terms And
A different notation acquisition method realized by a feature acquisition unit, a machine learning unit, and an output unit,
A feature acquisition step of acquiring a plurality of features including at least one of a character type related feature, a dictionary related feature, and a similarity feature for each term pair of the storage medium by the feature acquisition unit;
For each term pair, the machine learning unit uses the two or more learning data of the storage medium and the plurality of features acquired in the feature acquisition step to perform the storage by the supervised machine learning method. A machine learning step for determining whether each term pair in the medium is a term pair with a different notation;
An output step of outputting a determination result in the machine learning step by the output unit ;
The dictionary-related features are:
Information indicating whether or not the representative notation of two terms in a term pair is the same,
In the feature acquisition step,
For each term pair of the storage medium, the representative notation of two terms included in the term pair is acquired from the term dictionary, it is determined whether the two representative notations acquired are the same, and the determination result is Different notation acquisition method acquired as dictionary-related features . On the storage medium,
A term pair whose edit distance is 1 or more,
A term dictionary capable of storing one or more term information having a term and a representative representation of the term; and
Indicates the character type-related feature that is a feature related to the character type of the edited part that is a different character of the term pair, the dictionary-related feature that is a feature acquired by using the term dictionary, and the similarity of two terms that constitute the term pair Stores two or more learning data in which a plurality of features including one or more features of similarity features that are features and positive / negative information that is information indicating whether the term pair is a different pair of terms And
Computer
A feature acquisition unit that acquires a plurality of features including at least one of a character type related feature, a dictionary related feature, and a similarity feature for each term pair of the storage medium;
For each term pair, each term pair of the storage medium is different by supervised machine learning using two or more learning data of the storage medium and a plurality of features acquired by the feature acquisition unit. A machine learning unit that determines whether or not the term pair is a notation,
A program for causing the machine learning unit to function as an output unit that outputs a determination result ,
The dictionary-related features are:
Information indicating whether or not the representative notation of two terms in a term pair is the same,
The feature acquisition unit
For each term pair in the term pair storage unit, representative representations of two terms included in the term pair are acquired from the term dictionary, and it is determined whether or not the two representative representations acquired are the same. A program that causes a computer to function as a dictionary-related feature .