JP5042268B2 - Proper word acquisition device, machine learning device and method - Google Patents

Description

  The present invention relates to a proper word acquisition device, a machine learning device, and a method.

  In recent years, in natural language processing, there has been a great deal of research on acquisition of semantic knowledge about specific expressions such as names of people and place names and relationships between them. Recognition of specific expressions and relationships, which are the subjects of these studies, plays a major role in natural language processing applications such as search systems and automatic text summarization.

  Conventionally, recognition of specific expressions and relationships has been obtained from semi-supervised learning by a bootstrap method from a large corpus. Non-patent document 1 describing a technique for learning a semantic category using a large-scale Japanese search log and dealing with a so-called semantic drift problem is described as a document describing a technique for acquiring such a specific expression or relationship. Are known.

  Here, the semantic drift problem uses a generic pattern when a pattern that co-occurs with teacher data having a category is extracted from the corpus and the extracted pattern is a pattern that matches many words. In this case, a large number of words in a category different from that of the original teacher data are acquired, and the accuracy of acquiring words in the same category as the teacher data is greatly reduced.

  In Non-Patent Document 1, as a coping method for the semantic drift problem, a pattern that co-occurs with a number of times that is greater than (for example, twice or more) the number of co-occurrence patterns that occur most frequently among the patterns that co-occur with teacher data The technique of not extracting as a generic pattern is described. In this way, as the iteration for extracting the pattern progresses, the upper limit value that is determined as a generic pattern also increases, so at the beginning of the iteration, a pattern with a high precision and a low recall is extracted and gradually increased in frequency. A recall pattern can be extracted.

Mamoru Komachi, “Improvement of semi-supervised acquisition of semantic knowledge from search logs”, Journal of Population Intelligence, Vol. 23, No. 3 F (2008), pp. 217-225

  However, even in the method described in Non-Patent Document 1, there are many variables (for example, the number of iterations, addition of high-order words, how many seeds are given, etc.) to match the data used as a corpus. The above variables) must be set, and the accuracy varies greatly depending on the setting of these variables. Therefore, it is essential to adjust the variables optimally, which is an obstacle to actual operation.

  Therefore, there is a need for an apparatus that can accurately and accurately acquire words in the same category as the teacher data without adjusting many variables.

  An object of the present invention is to provide an apparatus and method that can accurately and appropriately acquire words in the same category as teacher data without adjusting many variables.

  The present invention provides the following solutions.

(1) An appropriate word acquisition device that acquires an appropriate pattern-corresponding word based on a pattern that co-occurs with teacher data for each category given in advance and a pattern-corresponding word that is associated with the pattern, A network score storage unit that regards the corresponding word and the pattern as nodes and stores an edge score, which is a score of an edge connecting the nodes, in association with the edge; and one of the pattern corresponding word and the pattern An initial scoring means for assigning a predetermined initial score to the pattern-corresponding word or the pattern based on the score assigned to the one of the patterns and the edge score, and the other of the pattern-corresponding word or the pattern A first scoring means for assigning a score to the first score; Second scoring means for assigning a score to the one of the pattern-corresponding word or the pattern based on the score given to the other by the alling means and the edge score; and the first scoring An appropriate word acquisition means for repeating the assignment of the score by the means and the assignment of the score by the second scoring means to acquire the pattern corresponding word satisfying a predetermined score condition, and the network score storage means Is stored in a matrix in which the pattern-corresponding word and the pattern are associated with each other, and an edge score between the pattern-corresponding word and the pattern is stored as an element in the matrix, and the first scoring means The second scoring means is connected to the matrix stored in the network score storage means. Te, has been determined based on the maximum eigenvalue λ of the matrix, proper word acquiring unit and performs the calculation of scores using a variable indicating the importance of the initial value.

  According to the configuration of (1), the proper word acquisition apparatus according to the present invention regards the pattern correspondence word and the pattern as nodes, and associates an edge score, which is a score of an edge connecting the nodes, with the edge. Network score storage means for storing a predetermined initial score to one of the pattern-corresponding word or pattern. Then, based on the score assigned to one of the pattern-corresponding words or patterns and the edge score, the score is assigned to the other of the pattern-corresponding words or patterns, and the score assigned to the other and the edge score Based on the pattern-corresponding word or pattern, a score is assigned to one of the patterns, and the addition of the score to the other and the provision of the score to the other are repeated to obtain a pattern-corresponding word that satisfies a predetermined score condition.

  That is, the appropriate word acquisition device according to the present invention regards the pattern-corresponding word and the pattern as a network structure, repeats scoring based on the edge score by the network score storage means and the initial score of the node, A pattern-corresponding word that satisfies the score condition is acquired. Therefore, the appropriate word acquisition device according to the present invention acquires the pattern-corresponding word based on the score representing the degree of association with the correct answer data (word that matches the category to be acquired) based on the initial score. With this variable, pattern-corresponding words in the same category as the category of the teacher data can be acquired. Therefore, the appropriate word acquisition apparatus according to the present invention can acquire words of the same category as the teacher data with high accuracy and appropriate without adjusting many variables.

  (2) The predetermined scoring condition includes the pattern-corresponding word score given by the first scoring means or the second scoring means, and the first scoring means immediately before giving the score. Or, when the difference from the score of the pattern corresponding word provided by the second scoring means is all equal to or less than a certain value, the score of the pattern corresponding word is equal to or higher than a predetermined value, The proper word acquisition device according to (1).

  According to the configuration of (2), in the appropriate word acquisition device described in (1), the difference between the score of the assigned pattern-corresponding word and the score of the pattern-corresponding word assigned immediately before giving the score is constant. When the value is equal to or less than the value, a pattern-corresponding word satisfying that the score of the pattern-corresponding word is equal to or greater than a predetermined value is acquired.

  Therefore, the proper word acquisition apparatus according to the present invention repeats scoring by regarding it as a network structure, the repeated scoring converges, and the pattern-corresponding word whose score indicating the degree of association with correct data is equal to or greater than a predetermined value Therefore, a pattern corresponding word in the same category as that of the teacher data can be acquired. Therefore, the appropriate word acquisition apparatus according to the present invention can appropriately acquire words of the same category as the teacher data with high accuracy based on the score.

(3) A machine learning device that extracts a pattern that co-occurs with teacher data for each category given in advance from a corpus and acquires a pattern-corresponding word associated with the extracted pattern, the pattern-corresponding word; A network score storage means that regards each of the patterns as a node and stores an edge score, which is a score of an edge connecting the nodes, in association with the edge; and a predetermined initial value in one of the pattern correspondence word or the pattern Based on the initial scoring means for assigning a score, the score assigned to the one of the pattern-corresponding word or the pattern, and the edge score, the score is assigned to the other of the pattern-corresponding word or the pattern First scoring means and the first scoring means By the second scoring means for assigning a score to the one of the pattern-corresponding words or the pattern based on the score given to the other by the above and the edge score, and by the first scoring means An appropriate word acquisition unit that repeats the score assignment and the score assignment by the second scoring unit and acquires the pattern-corresponding word satisfying a predetermined score condition, and the network score storage unit stores A network that associates the pattern-corresponding word with the pattern, stores an edge score of the pattern-corresponding word and the pattern as an element in the matrix, and the first scoring means and the first The second scoring means applies the row to the matrix stored in the network score storage means. Maximum eigenvalue is determined based on lambda, machine learning device characterized by the calculation of scores using a variable indicating the importance of the initial value of.

  According to the configuration of (3), the machine learning device according to the present invention regards a pattern correspondence word and a pattern as nodes, and associates an edge score, which is a score of an edge connecting the nodes, with the edge. Network score storage means for storing is provided, and a predetermined initial score is given to one of the pattern-corresponding word or the pattern. Then, based on the score assigned to one of the pattern-corresponding words or patterns and the edge score, the score is assigned to the other of the pattern-corresponding words or patterns, and the score assigned to the other and the edge score Based on the pattern-corresponding word or pattern, a score is assigned to one of the patterns, and the addition of the score to the other and the provision of the score to the other are repeated to obtain a pattern-corresponding word that satisfies a predetermined score condition. Then, the machine learning device according to the present invention adds the acquired pattern correspondence word to the teacher data.

  That is, the machine learning device according to the present invention regards the pattern-corresponding word and the pattern as a network structure, repeats scoring based on the edge score by the network score storage means and the initial score of the node, and obtains a predetermined score. Get a pattern-corresponding word that satisfies a condition. Therefore, the machine learning device according to the present invention acquires the pattern-corresponding word based on the score representing the degree of association with the correct answer data based on the initial score. Pattern-corresponding words can be acquired. Therefore, the machine learning device according to the present invention can accurately and accurately acquire words in the same category as the teacher data without adjusting many variables. Furthermore, since the acquired pattern correspondence word is a proper word that satisfies the score condition, the machine learning device according to the present invention can improve accuracy each time the acquired pattern correspondence word is added to the teacher data.

  (4) The predetermined score condition includes: the score of the pattern-corresponding word given by the first scoring means or the second scoring means; and the first scoring means immediately before giving the score. Or, when the difference from the score of the pattern corresponding word provided by the second scoring means is all equal to or less than a certain value, the score of the pattern corresponding word is equal to or higher than a predetermined value, The machine learning device according to (3).

  According to the configuration of (4), the machine learning device according to (3) is such that the difference between the score of the assigned pattern-corresponding word and the score of the pattern-corresponding word assigned immediately before giving the score is constant. When the value is equal to or less than the value, a pattern-corresponding word that satisfies that the score of the pattern-corresponding word is equal to or greater than a predetermined value is acquired.

  Therefore, the machine learning device according to the present invention repeats scoring by regarding it as a network structure, the repeated scoring converges, and a pattern corresponding word having a score representing a degree of association with correct data is a predetermined value or more. Since it is acquired, it is possible to acquire a pattern corresponding word in the same category as the category of the teacher data. Therefore, the machine learning device according to the present invention can accurately and accurately acquire words in the same category as the teacher data based on the score.

(5) In the appropriate word acquisition apparatus that acquires an appropriate pattern-corresponding word based on a pattern that co-occurs with teacher data for each category given in advance and a pattern-corresponding word associated with the pattern, the pattern-corresponding word And a method executed by a proper word acquisition device including network score storage means that regards the pattern as a node and stores an edge score that is a score of an edge connecting the nodes in association with the edge, Based on the step of assigning a predetermined initial score to one of the pattern correspondence word or the pattern, the score assigned to the one of the pattern correspondence word or the pattern, and the edge score, the pattern correspondence Assigning a score to the other of the words or the pattern And assigning a score to the one of the pattern-corresponding words or the pattern based on the score given to the other by the step of assigning a score to the other and the edge score, and to the other Repetitively repeating the step of assigning a score and the step of assigning a score to the one, and obtaining the pattern-corresponding word satisfying a predetermined score condition, and the network stored by the network score storage means, A matrix in which the pattern-corresponding word and the pattern are associated with each other, storing an edge score between the pattern-corresponding word and the pattern as an element in the matrix, and assigning a score to the other The assigning step corresponds to the matrix stored in the network score storage means. Te, has been determined based on the maximum eigenvalue λ of the matrix, wherein the performing the calculation of the score using a variable indicating the importance of the initial value.

  That is, the method according to the present invention regards the pattern-corresponding word and the pattern as a network structure, repeats scoring based on the edge score by the network score storage means and the initial score of the node, and sets a predetermined score condition. Get pattern-corresponding words that satisfy. Therefore, the method according to the present invention obtains the pattern correspondence word based on the score representing the degree of relevance with the correct answer data based on the initial score, so the pattern correspondence of the same category as the category of the teacher data is determined by the variable at the initial score. You can get a word. Therefore, the method according to the present invention can appropriately and accurately acquire words in the same category as the teacher data without adjusting many variables.

(6) In a machine learning device that extracts a pattern co-occurrence with teacher data for each given category from a corpus and obtains a pattern correspondence word associated with the extracted pattern, the pattern correspondence word and the pattern Is a method executed by a machine learning device including a network score storage unit that stores an edge score, which is a score of an edge connecting between the nodes, in association with the edge, the pattern corresponding word or Based on the step of assigning a predetermined initial score to one of the patterns, the score assigned to the one of the pattern-corresponding word or the pattern, and the edge score, the pattern-corresponding word or the pattern A step of assigning a score to the other, and a step A step of assigning a score to the one of the pattern-corresponding words or the pattern, and a step of assigning a score to the other, based on the score given to the other by the step of assigning a core and the edge score And assigning a score to the one, and obtaining the pattern-corresponding word satisfying a predetermined score condition, and the network stored by the network score storage means, The matrix corresponding to the pattern, and storing the edge score of the pattern corresponding word and the pattern as an element in the matrix, the step of assigning a score to the other and the step of assigning a score to the one, For the matrix stored in the network score storage means, It has been determined based on the large eigenvalue lambda, wherein the performing the calculation of the score using a variable indicating the importance of the initial value.

  That is, the method according to the present invention regards the pattern-corresponding word and the pattern as a network structure, repeats scoring based on the edge score by the network score storage means and the initial score of the node, and sets a predetermined score condition. Get pattern-corresponding words that satisfy. Therefore, the method according to the present invention obtains the pattern correspondence word based on the score representing the degree of relevance with the correct answer data based on the initial score, so the pattern correspondence of the same category as the category of the teacher data is determined by the variable at the initial score. You can get a word. Therefore, the method according to the present invention can appropriately and accurately acquire words in the same category as the teacher data without adjusting many variables. Furthermore, since the pattern correspondence word acquired by the method according to the present invention is an appropriate word that satisfies the score condition, the machine learning device that executes the method according to the present invention increases the accuracy every time the acquired pattern correspondence word is added to the teacher data. Can be improved.

  According to the present invention, the acquired pattern correspondence word and the extracted pattern are regarded as a network structure, scoring is repeated based on the edge score by the network score storage means and the initial score of the node, and a predetermined score condition Pattern-corresponding words that satisfy

  Furthermore, according to the present invention, repeated scoring converges, and a pattern-corresponding word having a score representing a degree of relevance with teacher data given in advance can be acquired.

  Therefore, according to the present invention, words in the same category as the teacher data can be acquired accurately and appropriately without adjusting many variables. Furthermore, the machine learning device according to the present invention can improve the accuracy each time the acquired pattern correspondence word is added to the teacher data.

It is a block diagram which shows the function structure of Example 1 of the appropriate word acquisition apparatus 10 which concerns on one Embodiment of this invention. It is a figure which shows an example in which the appropriate word acquisition apparatus 10 which concerns on one Embodiment of this invention calculates the score of a pattern corresponding word based on the network structure which considered the pattern corresponding word and the pattern each as a node. It is a figure which shows an example of the hardware constitutions of the appropriate word acquisition apparatus 10 which concerns on one Embodiment of this invention. It is a figure which shows network score DB21 of the appropriate word acquisition apparatus 10 which concerns on one Embodiment of this invention. It is a flowchart which shows the process of the appropriate word acquisition apparatus 10 which concerns on one Embodiment of this invention. It is a block diagram which shows the function structure of the machine learning apparatus 30 which concerns on one Embodiment of this invention. It is a figure which shows the example of teacher data DB41 of the machine learning apparatus 30 which concerns on one Embodiment of this invention.

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

[Example 1]
FIG. 1 is a block diagram showing a functional configuration of Example 1 of an appropriate word acquisition device 10 according to an embodiment of the present invention. The proper word acquisition device 10 includes a network score DB 21 as a network score storage unit, an initial scoring unit 11 as an initial scoring unit, a first scoring unit 12 as a first scoring unit, and a second scoring unit. As a second scoring unit 13 and an appropriate word acquisition unit 14 as appropriate word acquisition means. Then, the appropriate word acquisition apparatus 10 regards the pattern-corresponding word and the pattern as a network structure, repeats scoring based on the edge score by the network score DB 21 and the initial score of the node, and satisfies a predetermined score condition Get pattern correspondence word.

  The appropriate word acquisition device 10 acquires an appropriate pattern-corresponding word based on a pattern that co-occurs with teacher data for each given category and a pattern-corresponding word that is associated with the pattern. That is, the teacher data is data given in advance with a known category. Here, the pattern and the pattern-corresponding word may be a pattern and a pattern-corresponding word acquired from the corpus by the machine learning device, or may be created manually.

  The corpus is text data composed of natural language sentences. As the corpus, for example, a search query log that is a search query log for searching in a computer network may be used. Moreover, you may use the search click through log which is the log clicked from the web page searched by the search query. Click-through refers to which web page was actually clicked for a given search query, and for search queries that do not necessarily guarantee that the search results were correct, the page title and address , Information that refers to the text after looking at the snippet (summary of context around the search query). In particular, two different search queries that reach the exact same page are often synonymous, and the category can be recognized with high accuracy. Therefore, it is expected to improve both accuracy and coverage by using the search click-through log as a corpus in addition to the search query log.

  The pattern includes a co-occurrence pattern and an address pattern. Here, the co-occurrence pattern is obtained when, for example, a combination with a term appearing together with the teacher data is acquired, and in the obtained combination, an appearance part of the teacher data can be regarded as containing any replaceable term. This is a part excluding the appearance part of the teacher data. In addition, when a combination with a term that appears together with a pattern corresponding word corresponding to the co-occurrence pattern is acquired, and in the acquired combination, it can be considered that an appearance part of the pattern corresponding word includes any replaceable term The portion excluding the appearance portion of the pattern corresponding word is also included.

  The address pattern is, for example, the search click-through log, and the address information of the web page clicked through, for example, the URL (Uniform Resource Locator) of the web page.

  Examples of the pattern-corresponding word associated with the pattern include a term combined with the co-occurrence pattern and a search query corresponding to the address pattern. The pattern-corresponding word may include teacher data, which is correct data given in advance, or may not include teacher data.

  The network score DB 21 regards each pattern-corresponding word and pattern as a node, and stores an edge score that is a score of an edge connecting the nodes in association with the edge. That is, when the pattern correspondence word and the pattern are regarded as nodes as the network structure, the network score DB 21 stores the edge score in association with the edge connecting the nodes.

  Here, the edge score associated with the pattern-corresponding word and the edge of the pattern is obtained based on the frequency of the combination of the pattern-corresponding word and the pattern. For the edge score, for example, the frequency may be used as it is, it may be normalized to become a probability value based on the frequency, a self-mutual information amount may be used, or a log likelihood ratio may be used.

  This network structure has the property that pattern-corresponding words that co-occur with related patterns are related. That is, there is a high possibility that patterns co-occurring with the same pattern correspondence word are related to each other. A pattern corresponding word that co-occurs with many of the related patterns is likely to be related to the original pattern corresponding word. Similarly, pattern correspondence words that co-occur with the same pattern are likely to be related. A pattern that co-occurs with many of the related pattern correspondence words is likely to be related to the original pattern. Using such a property, it is possible to efficiently calculate a measure of the degree of association between pattern-corresponding words.

  That is, the appropriate word acquisition apparatus 10 regards the pattern and the pattern-corresponding word as a network structure, and calculates a score that is a measure of the degree of association between the pattern-corresponding word and the category to be acquired. Since the degree of association based on the calculated score represents the sameness between the category of the teacher data (the category to be acquired) and the category of the pattern-corresponding word, the appropriate word acquisition device 10 performs the teacher with high accuracy based on the calculated score. Pattern-corresponding words in the same category as the data can be acquired.

  Here, as a matter of course, since the teacher data is a word given in advance and is correct data (a word that matches the category to be acquired), the pattern-corresponding word not including the teacher data and the pattern are regarded as nodes. The score may be calculated in the network structure.

  The initial scoring unit 11 assigns a predetermined initial score to one of the pattern correspondence word or the pattern. That is, in order to analyze the network structure composed of nodes and edges, for example, an initial score is given to a pattern corresponding word as an initial score, and the analysis is started. Alternatively, as an initial score, for example, an initial score is given to the pattern and the analysis is started.

  As the predetermined initial score, for example, 1 is assigned to at least one correct answer data, and 0 is assigned to other data whose category is unknown. Here, the correct answer data refers to patterns that include teacher data and co-occur with pattern-corresponding words or teacher data in the same category as the teacher data. For example, in the case of a network structure including teacher data, it is only necessary to add 1 to at least one teacher data or a pattern co-occurring with the teacher data. In the case of a network structure not including teacher data, The selection of correct data among the pattern-corresponding words or patterns is accepted from the administrator, and 1 is assigned to at least one selected correct data.

  The first scoring unit 12 assigns a score to the other of the pattern-corresponding words or patterns based on the score assigned to one of the pattern-corresponding words or patterns and the edge score. That is, the first scoring unit 12 assigns a score to the pattern based on the score assigned to the pattern corresponding word and the edge score. Alternatively, a score is assigned to the pattern-corresponding word based on the score assigned to the pattern and the edge score. The given score is the initial score given by the initial scoring unit 11 for the first time, and the score given by the second scoring unit 13 for the second and subsequent times.

  The second scoring unit 13 assigns a score to one of the pattern-corresponding words or patterns based on the score given to the other by the first scoring unit 12 and the edge score. That is, the second scoring unit 13 gives a score to the pattern corresponding word based on the score given to the pattern by the first scoring unit 12 and the edge score. Or the 2nd scoring part 13 provides a score to a pattern based on the score provided to the pattern corresponding | compatible word by the 1st scoring part 12, and an edge score.

  The score given to a pattern is expressed by a matrix in which a pattern-corresponding word is a row and the pattern is a column.

Here, p _i is the current score of pattern i, x _j is the current score of pattern corresponding word j, and w _{i, j} is the edge score of pattern i and pattern corresponding word j. Also, let W be a matrix whose elements are w _{i, j} . This edge score is obtained by referring to the network score DB 21.

  The score given to the pattern corresponding word is calculated by Equation 2.

Here, y _j is an initial score of the pattern corresponding word j. α is a variable (for example, 0.9) indicating the importance of the initial value, and 0 <α <1. Here, α is α <1 / λ where λ is the maximum eigenvalue of the matrix W. As described above, since only one variable indicating the importance of correct answer data given first is set, the portability is excellent.

  In addition, the example of Numerical formula 1 and Numerical formula 2 is a case where an initial score is provided to the pattern corresponding word. When the initial score is given to the pattern, the score given to the pattern corresponding word is calculated by Equation 1 and the score given to the pattern is calculated by Equation 2. Here, x and p are interchanged, and y is the initial score of the pattern.

  The appropriate word acquisition unit 14 repeats the provision of the score to the pattern or the pattern-corresponding word by the first scoring unit 12 and the provision of the score to the pattern-corresponding word or the pattern by the second scoring unit 13, A pattern-corresponding word that satisfies the score condition is acquired.

  By using a technique of parallel processing programming to repeat the assignment of a score to a pattern and the assignment of a score to a pattern-corresponding word, the appropriate word acquisition device 10 can perform a search that expects a large amount of traffic such as a web search. The system can also handle this.

  The predetermined score condition is, for example, when the difference between the score of the assigned pattern-corresponding word and the score of the pattern-corresponding word given immediately before the score is all equal to or less than a certain value. The condition is that the score of the corresponding word is not less than a predetermined value. Here, the case where all the differences are equal to or less than a certain value means that the score value given repeatedly has converged, and that the score of the pattern-corresponding word is equal to or greater than a predetermined value means correct data It is highly likely that the category is the same as the category. That is, when the repeated scoring converges, the appropriate word acquisition unit 14 acquires pattern-corresponding words that are likely to be in the same category as the correct answer data.

  The predetermined score condition may be repeated a certain number of times (for example, 40 times), and when the score rank does not change, a pattern-corresponding word that is higher in the score rank may be acquired. By determining that the ranking does not change after a certain number of iterations, the amount of calculation can be reduced.

  FIG. 2 is a diagram illustrating an example in which the appropriate word acquisition apparatus 10 according to an embodiment of the present invention calculates a pattern-corresponding word and a score of the pattern-corresponding word based on a network structure in which each pattern is regarded as a node. .

  FIG. 2 (1) shows “xxx child” which is teacher data (for example, category is actress), pattern corresponding words “Nara”, “xxx child”, patterns “# + image”, “#”. FIG. 4 is a diagram illustrating that a network structure is configured by regarding “+ photo book”, “# + hotel”, and “# + tourism” as nodes and connecting nodes in a co-occurrence relationship with edges. That is, “xxx child” that is teacher data appears as “xxx child_image” and “xxx child_photo book” when the corpus is a search log (_ represents a space). ). Therefore, the patterns “# + image” and “# + photo book” are extracted. Based on the extracted patterns, “Nara_photo book”, “xxx child_image”, “xxx child_photo book” are extracted, and pattern corresponding words “Nara”, “xxx child” And represent it as a network structure. In addition, the patterns “# + hotel” and “# + tourism” co-occurring with “Nara” are also extracted.

  FIG. 2 (2) is a diagram showing the pattern correspondence word and the edge score of the pattern in a matrix format. The proper word acquisition device 10 stores such a reference edge score in the hard disk 1074 (see FIG. 3 described later) or the like based on the network score DB 21.

  In FIG. 2 (3), as the initial score for the first time, a score of 1 is assigned to the teacher data “xxx child”, and pattern corresponding words other than “xxx child” are “Nara”, “ XX child ”is given a score of 0, the pattern score is calculated based on the assigned score and the edge score, and the teacher data and the pattern correspondence are calculated based on the calculated pattern score and the edge score As a result of repeating the calculation of the word score, the difference between the score of the pattern-corresponding word given for the 40th time and the score of the pattern-corresponding word given for the 39th time immediately before giving the score are all constant values (for example, , 0.01) or less, the pattern corresponding word “XX child” satisfying that the score of the pattern corresponding word is equal to or higher than a predetermined value (for example, 0.3) is an appropriate pattern corresponding single It represents that can be acquired as. The proper word acquisition apparatus 10 stores such a calculation process in a hard disk 1074 (see FIG. 3 described later) or the like.

  FIG. 3 is a diagram illustrating an example of a hardware configuration of the proper word acquisition device 10 according to an embodiment of the present invention.

  The proper word acquisition apparatus 10 includes a central processing unit (CPU) 1010 (a plurality of CPUs such as CPU_A 1012 may be added in a multiprocessor configuration) constituting the control unit 1000, a bus line 1005, a communication I / F 1040, a main memory 1050. , A BIOS (Basic Input Output System) 1060, an I / O controller 1070, a hard disk 1074, an optical disk drive 1076, a semiconductor memory 1078, a display device 1080, and an input device 1090.

  The control unit 1000 is a part that controls the appropriate word acquisition device 10 in an integrated manner, and by reading and executing various programs stored in the hard disk 1074 as appropriate, it cooperates with the hardware described above, and relates to the present invention. Various functions are realized.

  The communication I / F 1040 is a network adapter that enables the proper word acquisition apparatus 10 to be connected to another server 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.

  The BIOS 1060 stores a boot program executed by the CPU 1010 when the appropriate word acquisition device 10 is activated, a program depending on the hardware of the appropriate word acquisition device 10, and the like.

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

  The hard disk 1074 stores a program for the proper word acquisition apparatus 10 to execute the function of the present invention, and stores a database such as the network score DB 21.

  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.

  The program provided to the appropriate word acquisition apparatus 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 read from the recording medium via the I / O controller 1070 or downloaded via the communication I / F 1040 to be installed and executed in the proper word acquisition apparatus 10.

  The display device 1080 displays a screen of a calculation processing result by the appropriate word acquisition device 10 and includes a display device such as a cathode ray tube display device (CRT) or a liquid crystal display device (LCD).

  The input device 1090 accepts input to the appropriate word acquisition device 10 and includes a keyboard and a mouse.

  FIG. 4 is a diagram showing the network score DB 21 of the proper word acquisition device 10 according to an embodiment of the present invention. The network score DB 21 associates a word with a corpus type (for example, a query for a search query log, a click for a click-through log, etc.), a pattern (for example, a co-occurrence pattern, an address pattern), a frequency, , Memorize the edge score. The frequency is a frequency at which a word appears together with a pattern, and the edge score is a score given to an edge created based on the frequency for each word, for example.

  FIG. 5 is a flowchart showing processing of the appropriate word acquisition device 10 according to an embodiment of the present invention. Note that this process starts upon receiving a program start command, executes the program, and ends.

  In step S101, the CPU 1010 gives a predetermined initial score to one of the pattern-corresponding word or pattern. More specifically, the CPU 1010 assigns 1 to at least one correct answer data (for example, teacher data) and 0 to other data, and stores them in the hard disk 1074. That is, when an initial score is given to a pattern-corresponding word, 1 is assigned to at least one correct answer data (for example, teacher data) among the pattern-corresponding words, and 0 is assigned to a pattern-corresponding word to which a category is to be assigned. . When an initial score is assigned to a pattern, 1 is assigned to at least one correct answer data (for example, a pattern co-occurring with teacher data), and 0 is assigned to another pattern. Thereafter, the CPU 1010 advances the processing to step S102.

  In step S102, the CPU 1010 assigns a score to the other of the pattern-corresponding words or patterns based on the score assigned to one of the pattern-corresponding words or patterns and the edge score. More specifically, the CPU 1010 determines that the initial score is a pattern-corresponding word based on the score given to the pattern-corresponding word in step S101 or step S103 and the edge score acquired with reference to the network score DB 21. Formula 1 is calculated and a score is given to the pattern. Alternatively, the CPU 1010 calculates Formula 1 when the initial score is a pattern based on the score given to the pattern in step S101 or step S103 and the edge score acquired with reference to the network score DB 21, and the pattern Give a score to the corresponding word. The CPU 1010 stores the assigned score in the hard disk 1074. Thereafter, the CPU 1010 advances the processing to step S103.

  In step S103, the CPU 1010 assigns a score to one of the pattern-corresponding words or patterns based on the score assigned to the other in step S102 and the edge score. More specifically, the CPU 1010 calculates Formula 2 when the initial score is a pattern-corresponding word based on the score given to the pattern in step S102 and the edge score acquired with reference to the network score DB 21. And give a score to the pattern-corresponding word. Alternatively, the CPU 1010 calculates Formula 2 when the initial score is a pattern based on the score given to the pattern-corresponding word in step S102 and the edge score acquired with reference to the network score DB 21 to obtain the pattern. Give a score. The CPU 1010 stores the assigned score in the hard disk 1074. Thereafter, the CPU 1010 advances the processing to step S104.

  In step S104, the CPU 1010 determines whether or not the score condition is satisfied. More specifically, the CPU 1010 has a constant difference between the score of the assigned pattern-corresponding word stored in the hard disk 1074 and the score of the pattern-corresponding word assigned immediately before giving the score for all the pattern-corresponding words. It is determined whether or not the value is less than or equal to. If the determination is YES, the process proceeds to step S105, and if the determination is NO, the process proceeds to step S102.

  In step S105, the CPU 1010 acquires a proper word. More specifically, the CPU 1010 acquires a pattern corresponding word stored in the hard disk 1074 and having a pattern corresponding word score equal to or higher than a predetermined value. Thereafter, the CPU 1010 ends the process.

[Example 2]
FIG. 6 is a block diagram showing a functional configuration of the machine learning device 30 according to the embodiment of the present invention. The machine learning device 30 includes a bootstrap unit 35, a proper word acquisition device 10, a teacher data adding unit 36, a teacher data DB 41, and a network score DB 21.

  The bootstrap unit 35 extracts words from the corpus by bootstrap based on the existing technology. For example, a corpus that matches the category of the teacher data acquired from the teacher data DB 41 (see FIG. 7 to be described later) is acquired from the corpus DB 42, the acquired corpus is morphologically analyzed, and the teacher data is extracted from the corpus based on the morphological analysis. And a pattern corresponding word associated with the extracted pattern is acquired. Then, the extracted pattern and the acquired pattern correspondence word are given to the appropriate word acquisition device 10.

  As described above, the appropriate word acquisition device 10 includes the network score DB 21, the initial scoring unit 11, the first scoring unit 12, the second scoring unit 13, and the appropriate word acquisition unit 14, Given pattern-corresponding words and patterns are regarded as a network structure, scoring is repeated based on the edge score by the network score DB 21 and the initial score of the node, and pattern-corresponding words satisfying a predetermined score condition are obtained.

  The teacher data adding unit 36 adds the pattern correspondence word acquired by the appropriate word acquisition device 10 to the teacher data DB 41. Therefore, the machine learning device 30 can efficiently create the teacher data DB 41 with high accuracy.

  FIG. 7 is a diagram illustrating an example of the teacher data DB 41 of the machine learning device 30 according to an embodiment of the present invention. The teacher data DB 41 stores categories in association with words.

  The hardware configuration of the machine learning device 30 is the same as the hardware configuration of the proper word acquisition device 10 (FIG. 3).

  According to the first embodiment, the appropriate word acquisition apparatus 10 regards the pattern-corresponding word and the pattern as nodes, and stores the edge score that is the score of the edge connecting the nodes in association with the edge. A DB 21 is provided, and a predetermined initial score is given to one of the pattern correspondence word or pattern. Then, based on the score assigned to one of the pattern-corresponding words or patterns and the edge score, the score is assigned to the other of the pattern-corresponding words or patterns, and the score assigned to the other and the edge score Based on the pattern-corresponding word or pattern, a score is assigned to one of the patterns, and the addition of the score to the other and the provision of the score to the other are repeated to obtain a pattern-corresponding word that satisfies a predetermined score condition. Therefore, the appropriate word acquisition device 10 acquires the pattern-corresponding word based on the score representing the degree of association with the correct answer data based on the initial score. Pattern-corresponding words can be acquired. Therefore, the appropriate word acquisition apparatus according to the present invention can acquire words of the same category as the teacher data with high accuracy and appropriate without adjusting many variables.

  Furthermore, the appropriate word acquisition device 10 determines that the pattern correspondence word when the difference between the score of the assigned pattern correspondence word and the score of the pattern correspondence word assigned immediately before the application of the score is equal to or less than a certain value. A pattern-corresponding word that satisfies that the score is equal to or greater than a predetermined value is acquired. Therefore, the appropriate word acquisition apparatus 10 repeats scoring according to the network structure, the repeated scoring converges, and a pattern corresponding word whose score representing the degree of association with teacher data given in advance is a predetermined value or more is selected. Since it is acquired, it is possible to acquire a pattern corresponding word in the same category as the category of the teacher data. Therefore, the appropriate word acquisition device 10 can appropriately acquire words of the same category as the teacher data with high accuracy based on the score.

  According to the second embodiment, the machine learning device 30 includes the bootstrap unit 35, the appropriate word acquisition device 10, and the teacher data addition unit 36. The pattern extracted by the bootstrap unit 35 and the acquired pattern corresponding word To the appropriate word acquisition device 10, the appropriate word acquisition device 10 acquires an appropriate pattern correspondence word, and the teacher data addition unit 36 adds the acquired pattern correspondence word to the teacher data DB 41. Therefore, the machine learning device 30 acquires the pattern-corresponding words in the same category as the teacher data appropriately with high accuracy based on the network score DB 21 without adjusting many variables. Each time it is added to the DB 41, the accuracy can be improved, and the teacher data DB 41 can be created efficiently.

  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 Proper word acquisition apparatus 11 Initial scoring part 12 1st scoring part 13 2nd scoring part 14 Proper word acquisition part 21 Network score DB
30 Machine Learning Device 35 Bootstrap Unit 36 Teacher Data Adding Unit 41 Teacher Data DB

Claims (6)

An appropriate word acquisition device that acquires an appropriate pattern correspondence word based on a pattern that co-occurs with teacher data for each category given in advance and a pattern correspondence word associated with the pattern,
A network score storage means that regards the pattern-corresponding word and the pattern as nodes, and stores an edge score, which is a score of an edge connecting the nodes, in association with the edge;
An initial scoring means for giving a predetermined initial score to one of the pattern-corresponding word or the pattern;
A first scoring means for assigning a score to the other of the pattern-corresponding word or the pattern based on the score given to the one of the pattern-corresponding word or the pattern and the edge score;
A second scoring means for assigning a score to the one of the pattern-corresponding word or the pattern based on the score given to the other by the first scoring means and the edge score;
Appropriate word acquisition means for repeating the provision of the score by the first scoring means and the provision of the score by the second scoring means, and acquiring the pattern-corresponding word satisfying a predetermined score condition;
Equipped with a,
The network stored by the network score storage means is a matrix in which the pattern-corresponding word is associated with the pattern, and the edge score between the pattern-corresponding word and the pattern is stored as an element in the matrix,
The first scoring means and the second scoring means are determined based on the maximum eigenvalue λ of the matrix with respect to the matrix stored in the network score storage means. A proper word acquisition device, wherein a score is calculated using a variable indicating degree .   The predetermined score condition includes a score of the pattern-corresponding word given by the first scoring means or the second scoring means, and the first scoring means or the first score just before giving the score. The score of the pattern-corresponding word is equal to or greater than a predetermined value when all the differences from the score of the pattern-corresponding word provided by the second scoring means are equal to or less than a predetermined value. The proper word acquisition device according to 1. A machine learning device that extracts a pattern that co-occurs with teacher data for each given category from a corpus and acquires a pattern-corresponding word that is associated with the extracted pattern,
A network score storage means that regards the pattern-corresponding word and the pattern as nodes, and stores an edge score, which is a score of an edge connecting the nodes, in association with the edge;
An initial scoring means for giving a predetermined initial score to one of the pattern-corresponding word or the pattern;
A first scoring means for assigning a score to the other of the pattern-corresponding word or the pattern based on the score given to the one of the pattern-corresponding word or the pattern and the edge score;
A second scoring means for assigning a score to the one of the pattern-corresponding word or the pattern based on the score given to the other by the first scoring means and the edge score;
Appropriate word acquisition means for repeating the provision of the score by the first scoring means and the provision of the score by the second scoring means, and acquiring the pattern-corresponding word satisfying a predetermined score condition;
Equipped with a,
The network stored by the network score storage means is a matrix in which the pattern-corresponding word is associated with the pattern, and the edge score between the pattern-corresponding word and the pattern is stored as an element in the matrix,
The first scoring means and the second scoring means are determined based on the maximum eigenvalue λ of the matrix with respect to the matrix stored in the network score storage means. A machine learning device that calculates a score using a variable indicating degree .   The predetermined score condition includes a score of the pattern-corresponding word given by the first scoring means or the second scoring means, and the first scoring means or the first score just before giving the score. The score of the pattern-corresponding word is equal to or greater than a predetermined value when all the differences from the score of the pattern-corresponding word provided by the second scoring means are equal to or less than a predetermined value. 3. The machine learning device according to 3. In an appropriate word acquisition device that acquires an appropriate pattern-corresponding word based on a pattern that co-occurs with teacher data for each given category and a pattern-corresponding word that is associated with the pattern, the pattern-corresponding word, and It is a method executed by an appropriate word acquisition device including network score storage means that regards a pattern as a node and stores an edge score that is a score of an edge connecting the nodes in association with the edge,
Giving a predetermined initial score to one of the pattern-corresponding word or the pattern;
Assigning a score to the other of the pattern-corresponding word or the pattern based on the score given to the one of the pattern-corresponding word or the pattern and the edge score;
Assigning a score to the pattern-corresponding word or the pattern based on the score given to the other by the step of assigning a score to the other and the edge score; and
Repeating the step of assigning a score to the other and the step of assigning a score to the one, and obtaining the pattern-corresponding word satisfying a predetermined score condition;
Equipped with a,
The network stored by the network score storage means is a matrix in which the pattern-corresponding word is associated with the pattern, and the edge score between the pattern-corresponding word and the pattern is stored as an element in the matrix,
The step of assigning a score to the other and the step of assigning a score to the other are determined on the basis of the maximum eigenvalue λ of the matrix for the matrix stored in the network score storage means. A method of calculating a score by using a variable indicating the importance of the score . In a machine learning apparatus that extracts a pattern-corresponding word associated with the extracted pattern, the pattern-corresponding word and the pattern are respectively extracted from a corpus, and a pattern that co-occurs with teacher data for each category given in advance. It is a method executed by a machine learning device that includes a network score storage unit that considers a node and stores an edge score that is a score of an edge connecting the nodes in association with the edge,
Giving a predetermined initial score to one of the pattern-corresponding word or the pattern;
Assigning a score to the other of the pattern-corresponding word or the pattern based on the score given to the one of the pattern-corresponding word or the pattern and the edge score;
Assigning a score to the pattern-corresponding word or the pattern based on the score given to the other by the step of assigning a score to the other and the edge score; and
Repeating the step of assigning a score to the other and the step of assigning a score to the one, and obtaining the pattern-corresponding word satisfying a predetermined score condition;
Equipped with a,
The network stored by the network score storage means is a matrix in which the pattern-corresponding word is associated with the pattern, and the edge score between the pattern-corresponding word and the pattern is stored as an element in the matrix,
The step of assigning a score to the other and the step of assigning a score to the other are determined on the basis of the maximum eigenvalue λ of the matrix for the matrix stored in the network score storage means. A method of calculating a score by using a variable indicating the importance of the score .

Images