JP5412137B2 - Machine learning apparatus and method - Google Patents

Description

  The present invention relates to a machine learning apparatus and method. In particular, the present invention relates to a machine learning apparatus and method for adding machine learning teacher data.

  Conventionally, in machine learning in which a computer extracts useful rules based on sample data, a large amount of sample data is required to extract useful rules. To create learning data that is a large amount of sample data, Requires a lot of manpower. There is a bootstrap technique as a technique for eliminating this enormous manpower.

  The bootstrap method creates a large amount of learning data by extracting data that conforms to a rule based on a small amount of initial learning data and adding it to the learning data. In this bootstrap method, even if the data is inappropriate, it is added to the learning data if it conforms to the rules based on the learning data. Therefore, data that conforms to the rules based on the learning data including inappropriate data is extracted. As a result, the large amount of learning data created contained a large amount of inappropriate data. In such a bootstrap technique, the invention of Patent Document 1 is known in which inappropriate data is removed when learning data is added.

  The invention described in Patent Document 1 divides learning data into a plurality of sub-learning data in order to avoid incorporating erroneously predicted data into learning data in the process of growing learning data by the bootstrap method. And an opportunity to perform machine learning based on learning data that does not include inappropriate data. Then, comprehensive evaluation prediction is performed using evaluation prediction based on the obtained plurality of learning results, and the data is incorporated into the learning data.

JP 2005-92253 A

  However, as in Patent Document 1, even if the learning data is divided into a plurality of sub-learning data, there is no opportunity to perform machine learning based on learning data that does not include any inappropriate data. Depending on the correct evaluation prediction, erroneous data may be incorporated into the learning data. In this way, when machine learning is performed such that erroneously evaluated data is incorporated into learning data, there is an opportunity to perform machine learning based on learning data that does not include inappropriate data even if divided into sub-learning data Since it decreases, the reliability of the finally obtained learning data may be low.

  Therefore, a machine learning device that omits inappropriate terms from terms added to teacher data for machine learning is desired.

  The present invention provides a machine learning apparatus and method for improving the accuracy of teacher data by omitting inappropriate terms from terms added to teacher data for machine learning, adding terms that match teacher data, and improving the accuracy of teacher data. For the purpose.

  The present invention provides the following solutions.

  (1) A machine learning device in which a computer learns based on teacher data for each category given in advance, and the teacher data storage means for storing the teacher data matches the category of the teacher data stored in advance. Morpheme analysis means for morphological analysis of the text of the selected category, feature extraction means for extracting the features of the morpheme analyzed by the morpheme, and information on morphemes before and after the morpheme having a specific feature among the extracted features Pre- and post-information extraction means for extracting certain pre- and post-information, specific expression acquisition means for acquiring a specific expression based on the extracted pre- and post-information, and specific expression storage processing for storing the acquired unique expression in the specific expression storage means Means for acquiring teacher data from the teacher data storage means, and the acquired teacher data A combination acquisition unit that acquires a combination with terms appearing in the pattern, a pattern extraction unit that extracts an appearance part of the teacher data as an arbitrary replaceable term in the acquired combination, and extracted A pattern correspondence term acquisition means for acquiring a pattern correspondence term, which is a term included in the replaceable portion of the pattern, from a corpus, and the acquired pattern correspondence term is based on the specific expression storage means A pattern-corresponding term determining unit that determines whether or not the expression matches, and a pattern-corresponding term that matches based on the determination is acquired, and the acquired matching pattern-corresponding term is added to the teacher data storage unit as teacher data A machine learning device comprising: a teacher data adding means.

  According to the configuration of (1), the machine learning device according to the present invention is a machine learning device in which a computer learns based on teacher data for each category given in advance, and teacher data storage means for storing teacher data Morphological analysis is performed on the text in the category that matches the category of the pre-stored teacher data, the morpheme features extracted from the morpheme analysis, and the morpheme before and after the morpheme having a specific feature among the extracted features The before and after information that is the information is extracted, a unique expression is acquired based on the extracted before and after information, and the acquired unique expression is stored in the specific expression storage means. Then, teacher data (for example, XX) is acquired from the teacher data storage means, and combinations (for example, “XX account”, “XX branch”,...) With terms that appear together with the acquired teacher data are acquired. In the obtained combination, a pattern in which an arbitrary replaceable term is included in the appearance part of the teacher data (for example, “# account”, “# branch”, where # is an arbitrary replaceable term) A pattern-corresponding term is extracted from the corpus (for example, “XX account”, “Tokyo branch”). ) And the terms that fall into the replaceable part corresponding to the extracted pattern are XX and Tokyo). Then, it is determined whether or not the acquired pattern correspondence term (for example, xx, Tokyo) is suitable for a specific expression to be machine-learned based on the specific expression storage means, and the pattern correspondence term that is suitable based on the determination (For example, xx) is acquired, and the acquired matching pattern correspondence term (for example, xx) is added to the teacher data storage means as teacher data.

  That is, the machine learning device according to the present invention acquires a pattern-corresponding term of a pattern that appears together with teacher data from the corpus, and the acquired pattern-corresponding term is a unique expression that should be machine-learned based on the specific-expression storage means The pattern correspondence terms that match are added to the teacher data storage means, and the pattern correspondence terms that do not match are not added to the teacher data storage means. Therefore, the machine learning device according to the present invention eliminates inappropriate terms from terms to be added to teacher data for machine learning, adds terms that match teacher data, and improves the accuracy of teacher data. Can do.

  (2) The machine learning device according to (1), wherein the corpus is a search log storing a search query.

  According to the configuration of (2), the corpus acquired by the machine learning device according to (1) is a search log storing a search query. In general, for example, a search log of a Web search is a search site in which information that a general user wants to search is input to a search window. In order to narrow down search results, a plurality of search queries (search keywords) are separated by spaces. Often used to input. Therefore, the search log is appropriately segmented in units of words, and it is easy to extract words and handle them easily. From the search log storing such a search query, the machine learning device extracts a pattern that appears together with the teacher data, and acquires a pattern corresponding term of the extracted pattern. Therefore, the machine learning device according to the present invention acquires the term to be added to the teacher data for machine learning from an appropriate corpus, so the inappropriate term is omitted, the term that matches the teacher data is added, and the teacher data Accuracy can be improved.

  (3) A machine learning method in which a computer learns based on teacher data storage means for storing teacher data for each category given in advance, and the category text that matches the category of the teacher data stored in advance is stored. A step of performing morpheme analysis, a step of extracting features of the morpheme analyzed by the morpheme, a step of extracting front-and-rear information which is information of morphemes before and after the morpheme having a specific feature among the extracted features, A step of acquiring a unique expression based on the extracted before and after information, and a step of storing the acquired unique expression in a specific expression storage unit, and acquiring teacher data from the teacher data storage unit And a step of acquiring a combination of terms that appear together with the acquired teacher data, and the acquired combination Then, the step of extracting the appearance part of the teacher data as a pattern regarded as containing any replaceable term, and the pattern corresponding term that is the term that falls within the replaceable part of the extracted pattern, A step of acquiring from the inside, a step of determining whether the acquired pattern-corresponding term conforms to the specific expression based on the specific-expression storage means, and a pattern-corresponding term that conforms to the determination A machine learning method comprising: acquiring the acquired matching pattern correspondence term as teacher data in the teacher data storage means.

  According to the configuration of (3), the machine learning method according to the present invention is based on the teacher data storage means for storing the teacher data for each category given in advance. Morphological analysis is performed on the text of the matched category, the features of the morpheme analyzed are extracted, and the information before and after the morpheme before and after the morpheme having the specific feature is extracted and extracted. A unique expression is acquired based on the before and after information, and the acquired unique expression is stored in the specific expression storage means. Then, teacher data (for example, XX) is acquired from the teacher data storage means, and combinations (for example, “XX account”, “XX branch”,...) With terms that appear together with the acquired teacher data are acquired. In the obtained combination, a pattern in which an arbitrary replaceable term is included in the appearance part of the teacher data (for example, “# account”, “# branch”, where # is an arbitrary replaceable term) A pattern-corresponding term is extracted from the corpus (for example, “XX account”, “Tokyo branch”). ) And the terms that fall into the replaceable part corresponding to the extracted pattern are XX and Tokyo). Then, it is determined whether or not the acquired pattern correspondence term (for example, xx, Tokyo) is suitable for a specific expression to be machine-learned based on the specific expression storage means, and the pattern correspondence term that is suitable based on the determination (For example, xx) is acquired, and the acquired matching pattern correspondence term (for example, xx) is added to the teacher data storage means as teacher data. Therefore, the machine learning method according to the present invention eliminates inappropriate terms from terms added to teacher data for machine learning, adds terms that match teacher data, and improves the accuracy of teacher data. Can do.

  According to the present invention, it is possible to omit inappropriate terms from terms added to teacher data for machine learning, add terms that match teacher data, and improve the accuracy of teacher data.

It is a functional block diagram which shows the function of the machine learning apparatus 10 which concerns on one Embodiment of this invention. It is a figure which shows an example of the hardware constitutions of the machine learning apparatus 10 which concerns on one Embodiment of this invention. It is a figure which shows corpus DB32 which concerns on one Embodiment of this invention. It is a figure which shows pattern DB33 which concerns on one Embodiment of this invention. It is a figure which shows NE dictionary DB41 concerning one Embodiment of this invention. It is a flowchart which shows the processing content of the machine learning apparatus 10 which concerns on one Embodiment of this invention. It is a flowchart which shows the NE dictionary creation process of the machine learning apparatus 10 which concerns on one Embodiment of this invention. It is a figure which shows teacher data DB31 which concerns on one Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 is a functional block diagram illustrating functions of a machine learning device 10 according to an embodiment of the present invention. A machine learning device 10 according to an embodiment of the present invention includes a teacher data DB 31 as a teacher data storage unit, a NE (Named Entity) dictionary DB 41 as a specific expression storage unit, a teacher data acquisition unit 11 as a teacher data acquisition unit, A combination acquisition unit 12 as a combination acquisition unit, a pattern extraction unit 13 as a pattern extraction unit, a pattern corresponding term acquisition unit 14 as a pattern corresponding term acquisition unit, a pattern corresponding term determination unit 15 as a pattern corresponding term determination unit, and teacher data A teacher data adding unit 16 is provided as an adding means. Furthermore, a pattern DB 33 for storing patterns, a morpheme analysis unit 21 as a morpheme analysis unit, a feature extraction unit 22 as a feature extraction unit, and a front-rear information extraction unit as a feature extraction unit for storing expressions unique to the NE dictionary DB 41 An extraction unit 23, a specific expression acquisition unit 24 as a specific expression acquisition unit, and a specific expression storage processing unit 25 as a specific expression storage processing unit.

  The teacher data DB 31 stores teacher data for each category. The initial value of the teacher data is correct case data. For example, when the teacher data stores a specific expression of the financial category, the initial value is “XX” when the well-known bank name is “XX” as the correct example data of the specific expression of the financial category. It is remembered.

  The NE dictionary DB 41 performs morphological analysis on the text of the category that matches the category of the teacher data, extracts the features of the analyzed morpheme, and includes information on the extracted features and the morpheme having a specific feature among the extracted features. And the unique expression acquired based on the information of the later morpheme is stored. Information on the extracted features includes notation and part-of-speech tags given in morphological analysis, character types that are features (Hiragana, Katakana, Kanji, English, uppercase, lowercase, number of characters, etc.), first character of the notation, last character of the notation, default tag (Count in the correct answer corpus and tag the majority). Examples of morphemes having specific features include nouns and capital letters. The information of the morpheme before and after is a morpheme that co-occurs with a morpheme having a specific feature, a morpheme immediately before and after a morpheme having a specific feature, or the like. Specific expressions include, for example, proper noun expressions such as person names, place names, and organization names, time expressions such as dates and times, and numerical expressions such as amounts and ratios. The NE dictionary DB 41 stores a unique expression for each category (for example, XX, XX, etc., where XX is also a well-known bank name). In addition, information on the extracted features possessed by the unique expression in the category and information on the previous and subsequent morphemes are stored in association with the unique expression. And the type of the specific expression calculated | required based on the information regarding the extracted feature and the information on the previous and subsequent morphemes (for example, specific, ambiguous expression, frequent expression, etc.) is stored.

  The teacher data acquisition unit 11 acquires teacher data (for example, OO) from the teacher data DB 31.

  The combination acquisition unit 12 acquires a combination with a term that appears together with the acquired teacher data (for example, OO). For example, a combination of “XX account”, “XX branch”, etc. appearing together with “XX” is acquired.

  In the acquired combination, the pattern extraction unit 13 extracts the appearance portion of the teacher data as a pattern that is considered to contain any replaceable term. For example, in a combination of “XXX account”, “XX branch”, etc., a pattern (for example, “# account”, “# Branch” etc.). The pattern DB 33 stores the extracted pattern. When the corpus is a search log, the pattern extraction unit 13 extracts all character strings other than the character string of the teacher data as a pattern from the search log.

  The pattern-corresponding term acquisition unit 14 acquires a pattern-corresponding term that is a term included in a replaceable portion of the extracted pattern from the corpus. A pattern corresponding term (for example, xx, Tokyo) corresponding to # of the extracted pattern (for example, “# account”, “# branch”) is acquired from the corpus. The corpus can be acquired from the corpus DB 32 including, for example, Web data and a search log. In particular, in the search log, the user designates word division, and it is easy to extract words and handle them easily.

  The pattern corresponding term determination unit 15 determines whether or not the acquired pattern corresponding term (for example, xx, Tokyo) matches the unique expression based on the NE dictionary DB 41. The determination as to whether or not the acquired pattern correspondence term conforms to a specific expression is performed when the acquired pattern correspondence term is stored as a unique expression in the NE dictionary DB 41 and matches the category to be acquired. Is determined to be suitable. Further, even if the acquired pattern correspondence term is not stored as a unique expression in the NE dictionary DB 41, it is determined that the acquired pattern correspondence term and the frequency of occurrence of the pattern are equal to or higher than a predetermined threshold value. To do. Specifically, when the acquired pattern correspondence term is “Tokyo”, “Tokyo” is stored in the NE dictionary DB 41, but is stored as an expression that does not match the category to be acquired. Therefore, it is determined that “Tokyo” does not match the unique expression. On the other hand, if it is stored that the acquired pattern correspondence term “XX” matches the category to be acquired, it is determined that it matches the unique expression. Further, even if the acquired pattern correspondence term is a pattern correspondence term that should originally be obtained, if it is an unusual expression (in this case, for example, a bank name in another country), it is stored in the NE dictionary as a unique expression. However, even in this case, machine learning is repeated, and it is determined that it is suitable when the appearance frequency is equal to or higher than a predetermined threshold.

  The teacher data adding unit 16 acquires a matching pattern correspondence term (for example, xx) based on the determination, and adds the acquired matching pattern correspondence term (for example, xx) to the teacher data DB 31.

  The morpheme analysis unit 21 performs a morphological analysis on a category text (for example, a Web document) that matches the category of the teacher data, and the feature extraction unit 22 extracts the features of the morpheme analyzed by the morpheme analysis unit 21. The before-and-after information extracting unit 23 extracts the before-and-after information that is the information of the morpheme before and after the morpheme having the specific feature from the features extracted by the feature extracting unit 22, The unique expression is acquired based on the unique expression, and the unique expression acquired by the specific expression storage processing unit 25 is stored in the NE dictionary DB 41.

  Here, the morpheme is the smallest linguistic unit having a meaning, and is a linguistic unit used when a sentence written in a natural language is divided. For example, “Today is a good weather” is divided into morphemes of “Today / Has / Good / Weather / Is it”. Features that characterize morphemes include, for example, character types, characters, default tags, additional information, and fine classification information in addition to parts of speech and word forms.

  In the storage of the NE dictionary DB 41, based on a feature map that is a correspondence table between character strings representing features and identifier IDs, information on the extracted features is coded, and the representation of the features written in the character strings is digitized and stored. May be. For example, based on the feature map of common noun: 53, particle: 54, etc., information on the feature extracted for the unique expression “XX” (for example, common noun, kanji, etc.) is encoded, and the feature of “XX” It is digitized like “53: 1, 54: 0”. Here, the meaning is 1 and the meaning is 0. In this way, rapid processing by a computer is possible by digitizing.

  FIG. 2 is a diagram illustrating an example of a hardware configuration of the machine learning device 10 according to an embodiment of the present invention. The machine learning device 10 includes a CPU (Central Processing Unit) 1010 (in the multiprocessor configuration, a plurality of CPUs such as a CPU 1012 may be added), a bus line 1005, a communication I / F 1040, a main memory 1050, a BIOS (Basic Input Output System). ) 1060, USB port 1090, I / O controller 1070, input means such as a keyboard and mouse 1100, and a display device 1022.

  The CPU 1010 controls the machine learning device 10 in an integrated manner, and appropriately reads and executes various programs stored in the hard disk 1074, thereby realizing various functions according to the present invention in cooperation with the hardware described above. Yes.

  Storage means such as a tape drive 1072, a hard disk 1074, an optical disk drive 1076, and a semiconductor memory 1078 can be connected to the I / O controller 1070.

  The BIOS 1060 stores a boot program executed by the CPU 1010 when the machine learning device 10 is started, a program depending on the hardware of the machine learning device 10, and the like.

  The hard disk 1074 stores a program for the machine learning device 10 to execute the functions of the present invention, and can constitute various databases such as a teacher data DB 31, a pattern DB 33, and an NE dictionary DB 41.

  As the optical disc drive 1076, for example, a DVD-ROM drive, a CD-ROM drive, a DVD-RAM drive, or a CD-RAM drive can be used. In this case, the optical disk 1077 corresponding to each drive is used. A program or data may be read from the optical disk 1077 by the optical disk drive 1076 and provided to the main memory 1050 or the hard disk 1074 via the I / O controller 1070. Similarly, the tape medium 1071 corresponding to the tape drive 1072 can be used mainly for backup.

  The program provided to the machine learning device 10 is provided by being stored in a recording medium such as the hard disk 1074, the optical disk 1077, or a memory card. This program may be installed in the machine learning device 10 and executed by being read from a recording medium via the I / O controller 1070 or downloaded via the communication I / F 1040.

  The aforementioned program may be stored in an internal or external storage medium. Here, in addition to the hard disk 1074, the optical disk 1077, or the memory card, a magneto-optical recording medium such as an MD or a tape medium can be used as the storage medium. Alternatively, a storage device such as a hard disk or an optical disk library provided in a server system connected to a dedicated communication line or the Internet may be used as a recording medium, and the program may be provided to the machine learning device 10 via the communication line.

  Here, the display device 1022 displays a screen of the calculation processing result by the machine learning device 10, and includes a display device such as a cathode ray tube display device (CRT) or a liquid crystal display device (LCD).

  The communication I / F 1040 is a network adapter for enabling the machine learning device 10 to be connected to the corpus DB 32 or the like via a dedicated network or a public network. The communication I / F 1040 may include a modem, a cable modem, and an Ethernet (registered trademark) adapter.

  FIG. 3 is a diagram showing the corpus DB 32 according to an embodiment of the present invention.

  The corpus DB 32 stores a search log. The search log stores a search query input when the user searches for information. For example, when the user searches for the location of the branch of XX bank, the inputted “XX branch” is stored.

  FIG. 4 is a diagram showing a pattern DB 33 according to an embodiment of the present invention.

  The pattern DB 33 stores patterns. The pattern is a pattern that appears together with the teacher data. A pattern (for example, “# account”) that appears with the teacher data (for example, XX) extracted by the pattern extraction unit 13 from a corpus (for example, a search log including “XX account”, “XX branch”, etc.) "," # Branch "). Here, # represents the position where the pattern corresponding term appears. Note that the pattern DB 33 may store a pattern correspondence term and frequency together with the pattern. The pattern correspondence term is an acquired term corresponding to the pattern, and the frequency is the number of occurrences of the pattern correspondence term appearing together with the pattern. is there.

  FIG. 5 is a diagram showing the NE dictionary DB 41 according to an embodiment of the present invention.

  The NE dictionary DB 41 stores unique expressions extracted from the text of the category that matches the category of the teacher data, and stores the frequency of appearance in the text and the features of the extracted unique expressions in association with each other. Has been. In the example of FIG. 5, “XX” is a unique expression that appears at a frequency of 869 in the text of the financial category, and the feature is a proper noun, kanji, organization name,..., “XX” is the text of the financial category. This is a unique expression that appears at a frequency of 783, indicating that the feature is a proper noun, English letter, capital letter, organization name,. “Tokyo” is a unique expression that appears at a frequency of 300 in the text of the financial category, and indicates that the feature is an expression indicating a proper noun, a kanji character, a place, and so on. In addition to this information, the unique expression that appears with the unique expression of the account is the unique expression that represents the financial institution (highly likely), and the unique expression that appears with the unique expression of the branch May record the co-occurrence between unique expressions, such as is a unique expression (highly likely).

  FIG. 6 is a flowchart showing the processing contents of the machine learning device 10 according to the embodiment of the present invention. This process starts upon receiving a program start instruction, and ends upon receiving a program end instruction.

  In step S101, the CPU 1010 acquires teacher data from the teacher data DB 31. Thereafter, the CPU 1010 advances the processing to step S102.

  In step S102, the CPU 1010 acquires a corpus. More specifically, the CPU 1010 acquires a corpus having the same category as that of the teacher data from the corpus DB 32. Thereafter, the CPU 1010 advances the processing to step S103.

  In step S103, the CPU 1010 extracts a pattern from the acquired corpus. More specifically, the CPU 1010 searches the acquired corpus with the teacher data, acquires a combination with a term that appears together with the teacher data, and can arbitrarily replace the appearance portion of the teacher data in the acquired combination. It is extracted as a pattern that is considered to contain a simple term. Thereafter, the CPU 1010 advances the processing to step S104.

  In step S104, the CPU 1010 acquires a pattern correspondence term. More specifically, the CPU 1010 further searches the corpus DB 32 based on the extracted pattern, and acquires a pattern corresponding term that is a term that falls within a replaceable portion of the extracted pattern. Thereafter, the CPU 1010 advances the processing to step S105.

  In step S105, the CPU 1010 determines whether or not the acquired pattern correspondence term is stored in the NE dictionary DB 41. More specifically, the CPU 1010 searches the NE dictionary DB 41 with the acquired pattern correspondence term and determines whether or not it is stored. If the determination is YES, the process proceeds to step S106, and if the determination is NO, the process proceeds to step S108.

  In step S106, the CPU 1010, in the NE dictionary DB 41, uses the frequency associated with the searched unique expression and the co-occurrence of the pattern corresponding to the acquired pattern corresponding term to determine the specific term that the pattern corresponding term should acquire. Judge whether it is an expression. If the determination is YES, the process proceeds to step S107, and if the determination is NO, the process ends. More specifically, in the judgment, for example, “XX” appears in the text of the financial category with a frequency of 783, is a high frequency, and is a pattern corresponding term acquired from the pattern of “# account”. From the viewpoint of co-occurrence, it appears that the account has a unique expression of an account, so it is highly likely that it is a financial institution, and is determined to be a term belonging to the financial category. On the other hand, “Tokyo” has a low frequency of 300 and is a pattern-corresponding term acquired from the pattern of “# branch”. Therefore, it is determined that this is not a term to be acquired. In addition, what is necessary is just to set suitably the evaluation of how much frequency and co-occurrence are suitable, without being limited to this.

  In step S107, the CPU 1010 adds the acquired pattern correspondence term to the teacher data DB 31 as a suitable term.

  In step S108, the CPU 1010 determines whether or not the frequency with which the acquired pattern correspondence term appears together with the pattern in the pattern DB 33 is a predetermined number of times or more. If the determination is YES, the process proceeds to step S107, and if the determination is NO, the process ends.

  FIG. 7 is a flowchart showing NE dictionary creation processing of the machine learning device 10 according to an embodiment of the present invention. This process starts upon receiving a program start instruction, and ends upon receiving a program end instruction.

  In step S201, the CPU 1010 performs a morphological analysis on a category text (for example, a Web document) that matches the category of the teacher data. Thereafter, the CPU 1010 shifts the processing to step S202.

  In step S202, the CPU 1010 extracts the features of the morphemes subjected to the morphological analysis. Thereafter, the CPU 1010 advances the processing to step S203.

  In step S203, the CPU 1010 extracts the before / after information. More specifically, the CPU 1010 extracts morpheme information that appears before and after a morpheme having a specific feature among morpheme features. At this time, the appearance frequency of the morpheme is also counted. Thereafter, the CPU 1010 advances the processing to step S204.

  In step S <b> 204, the CPU 1010 acquires a unique expression based on the front and rear information. Thereafter, the CPU 1010 advances the processing to step S205.

  In step S205, the CPU 1010 stores the acquired unique expression in the NE dictionary DB 41. The NE dictionary DB 41 stores the appearance frequency and features of unique expressions. Thereafter, the CPU 1010 ends the process.

  FIG. 8 is a diagram showing the teacher data DB 31 according to an embodiment of the present invention.

  The teacher data DB 31 stores teacher data for each category. For example, in the example of FIG. 8, the category of the teacher data DB 31 is “Finance”, and “XX” whose attribute is a bank name is stored in advance as teacher data. The machine learning device 10 extracts the pattern of “# account” and “# branch” from the corpus DB 32 based on the teacher data “XX”, and stores the corpus DB 32 with the extracted pattern of “# account” and “# branch”. Search and obtain the pattern corresponding terms “XX” and “Tokyo” of the pattern. Then, the machine learning device 10 determines whether or not the acquired “xx” and “Tokyo” match the unique expression based on the NE dictionary DB 41 with reference to the feature and co-occurrence degree, Inappropriate “Tokyo” is omitted, and “xx” that matches is added to the teacher data DB 31. The example of FIG. 8 shows that “XX” is added, the attribute of “XX” that is teacher data is added to “XX”, and the bank name “XX” is added as new teacher data. ing.

  According to the embodiment, the machine learning device 10 according to the present invention includes a teacher data DB 31 that stores teacher data, and morphological analysis of morphological analysis of text in a category that matches a category of teacher data stored in advance. Extracted before and after information that is information on the morpheme before and after the morpheme that has a specific feature among the extracted features, and obtained a unique expression based on the extracted before and after information The unique expression is stored in the NE dictionary DB 41. Then, teacher data (for example, XX) is acquired from the teacher data DB 31, and combinations (for example, “XX account”, “XX branch”,...) With terms that appear together with the acquired teacher data are acquired. In the obtained combination, a pattern in which an arbitrary replaceable term is included in the appearance part of the teacher data (for example, “# account”, “# branch”, where # is an arbitrary replaceable term Pattern matching terms that are terms that can be replaced in the extracted pattern are obtained from the corpus (for example, a corpus consisting of “XX account” and “Tokyo branch”). If acquired, the terms that fall into the replaceable part corresponding to the extracted pattern are XX and Tokyo). Then, it is determined whether or not the acquired pattern correspondence term (for example, xx, Tokyo) matches a specific expression to be machine-learned based on the NE dictionary DB 41, and a term that matches based on the determination (for example, Xx) is acquired, and the acquired matching term (for example, xx) is added to the teacher data storage means as teacher data. Therefore, the machine learning device 10 according to the present invention omits inappropriate terms from terms added to teacher data for machine learning, adds terms that match teacher data, and improves the accuracy of the teacher data. be able to.

  Furthermore, the machine learning device 10 according to the present invention extracts a pattern that appears together with the teacher data from the corpus DB 32 that stores a search query that is appropriately segmented in units of words, and extracts pattern corresponding terms of the extracted pattern. get. Therefore, the machine learning device 10 according to the present invention acquires the term to be added to the teacher data for machine learning from an appropriate corpus, so the inappropriate term is omitted, the term that matches the teacher data is added, and the teacher Data accuracy can be improved.

  As mentioned above, although embodiment of this invention was described, this invention is not restricted to embodiment mentioned above. The effects described in the embodiments of the present invention are only the most preferable effects resulting from the present invention, and the effects of the present invention are limited to those described in the embodiments of the present invention. is not.

DESCRIPTION OF SYMBOLS 10 Machine learning apparatus 11 Teacher data acquisition part 12 Combination acquisition part 13 Pattern extraction part 14 Pattern corresponding term acquisition part 15 Pattern corresponding term determination part 16 Teacher data addition part 21 Morphological analysis part 22 Feature extraction part 23 Before and behind information extraction part 24 Specific expression Acquisition unit 25 Named entity storage processing unit 31 Teacher data DB
32 Corpus DB
33 Pattern DB
41 NE Dictionary DB

Claims (6)

A machine learning device in which a computer learns based on teacher data for each category given in advance,
Teacher data storage means for storing the teacher data;
Pre- and post-information that is information on morphemes before and after a morpheme having a specific feature among morpheme features obtained by morphological analysis of a category text that matches the category of the teacher data stored in advance A unique expression storage means for storing a unique expression acquired based on the preceding and following information which is a morpheme co-occurring with a morpheme having a specific feature or a morpheme immediately before and immediately after a morpheme having the specific feature Specific expression storage processing means for storing in,
Teacher data acquisition means for acquiring teacher data from the teacher data storage means;
A combination acquisition means for acquiring a combination with a term that appears together with the acquired teacher data;
In the acquired combination, pattern extraction means for extracting the appearance part of the teacher data as a pattern that is considered to contain any replaceable term; and
Pattern-corresponding term acquisition means for acquiring a pattern-corresponding term that is a term that is included in the replaceable portion of the extracted pattern from a corpus;
Pattern-corresponding term determining means for determining whether or not the pattern-corresponding term is stored as a unique expression in the specific-expression storage means;
Teacher data adding means for acquiring a pattern corresponding term determined to be stored in the specific expression storage means by the pattern corresponding term determining means, and adding the acquired pattern corresponding term as teacher data to the teacher data storage means; ,
A machine learning device comprising: For each unique expression acquired from the text, the specific expression storage processing means stores in the specific expression storage means the frequency at which the specific expression appears in the text,
When the pattern-corresponding term is stored as a unique expression in the specific-expression storage unit, the pattern-corresponding term determination unit is associated with the specific expression and the frequency stored in the specific-expression storage unit is Determine whether it is greater than or equal to a predetermined frequency threshold,
The teacher data adding unit adds the pattern corresponding term determined by the pattern corresponding term determining unit as being equal to or higher than the frequency threshold to the teacher data storage unit as teacher data. Machine learning device. The specific expression storage processing unit stores a co-occurrence degree indicating a degree of appearance of a plurality of specific expressions in the teacher data category in the specific expression storage unit;
When the pattern-corresponding term is stored in the specific-expression storage unit, the pattern-corresponding term determination unit indicates the co-occurrence degree between the pattern-corresponding term and the term acquired by the combination acquisition unit. Determining whether the acquired co-occurrence degree is equal to or greater than a predetermined co-occurrence degree threshold,
The teacher data adding means adds, to the teacher data storage means, pattern correspondence terms determined by the pattern correspondence term determination means as being equal to or greater than the co-occurrence degree threshold value as teacher data. 2. The machine learning device according to 2. For each pattern-corresponding term, it further comprises a pattern storage means for storing the number of appearances that is the number of the pattern-corresponding term acquired from the corpus by the pattern-corresponding term acquiring means,
The pattern correspondence term determining means, when the pattern correspondence term is not stored in the specific expression storage means, the number of appearances stored in the pattern storage means in association with the pattern correspondence term is a predetermined appearance threshold Determine whether or not
The said teacher data addition means adds the pattern corresponding term determined to be more than the said appearance threshold value by the said pattern corresponding term determination means to the said teacher data storage means as teacher data. The machine learning device according to any one of the above.   The machine learning device according to claim 1, wherein the corpus is a search log storing a search query. A machine learning method in which a computer learns based on teacher data storage means for storing teacher data for each category given in advance,
Pre- and post-information that is information on morphemes before and after a morpheme having a specific feature among morpheme features obtained by morphological analysis of a category text that matches the category of the teacher data stored in advance A unique expression storage means for storing a unique expression acquired based on the preceding and following information which is a morpheme co-occurring with a morpheme having a specific feature or a morpheme immediately before and immediately after a morpheme having the specific feature The step of storing in
Obtaining teacher data from the teacher data storage means;
Obtaining a combination of terms that appear with the acquired teacher data;
In the acquired combination, extracting the appearance part of the teacher data as a pattern considered to include any replaceable term; and
Obtaining a pattern-corresponding term that is a term that falls within the replaceable portion of the extracted pattern from a corpus;
Determining whether the pattern-corresponding term is stored as a unique expression in the specific expression storage means;
Obtaining a pattern-corresponding term determined to be stored in the specific expression storage means by the determination, and adding the acquired pattern-corresponding term as teacher data to the teacher data storage means;
A machine learning method comprising:

Images