JP2008234618A - Knowledge extracting device, knowledge extracting method and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a knowledge extracting method capable of further highly accurately extracting knowledge at a high speed, by using text data and structured data. <P>SOLUTION: This knowledge extracting device has a text dividing part 122 dividing the text data into words, a word pair making part 126 making a word pair by combining the words of the text data, a co-occurrence score calculating part 128 calculating a co-occurrence score of indicating an appearance degree of both composing words composing the word pair in a plurality of text data, a priority determining part 130 determining priority of the word pair based on the co-occurrence score, a text-time series data corresponding part 150 acquiring time series data corresponding to the composing word of the word pair, and a correlation coefficient calculating part 140 calculating a correlation coefficient of the word pair by using the time series data associated with the composing word of the word pair according to the determined priority. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a knowledge extraction device, a knowledge extraction method, and a computer program.

  Data mining, which is an analysis method that extracts relevance and law from a large amount of data, is used to discover effective business patterns. For example, in the area of consumer finance, credit quality of a customer can be measured by performing data mining using a large amount of customer attribute data (for example, income, age, etc.).

  With the recent rise of the Internet and the increase in storage capacity, a large amount of text and structured data such as POS data coexist recently. Conventionally, among coexisting data, text data is analyzed separately using text mining, structured data is analyzed using data mining, and so on. For example, when measuring the creditworthiness of the customer described above, only attribute data, which is structured data, is used, and the content of the customer's remarks is not used.

  However, by analyzing these information in an integrated manner, more accurate knowledge extraction and high-speed knowledge extraction can be expected. For such a situation, a method of analyzing using text data and structured data has been proposed (for example, Patent Documents 1 to 4). For example, when the structured data is time-series data, for example, in a news article, a company's profit and the exchange rate may be spoken in relation to each other. In this case, it is presumed that there is some correlation between the company profit of the company and the time series data of the exchange rate. Therefore, it is possible to infer the correlation between the company profit and the time series data of the exchange rate from the text data of the news article.

JP-A-2005-78240 JP 2000-348015 A JP 2001-134575 A JP 2005-135167 A

  However, the conventional method for analyzing both text data and structured data is to analyze each data separately and acquire knowledge from the analysis result. For this reason, the number of processes for extracting new knowledge increases, and it takes too much processing time to discover correlations between data, particularly time-series data. Therefore, it is not easy to extract new knowledge.

  Therefore, the present invention has been made in view of the above problems, and an object of the present invention is to extract knowledge with high accuracy and high speed using text data and structured data. It is an object of the present invention to provide a new and improved knowledge extraction apparatus and knowledge extraction method.

  In order to solve the above-described problem, according to an aspect of the present invention, a knowledge extraction device is provided that extracts knowledge using text data and time-series data. Such a knowledge extraction device includes a text dividing unit that divides text data into words, a word pair creation unit that creates word pairs by combining words in text data, and constituent words that constitute word pairs in a plurality of text data. A co-occurrence score calculation unit that calculates a co-occurrence score indicating the degree of appearance, a priority determination unit that determines the priority of word pairs based on the co-occurrence score, and time-series data corresponding to the constituent words of the word pair A text / time-series data correspondence unit to be acquired; and a correlation coefficient calculation unit that calculates a correlation coefficient of the word pair using time-series data associated with the constituent words of the word pair according to the determined priority order; It is characterized by providing.

  According to the present invention, in order to find correlated data from a large amount of time-series data, first, a plurality of text data input to the knowledge extraction device is divided into words by a text dividing unit, and the divided words are combined. A co-occurrence score is calculated for each word pair. The co-occurrence score indicates how much one constituent word that constitutes a word pair and another constituent word appear together. In accordance with the priority order determined based on the co-occurrence score, the correlation coefficient of the word pair is calculated from the time series data. Thus, according to the knowledge extraction apparatus of the present invention, by finding a pair of time series data that seems to be highly correlated by text analysis using text data, and calculating correlation coefficients in the order that the correlation is likely to be high. The correlation between time series data can be found with high accuracy and high speed.

  Here, the priority order determination unit can assign priorities in descending order of the absolute value of the co-occurrence score, for example. Here, when the degree of appearance in each text data is linked between the constituent words (that is, when the constituent words appear many together or rarely appear together in each text data), the co-occurrence score is Suppose it is expensive. At this time, it is presumed that the word pair having a higher absolute value of the co-occurrence score is more likely to be correlated, and the correlation coefficient is preferentially calculated from the time series data of the word pair.

  In addition, the knowledge extraction device stores a time-series data storage unit that associates and stores time-series data and a time-series data ID for specifying the time-series data, and stores a word and a time-series data ID in association with each other. A table storage unit. Here, the correlation coefficient calculation unit may not calculate the correlation coefficient of the word pair when the time-series data ID corresponding to the constituent words of the word pair is not stored in the correspondence table storage unit. Further, the correlation coefficient calculation unit may not calculate the correlation coefficient of the word pair when the time series data ID corresponding to the constituent words of one word pair is the same. Thus, before calculating the correlation coefficient, by excluding the case where the correlation coefficient cannot be calculated from the time series data or the case where the correlation coefficient is calculated from the same time series data, the correlation is improved more efficiently. A number can be calculated.

  The correlation coefficient calculation unit sequentially generates a predetermined delay between the time series data, calculates a correlation coefficient for each delay, calculates a plurality of correlation coefficients for one word pair, Of these correlation coefficients, the correlation coefficient having the maximum absolute value may be used as the correlation coefficient of the word pair. In this way, by causing a delay in one time-series data with respect to other time-series data, a word pair in which a strong correlation appears can be found, and knowledge can be extracted with higher accuracy.

  Furthermore, a correlation determination unit that determines whether or not the constituent words of the word pair have a correlation based on the correlation coefficient of the word pair can be provided. Here, the correlation determination unit may determine that there is a correlation when the absolute value of the correlation coefficient is equal to or greater than a predetermined value. Thereby, the knowledge extraction apparatus can select a highly correlated word pair and provide it to the user as new knowledge.

  In order to solve the above-described problem, according to another aspect of the present invention, a computer program that causes a computer to function as the above-described knowledge extraction device is provided. The computer program is stored in a storage device included in the computer, and read and executed by a CPU included in the computer, thereby causing the computer to function as the knowledge extraction device. A computer-readable recording medium in which a computer program is recorded can also be provided. The recording medium is, for example, a magnetic disk or an optical disk.

  Furthermore, in order to solve the above-described problem, according to another aspect of the present invention, a knowledge extraction method for extracting knowledge using text data and time-series data is provided. The knowledge extraction method includes a text dividing step for dividing text data into words, a word pair creating step for creating a word pair by combining the divided words, and a word pair created in a plurality of text data A co-occurrence score calculating step for calculating a co-occurrence score indicating the degree of appearance of the words, a priority order determining step for determining the priority order of the word pair based on the co-occurrence score, Correlation coefficient for calculating the correlation coefficient of a word pair using the time-series data associated with the constituent words of the word pair according to the text / time-series data correspondence step for acquiring the series data and the determined priority order And a calculating step.

  According to the present invention, in order to find correlated data from a large amount of time-series data, first, a plurality of text data input to the knowledge extraction device is divided into words, and a word pair obtained by combining the divided words. A co-occurrence score is calculated for each. The co-occurrence score indicates how much one constituent word that constitutes a word pair and another constituent word appear together. In accordance with the priority order determined based on the co-occurrence score, the correlation coefficient of the word pair is calculated from the time series data. Thus, according to the knowledge extraction method of the present invention, by finding a pair of time series data that seems to be highly correlated by text analysis using text data, and calculating correlation coefficients in the order that the correlation is likely to be high. The correlation between time series data can be found with high accuracy and high speed.

  Here, in the priority order determination step, priorities can be assigned in descending order of the absolute value of the co-occurrence score. As a result, if the co-occurrence score is high when the degree of occurrence in each text data is linked between the constituent words, the word pair having a higher absolute value of the co-occurrence score is more likely to be correlated. The correlation coefficient is calculated from the time series data of the word pair.

  In the correlation coefficient calculating step, the time series data ID for specifying the time series data corresponding to the constituent words of the word pair is stored in the correspondence table storage unit that stores the word and the time series data ID in association with each other. If not, the correlation coefficient of the word pair may not be calculated. Further, in the correlation coefficient calculating step, when the time series data corresponding to the constituent words of one word pair are the same, the correlation coefficient of the word pair may not be calculated. Thus, before calculating the correlation coefficient, by excluding the case where the correlation coefficient cannot be calculated from the time series data or the case where the correlation coefficient is calculated from the same time series data, the correlation is improved more efficiently. A number can be calculated.

  The correlation coefficient calculating step calculates a plurality of correlation coefficients for one word pair by sequentially generating a predetermined delay between time series data and calculating a correlation coefficient for each delay. The step of setting the correlation coefficient having the maximum absolute value among the plurality of calculated correlation coefficients as the correlation coefficient of the word pair may be included. As a result, it is possible to find word pairs in which strong correlation is caused by delaying one time-series data with respect to other time-series data, and it is possible to extract knowledge with higher accuracy.

  The knowledge extraction method of the present invention may further include a correlation determination step of determining whether or not the constituent words of the word pair have a correlation based on the correlation coefficient of the word pair. Thereby, the knowledge extraction apparatus can select a highly correlated word pair and provide it to the user as new knowledge.

  In order to solve the above-mentioned problem, according to another aspect of the present invention, there is provided a knowledge extraction device that extracts knowledge using a plurality of user input models that are text data, time-series data, and mutually correlated information. Provided. Such a knowledge extraction device includes at least two or more text dividing units that divide text data into constituent words, an appearance frequency measuring unit that measures the appearance frequency of constituent words of text data, and an input device for inputting a user input model. Co-occurrence score indicating the degree of appearance of both the user input model and the time-series data for each user input model based on the model input unit to which the user input model is input and the appearance frequency of the constituent words of the text data Correlation score indicating the strength of correlation of each time-series data pair created by combining arbitrary two time-series data based on the calculated co-occurrence score The time series data correlation score calculation unit for calculating the correlation coefficient of the time series data pair based on the correlation score of the time series data pair A priority determining section for determining the priority of output, and the correlation coefficient calculation unit for calculating a correlation coefficient time series data pair when according to the determined priorities, characterized in that it comprises a.

  According to the present invention, first, a plurality of text data input to the knowledge extracting device is divided into constituent words by the text dividing unit, and the appearance frequency of the constituent words in each text data is measured by the appearance frequency measuring unit. On the other hand, in order to find correlated data from a large amount of time-series data, when a user input model that the user considers to be correlated with each other is input to the knowledge extraction device, the frequency of appearance of constituent words of text data measured in advance Based on the above, a co-occurrence score of the user input model and the time series data is calculated. Then, based on the calculated co-occurrence score, a correlation score of each time-series data pair created by combining arbitrary two time-series data is calculated. As described above, according to the knowledge extracting apparatus of the present invention, the text data, the user input model, and the time series data are integrated and analyzed, whereby the correlation between the time series data can be found with high accuracy. In addition, since the correlation coefficient of the time series data pair is calculated according to the priority order after the priority order for calculating the correlation coefficient of the time series data pair is determined using the user input model correlated with each other, The correlation coefficient calculation process can be reduced without calculating the correlation coefficient of the time series data pair, and the correlation between the time series data can be found at high speed.

  Here, the time-series data co-occurrence score calculation unit calculates the appearance frequency of the constituent words of the text data that are the same as the user input model in the text data and the constituent words of the text data that are the same as the words corresponding to the time-series data. A co-occurrence score between the user input model and the time series data can be calculated using the appearance frequency in the text data. Thus, by calculating the co-occurrence score of the user input model and the time series data using the appearance frequency of the constituent words of the text data, the priority order for calculating the correlation coefficient of the time series data pair is determined. In addition to the correlation of the model obtained from the user input model, the relationship between the constituent words obtained from the text data can be reflected.

  In addition, the knowledge extraction device of the present invention relates to a time series data storage unit that stores time series data and a time series data ID for specifying the time series data in association with each other, and associates a word with the time series data ID. A word list composed of one or more words stored in the corresponding notation part corresponding to the same time series data ID for the correspondence table storage unit to be stored and the time series data ID stored in the time series data storage unit And a time-series data word list acquisition unit for acquiring. At this time, the time-series data co-occurrence score calculation unit calculates a co-occurrence score between each word constituting the word list of the time-series data and the user input model, and the absolute value of the calculated co-occurrence scores is the maximum. Can be the co-occurrence score of the time-series data and the model. Thereby, even when there are a plurality of words having the same time-series data ID, a co-occurrence score with the user input model can be determined for each time-series data.

  Furthermore, the time series data correlation score calculation unit creates a user input model pair composed of a first user input model and a second user input model, which are arbitrary two user input models, by combining the user input models, For a time-series data pair consisting of first time-series data and second time-series data, which is a correlation score calculation target, a co-occurrence score between the first user input model and the first time-series data, and second The product of the co-occurrence score of the user input model and the second time series data, the co-occurrence score of the second user input model and the first time series data, the first user input model and the 2 is multiplied by the co-occurrence score with the time series data of 2, and the maximum value among the calculated multiplication values for the user input pairs made up of any two arbitrary user input models is calculated. Or as a correlation score of the time-series data pair is the score calculation target.

  The priority order determination unit can assign priorities in descending order of the correlation score of the time-series data pair. Here, since each user input model is correlated information, it is considered that the user input model and time series data are linked (for example, in text data, both the user input model and time series data are If the co-occurrence score between the user input model and the time series data is high when it appears frequently (or rarely appears together), the user input model and the time series data are likely to be correlated Conceivable. The correlation score indicating the strength of correlation of each time series data pair is calculated by, for example, multiplication of the co-occurrence score, and the larger the value, the stronger the correlation of the time series data pair is considered, and the time series having a higher correlation score. The calculation of the correlation coefficient by the data pair is performed preferentially.

  Further, the correlation coefficient calculation unit sequentially generates a predetermined delay between the time series data, and calculates a correlation coefficient for each delay, thereby obtaining a plurality of correlation coefficients for one time series data pair. The correlation coefficient having the maximum absolute value among the plurality of correlation coefficients can be calculated as the correlation coefficient of the time series data pair. By causing one time-series data to be delayed with respect to other time-series data, a time-series data pair in which strong correlation appears can be found, and knowledge can be extracted with higher accuracy.

  Furthermore, the knowledge extraction apparatus of the present invention further includes a correlation determination unit that determines whether or not there is a correlation between the time series data constituting the time series data pair based on the correlation coefficient of the time series data pair. May be. The correlation determination unit can determine that there is a correlation when the absolute value of the correlation coefficient is equal to or greater than a predetermined value. Thereby, the knowledge extraction apparatus can select a time-series data pair with high correlation and provide it to the user as new knowledge.

  In order to solve the above-described problem, according to another aspect of the present invention, a computer program that causes a computer to function as the above-described knowledge extraction device is provided. The computer program is stored in a storage device included in the computer, and read and executed by a CPU included in the computer, thereby causing the computer to function as the knowledge extraction device. A computer-readable recording medium in which a computer program is recorded can also be provided. The recording medium is, for example, a magnetic disk or an optical disk.

  Furthermore, in order to solve the above-described problem, according to another aspect of the present invention, there is provided a knowledge extraction method for extracting knowledge using a plurality of user input models that are text data, time series data, and mutually correlated information. Provided. The knowledge extraction method includes at least two or more text dividing steps for dividing text data into constituent words, an appearance frequency measuring step for measuring the appearance frequency of constituent words of the text data, and an input device for inputting a user input model. A co-occurrence score indicating the degree of appearance of both the user input model and the time series data for each user input model based on the model input step in which the user input model is input and the appearance frequency of the constituent words of the text data A time series data co-occurrence score calculation step to calculate, and a correlation score indicating the strength of correlation of each time series data pair created by combining any two time series data based on the calculated co-occurrence score Based on the time series data correlation score calculation step to be calculated and the correlation score of the time series data pair, A priority order determining step for determining a priority order for calculating the correlation coefficient of the column data pair, and a correlation coefficient calculating step for calculating the correlation coefficient of the time-series data pair according to the determined priority order. It is characterized by.

  According to the present invention, first, a plurality of input text data is divided into constituent words, and the appearance frequency of the constituent words in each text data is measured. On the other hand, in order to find correlated data from a large amount of time-series data, when a user input model that is considered to be mutually correlated information by a user is input, based on the appearance frequency of constituent words of text data measured in advance, A co-occurrence score of the user input model and time series data is calculated. Then, based on the calculated co-occurrence score, a correlation score of each time-series data pair created by combining arbitrary two time-series data is calculated. As described above, according to the knowledge extraction method of the present invention, the text data, the user input model, and the time series data are integrated and analyzed, whereby the correlation between the time series data can be found with high accuracy. In addition, since the correlation coefficient of the time series data pair is calculated according to the priority order after the priority order for calculating the correlation coefficient of the time series data pair is determined using the user input model correlated with each other, The correlation coefficient calculation process can be reduced without calculating the correlation coefficient of the time series data pair, and the correlation between the time series data can be found at high speed.

  Here, the time series data co-occurrence score calculation step includes the appearance frequency of text data constituent words that are the same as the user input model in the text data, and text data constituent words that are the same as the words corresponding to the time series data. A co-occurrence score between the user input model and the time series data can be calculated using the appearance frequency in the text data. Thus, by calculating the co-occurrence score of the user input model and the time series data using the appearance frequency of the constituent words of the text data, the priority order for calculating the correlation coefficient of the time series data pair is determined. In addition to the correlation of the model obtained from the user input model, the relationship between the constituent words obtained from the text data can be reflected.

  Further, the knowledge extraction method of the present invention relates a word and a time-series data ID with respect to the time-series data ID of the time-series data storage unit stored in association with the time-series data in order to identify the time-series data. It may further include a time-series data word list acquisition step of acquiring one or more words corresponding to the same time-series data ID from the stored corresponding notation parts and creating a word list. At this time, the time series data co-occurrence score calculating step calculates a co-occurrence score between each word constituting the word list of the time series data and the user input model, and the absolute value of the calculated co-occurrence scores is the maximum. Can be the co-occurrence score of the time-series data and the model. Thereby, even when there are a plurality of words having the same time-series data ID, a co-occurrence score with the user input model can be determined for each time-series data.

  The time series data correlation score calculating step combines the user input models to create a user input model pair including a first user input model and a second user input model which are arbitrary two user input models. Co-occurrence of the first user input model and the first time-series data for the pair creation step and the time-series data pair consisting of the first time-series data and the second time-series data, which is a correlation score calculation target A score, a multiplication value of the co-occurrence score of the second user input model and the second time series data, a co-occurrence score of the second user input model and the first time series data, the first A user comprising a multiplication step for calculating a multiplication value of the user input model and the co-occurrence score of the second time-series data, and any two arbitrary user input models assumed A correlation score determining step of the correlation score time series data pairs a maximum value among the calculated multiplication value is a correlation score calculation target for the force pairs may include.

  Further, the priority order determining step can assign priorities in descending order of the correlation score of the time series data pair. The correlation score indicating the correlation strength of each time series data pair is calculated by, for example, multiplication of the co-occurrence score, and it is considered that the correlation between the time series data pairs is stronger as the value is larger. From this, the calculation of the correlation coefficient by the time series data pair having a higher correlation score is preferentially performed.

  Further, the correlation coefficient calculating step sequentially generates a predetermined delay between the time series data, and calculates a correlation coefficient for each delay, thereby obtaining a plurality of correlation coefficients for one time series data pair. And a step of calculating a correlation coefficient having the maximum absolute value among the plurality of calculated correlation coefficients as a correlation coefficient of the time-series data pair. By causing one time-series data to be delayed with respect to other time-series data, a time-series data pair in which strong correlation appears can be found, and knowledge can be extracted with higher accuracy.

  Further, it may further include a correlation determination step for determining whether or not there is a correlation between the time series data constituting the time series data pair based on the correlation coefficient of the time series data pair. Thereby, a time-series data pair with high correlation can be selected and provided to the user as new knowledge.

  As described above, according to the present invention, it is possible to provide a knowledge extraction apparatus and a knowledge extraction method that can extract knowledge with high accuracy and high speed using text data and structured data. .

  Exemplary embodiments of the present invention will be described below in detail with reference to the accompanying drawings. In addition, in this specification and drawing, about the component which has the substantially same function structure, duplication description is abbreviate | omitted by attaching | subjecting the same code | symbol.

(First embodiment)
First, a schematic configuration of the knowledge extraction device according to the first embodiment of the present invention will be described with reference to FIGS. FIG. 1 is a block diagram showing a schematic configuration of the knowledge extracting apparatus according to the present embodiment. FIG. 2 shows an example of a table representing the appearance frequency of words constituting a document that is text data. FIG. 3 shows an example of the configuration of the time-series data storage unit. FIG. 4 shows an example of a configuration of a correspondence table storage unit that stores a correspondence table in which words and time-series data are associated with each other.

<Schematic configuration of knowledge extraction device>
As shown in FIG. 1, the knowledge extraction apparatus 100 according to the present embodiment includes a text input unit 110, a text analysis unit 120, a priority order determination unit 130, a correlation coefficient calculation unit 140, text / time-series data. A correspondence unit 150, a correlation determination unit 160, a time-series data storage unit 170, and a correspondence table storage unit 180 are provided.

  The text input unit 110 is an input unit for inputting text data. As the text data, for example, articles and documents of newspapers and magazines can be used.

  The text analysis unit 120 is a functional unit that analyzes the co-occurrence relationship for words constituting text data. For example, the text division unit 122, the appearance frequency measurement unit 124, the word pair creation unit 126, and the co-occurrence score And a calculation unit 128.

  The text dividing unit 122 is a functional unit that divides text data into words, and can be divided using a technique such as morphological analysis.

  The appearance frequency measurement unit 124 is a functional unit that measures the frequency with which words appear in text data. The appearance frequency measuring unit 124 measures the appearance frequency of the words divided by the text dividing unit 122, and creates a matrix 125 for the word appearance frequency as shown in FIG. The matrix 125 in FIG. 2 shows the number of appearances of the word 125b measured for each text data by the appearance frequency measuring unit 124. The document ID 125a is an ID for specifying text data.

  The word pair creation unit 126 is a functional unit that creates one word pair by combining two words divided by the text division unit 122. The knowledge extraction apparatus 100 according to the present embodiment extracts a word pair having a strong correlation among the word pairs created by the word pair creation unit 126.

  The co-occurrence score calculation unit 128 is a functional unit that calculates a co-occurrence score indicating the co-occurrence relationship of the constituent words of the word pair created by the word pair creation unit 126. Details of the co-occurrence score will be described later.

  The priority order determination unit 130 is a functional unit that determines the order in which correlation coefficients are calculated in the correlation coefficient calculation unit 140 described later. The priority order determination unit 130 of this embodiment is determined based on the co-occurrence score calculated by the co-occurrence score calculation unit 128.

  The correlation coefficient calculation unit 140 is a functional unit that calculates the correlation coefficient of the time series data associated with the constituent words of the word pair. Correlation coefficient calculation section 140 calculates correlation coefficients in descending order of word pairs determined by priority determination section 130. Details of the correlation coefficient calculation method by the correlation coefficient calculation unit 140 will be described later.

  The text / time-series data correspondence unit 150 is a functional unit that acquires time-series data associated with the constituent words of the word pair. The text / time-series data correspondence unit 150 acquires time-series data corresponding to the constituent words of the word pair received from the correlation coefficient calculation unit 140 from the time-series data storage unit 170 and the correspondence table storage unit 180 described later. . Then, the text / time-series data correspondence unit 150 outputs the acquired time-series data to the correlation coefficient calculation unit 140.

  The correlation determination unit 160 is a functional unit that determines whether or not the constituent words of the word pair have a correlation based on the calculated correlation coefficient of the word pair. When the absolute value of the correlation coefficient calculated by the correlation coefficient calculation unit 140 is greater than or equal to a predetermined value, the correlation determination unit 160 of this embodiment determines that the constituent words of this word pair have a high correlation. Correlation determining section 160 outputs word pairs having a strong correlation as new knowledge.

  The time-series data storage unit 170 is a storage unit that stores time-series data stored in association with words, and includes, for example, a memory such as a RAM or a hard disk. As shown in FIG. 3, the time-series data storage unit 170 includes, for example, a time-series data ID 171 that is an ID for specifying time-series data and time-series data 172. The time series data 172 is a collection of data that changes with time, and stores data acquired at predetermined time intervals. Here, the time-series data 172 stored in the time-series data storage unit 170 is not limited to those acquired at the same time interval, and the acquisition start time and the acquisition end time may be different. Conceivable. However, in order to calculate the correlation coefficient from the time series data 172, it is necessary to use data acquired at the same timing. For this reason, the data acquisition time interval, the acquisition start time, the acquisition end time, and the like may be stored in the time-series data storage unit 170 in order to ensure data consistency.

  The correspondence table storage unit 180 is a storage unit that stores a correspondence relationship between words and time-series data, and includes, for example, a memory such as a RAM or a hard disk. As shown in FIG. 4, the correspondence table storage unit 180 includes, for example, a word 181 and a time series data ID 182. The word 181 stores a word having time series data. The time series data ID 182 is the same as the time series data ID 171 in FIG. 3 and is used to specify time series data corresponding to the word 181. Even when different words 181 have the same time-series data, that is, the same time-series data ID 182 is stored in the semantically identical words. In FIG. 4, the time-series data IDs corresponding to the words “exchange rate” and “exchange rate” are both “D1”.

  The schematic configuration of the knowledge extraction device 100 according to the present embodiment has been described above. Next, the knowledge extraction method according to the present embodiment will be described based on FIGS. 5 and 6. FIG. 5 is a flowchart showing the knowledge extraction method according to the present embodiment. FIG. 6 is an explanatory diagram for explaining the nature of the co-occurrence score.

<Knowledge extraction method>
In the knowledge extraction method according to the present embodiment, as shown in FIG. 5, first, text data is input to the text input unit 110 (S101). In this embodiment, it is assumed that a large amount of documents are used as text data. As the document, for example, an article or a book related to the time series data stored in the time series data storage unit 170 is used.

  Next, the text input unit 110 transmits the input text data to the text analysis unit 120. In the text analysis unit 120, first, the text division unit 122 divides the text data into words (S103). The text dividing unit 122 decomposes the sentence into words for each text data by, for example, morphological analysis. Then, how often the decomposed word appears in the text data is measured (S105). The appearance frequency of the word is expressed by measuring the number of times the word appears for each text data. Thereby, for example, a matrix 125 as shown in FIG. 2 is formed.

  Thereafter, a word pair is created by combining two of the words 125a appearing in the text data (S107). As many word pairs as possible are created for the word 125a appearing in the text data, for example. Then, based on the appearance frequency of the words shown in the matrix 125, a co-occurrence score between the constituent words constituting the word pair is calculated (S109). The co-occurrence score indicates the degree of occurrence of both words constituting the word pair in the text data, and is calculated from the following Equation 1 as the cosine distance of the N-dimensional vector in this embodiment.

Here, x _i and y _i are the number of times the word X and the word Y appear in the document with the document ID = i, respectively, and N is the total number of text data.

  The co-occurrence score takes a value of −1 ≦ (co-occurrence score) ≦ 1, and when the co-occurrence of the constituent words of the word pair is strong, the absolute value of the co-occurrence score increases. As described above, the co-occurrence score indicates how much a certain word tends to appear with other words. Here, FIG. 6 shows a graph in which the horizontal axis represents the document ID for specifying the text data, and the vertical axis represents the number of occurrences of the word. For example, in FIG. 6, the word X and the word Y show different waveforms. That is, when the appearance frequency of the word X is high, the appearance frequency of the word Y is high or low, and there is some relationship between the appearance frequency of the word X and the appearance frequency of the word Y in various documents. Is not seen. In this case, it can be said that the word X and the word Y do not co-occur. At this time, the co-occurrence score is close to zero.

  On the other hand, the word X and the word Z show almost the same waveform. This is because the number of times that the word X and the word Z appear both increases or decreases, and it can be said that the word X and the word Z co-occur. At this time, the co-occurrence score is close to 1. On the other hand, on the other hand, when the number of occurrences of word X in a document is high, the number of appearances of word W is low, and when the number of appearances of word X in another document is low, the number of appearances of word W is high. There may be a relationship such as At this time, there is a negative co-occurrence relationship between the words X and W, and the co-occurrence score is close to -1. Also in this case, it is considered that there is some relationship between the word X and the word W. Therefore, the magnitude of the absolute value of the co-occurrence score is important and will be used in the following processing.

  When the co-occurrence score of each word pair is calculated in step S109, the priority order for calculating the correlation coefficient for each word pair is determined (S111). In step S111, for example, it is assumed that priorities are assigned in descending order of the absolute value of the co-occurrence score. Here, a word pair list in which word pairs are arranged in descending order of priority may be created, and the following steps S113 to S121 may be performed in the order of description in the word pair list.

  Next, the correlation coefficients of word pairs are calculated in descending order of priority according to the priority order of step S111, but the following processes of steps S113 and S115 may be performed before calculating the correlation coefficient.

  First, since the correlation coefficient is calculated based on the time series data, it is checked whether or not the time series data of the constituent words of the word pair exists (S113). In step S113, when the text / time-series data correspondence unit 150 tries to acquire the time-series data ID 182 corresponding to the constituent word from the correspondence table storage unit 180, the constituent word is not registered in the word 181 or the constituent word Is registered in the word 181 but the time series data ID 182 is blank (no corresponding time series data), it is determined that there is no time series data, and the correlation coefficient of this word pair is not calculated. The process proceeds to step S123.

  Next, it is checked whether the time series data IDs of the two constituent words of the word pair are different (S115). The fact that the constituent words have the same time series data ID means that the constituent words are semantically substantially the same and the time series data are also the same. For example, in FIG. 4, the “exchange rate” and the “exchange rate” are associated with the time series data ID “D1”. Since there is no need to find a correlation for such a word pair, the process proceeds to step S123 without calculating a correlation coefficient.

  After performing such a check, a correlation coefficient is calculated for a word pair that satisfies the requirements of steps S113 and S115 (S117). For example, assume that the correlation coefficient between the word “exchange rate” and “Company X revenue” is calculated. First, the text / time-series data correspondence unit 150 acquires time-series data corresponding to each word. From the correspondence table storage unit 180, it can be seen that the time series ID of the time series data corresponding to the “exchange rate” is “D1”, and the time series ID of the time series data corresponding to “company X revenue” is “D2”. Therefore, the text / time-series data handling unit 150 acquires the time-series data of the time-series data IDs “D1” and “D2” from the time-series data storage unit 170 and transmits the time-series data to the correlation coefficient calculation unit 140. In the following, it is assumed that the data acquisition time interval, the acquisition start time, etc. of the time series data corresponding to the time series data IDs “D1” and “D2” are the same.

  The correlation coefficient calculation unit 140 that has received the time series data calculates a correlation coefficient from the following Equation 2.

Here, x bar is an average value of the time series data of the word X, and σ _x is a standard deviation of the time series data of the word X. Further, t is an index of time series data, and n is a sequence length of time series data.

  When the correlation coefficient is calculated from Equation 2, it is checked whether the absolute value of the correlation coefficient is equal to or greater than a predetermined value (S119). The correlation coefficient takes a value of −1 ≦ (correlation coefficient) ≦ 1, and the larger the absolute value, the stronger the correlation. Therefore, a word pair having a correlation coefficient whose absolute value is larger than a predetermined value is output as new knowledge (S121). On the other hand, when the absolute value of the correlation coefficient is smaller than the predetermined value, the new knowledge is not output and the process proceeds to step S123. Here, the predetermined value can be arbitrarily determined, but can be set to a value considered to have a strong correlation, for example, 0.7.

  When the correlation for one word pair is obtained, it is confirmed whether the process for obtaining the correlation for all word pairs for which the correlation coefficient is to be calculated has been completed (S123). The word pair for which the correlation coefficient is to be calculated may be, for example, an all-word pair, or a word pair having a high priority to a half word pair. The number of word pairs for calculating the correlation coefficient can be determined in consideration of the time that can be allocated to the knowledge extraction process, for example. Here, in the knowledge extraction apparatus 100 of the present embodiment, word pairs that are likely to be highly correlated are selected with priority using text data. Therefore, even when the number of word pairs for calculating the correlation coefficient is limited, there is a high possibility of finding a word pair having a high correlation.

  If it is determined in step S123 that the process for obtaining the correlation has been completed, the entire process is terminated. On the other hand, if there are still word pairs whose correlation coefficients are to be calculated, the process proceeds to step S125, and the next highest priority word pair is read from, for example, the word pair list (S125), and the process is repeated from step S113. .

  The knowledge extraction method according to the first embodiment of the present invention has been described above. According to the present embodiment, when obtaining a correlation from a large amount of time-series data, a text analysis is performed in advance from text data, and a pair of time-series data that seems to have a high correlation is determined with priority. Then, the correlation is obtained by calculating the correlation coefficient of the time series data in descending order of priority. By integrating and analyzing text data and time-series data, more accurate knowledge extraction can be expected, and correlation coefficients are calculated for all pairs of time-series data blindly. Compared to this, it is possible to find a correlation between time series data at high speed and extract new knowledge.

  Note that when calculating the correlation coefficient in step S117, it is possible to consider the delay of the time series data. When the delay L is taken into consideration, the correlation coefficient can be calculated from Equation 3 below.

  The maximum value of the delay is predetermined as MAX, the delay L is changed from 0 to MAX, and the correlation coefficient σ (L) is calculated from Equation 3 for each value. For example, in the correlation between “Company X revenue” and “Exchange rate”, the time series data is delayed to 0 weeks (no delay), 1 week, 2 weeks, etc. until the 10th week (maximum delay MAX). The correlation coefficient σ (L) is calculated respectively. Then, a value having the maximum absolute value among the calculated correlation coefficients σ (L) is determined as the correlation coefficient of the word pair. In this way, by taking into account the delay of the time series data, it is possible to find word pairs that show a strong correlation when the delay occurs, and to extract highly correlated knowledge with higher accuracy. Become.

(Second Embodiment)
Next, a knowledge extraction apparatus according to the second embodiment of the present invention will be described. Compared with the first embodiment, the knowledge extraction apparatus according to the present embodiment determines the priority for calculating the correlation coefficient using the user input model input from the user in addition to the text data. Is different. First, a schematic configuration of the knowledge extraction device 200 according to the present embodiment will be described based on FIG. FIG. 7 is a block diagram showing a schematic configuration of the knowledge extracting apparatus according to the present embodiment.

<Schematic configuration of knowledge extraction device>
As shown in FIG. 7, the knowledge extraction apparatus 200 according to the present embodiment includes a text input unit 211, a model input unit 213, a text analysis unit 220, a text / time-series data correspondence unit 231, and a time-series data word. List acquisition unit 233, time series data co-occurrence score calculation unit 240, time series data correlation score calculation unit 250, priority order determination unit 260, correlation coefficient calculation unit 270, correlation determination unit 280, text data A storage unit 291, a time series data storage unit 293, and a correspondence table storage unit 295 are provided.

  The text input unit 211 is an input unit to which text data is input from a text data storage unit 291 described later. As the text data, for example, articles and documents of newspapers and magazines can be used.

  The model input unit 213 is a functional unit that receives a user input model from a user and acquires a list of user input models. The user input model is information used to limit the data for which the co-occurrence score is calculated and to determine the priority for calculating the correlation coefficient of the time-series data. In this embodiment, the user input model is related to each other. It is a word that has. The user inputs a word using an external input device 205 such as a keyboard. The model input unit 213 acquires words (hereinafter referred to as “user input words”) input by the user using the input device 205 to create a word list, and a time-series data co-occurrence score calculation unit 240 described later. Output to.

  The text analysis unit 220 is a functional unit that analyzes the appearance frequency of words constituting the text data, and includes, for example, a text division unit 221 and an appearance frequency measurement unit 223. The text dividing unit 221 is a functional unit that divides text data into words, and can be divided using a technique such as morphological analysis. The appearance frequency measurement unit 223 is a functional unit that measures the frequency at which words appear in text data. The appearance frequency measuring unit 223 measures the appearance frequency for the words divided by the text dividing unit 221 and, for example, the word appearance frequency as shown in FIG. 2 as in the appearance frequency measuring unit 124 of the first embodiment. The matrix 125 is created.

  The text / time-series data correspondence unit 231 is a functional unit that acquires a correspondence relationship between time-series data and words. The text / time-series data correspondence unit 231 refers to a time-series data storage unit 293 and a correspondence table storage unit 295, which will be described later, in response to a request from the time-series data word list acquisition unit 233, which will be described later. The correspondence with the list is acquired and output to the time-series data word list acquisition unit 233. Also, the text / time-series data correspondence unit 231 outputs desired time-series data to the correlation coefficient calculation unit 270.

  The time series data word list acquisition unit 233 is a functional unit that acquires a word list of time series data. The time-series data word list acquisition unit 233 requests the text / time-series data correspondence unit 231 to obtain the correspondence between the time-series data and the words, and the time-series data obtained from the text / time-series data correspondence unit 231. A word list for time-series data is created from the correspondence between the word and the word, and is output to a time-series data co-occurrence score calculation unit 240 described later.

  The time-series data co-occurrence score calculation unit 240 calculates a co-occurrence relationship between the word list created by the time-series data word acquisition unit 233 and the word list including user input words input from the model input unit 213. Part. Details of the calculation of the co-occurrence score will be described later. The time series data co-occurrence score calculation unit 240 outputs the calculated co-occurrence score to the time series data correlation score calculation unit 250 described later.

  The time-series data correlation score calculation unit 250 is a functional unit that calculates a correlation score that is a priority when calculating a correlation coefficient for a pair of time-series data. A method for calculating the correlation score will be described later. The time-series data correlation score calculation unit 250 outputs the calculated correlation score to the priority order determination unit 260 described later.

  The priority order determination unit 260 is a functional unit that determines the order in which correlation coefficients are calculated in a correlation coefficient calculation unit 270 described later. The priority order determination unit 260 of this embodiment is determined based on the correlation score calculated by the time series data correlation score calculation unit 250.

  The correlation coefficient calculation unit 270 is a functional unit that calculates the correlation coefficient of the time series data pair. Correlation coefficient calculation section 270 calculates correlation coefficients in order from the time-series data pair having the highest priority determined by priority determination section 260.

  The correlation determination unit 280 is a functional unit that determines whether or not the time series data pair has a correlation based on the correlation coefficient of the time series data pair calculated by the correlation coefficient calculation unit 270. The correlation determination unit 280 of the present embodiment determines that the correlation of the time series data pair is high when the absolute value of the correlation coefficient calculated by the correlation coefficient calculation unit 270 is equal to or greater than a predetermined value. Correlation determining section 280 outputs time series data pairs having strong correlation as new knowledge.

  The text data storage unit 291 is a storage unit that stores text data such as articles and documents of newspapers and magazines, and includes a memory such as a RAM and a hard disk. In the present embodiment, the text data is stored in the text data storage unit 291 in advance, but the text data may be input from the outside like a user input model.

  The time-series data storage unit 293 is a storage unit that stores time-series data stored in association with words. The time series data storage unit 293 can have the same configuration as the time series data storage unit 170 of the first embodiment as shown in FIG.

  The correspondence table storage unit 295 is a storage unit that stores a correspondence relationship between words and time-series data. The correspondence table storage unit 295 can be configured in the same way as the correspondence notation unit 180 of the first embodiment as shown in FIG.

  The schematic configuration of the knowledge extraction apparatus 200 according to the present embodiment has been described above. Next, the knowledge extraction method according to the present embodiment will be described with reference to FIGS. FIG. 8 is a flowchart showing the knowledge extraction method according to the present embodiment. FIG. 9 is an explanatory diagram of words input to the model input unit 213. FIG. 10 is an explanatory diagram for processing of the time-series data correlation score calculation unit 250.

<Knowledge extraction method>
The knowledge extraction method according to the present embodiment analyzes text data stored in the text data storage unit 291 as preprocessing (S201 to S205). First, as shown in FIG. 8, the text data stored in the text data storage unit 291 is read by the text input unit 211 (S201). In this embodiment, a large amount of documents are used as text data. As the document, for example, an article or a book related to the time series data stored in the time series data storage unit 293 is used.

  Next, the text input unit 211 transmits the input text data to the text analysis unit 220. In the text analysis unit 220, first, the text division unit 221 divides the text data into words (S203). The text dividing unit 221 decomposes a sentence into words for each text data, for example, by morphological analysis. Then, how often the decomposed word appears in the text data is measured (S205). The appearance frequency of the word is expressed by measuring the number of times the word appears for each text data. Thereby, for example, a matrix 125 as shown in FIG. 2 is formed.

  The text data preprocessing (S201 to S205) is performed before calculating the co-occurrence score of the time-series data. Next, based on such pre-processing, the correlation of time series data is discovered and new knowledge is extracted (S211 to S229). First, a model is input by the user (S211). As described above, the model is information used to determine the priority for calculating the correlation coefficient of the time-series data while limiting the data for which the co-occurrence score is to be calculated. Relevant words. The user inputs mutually related words to the model input unit 213 using the input device 205.

  For example, as shown in FIG. 9, “interest rate” (U1) is related to “exchange” (U2), “economy” (U3), “inflation” (U4), and “inflation” (U4) is “unemployment”. It is assumed that the user thinks that there is a relation with “rate” (U5) and “raw material price” (U6). At this time, the user inputs certain concepts and words one after another without describing the detailed relationship between the words. That is, “exchange” (U2), “economy” (U3), “inflation” (U4) are input as words related to “interest rate” (U1), and “unemployment” is related to “inflation” (U4). An interrelated word list can be created in the process of inputting “rate” (U5) and “raw material price” (U6). In this way, a word list U (U1, U2,..., Um) composed of user input words is obtained.

  Here, it is necessary that at least two user input words are input. In addition, since the user input words are treated as having relevance in the knowledge extraction method of the present embodiment, the number of words (for example, several) is smaller than the number of weakly related words (for example, several hundred words). Ten words) It is possible to determine an appropriate priority by inputting a word that seems to have a strong correlation.

  Next, for each of the time series data stored in the time series data storage unit 295, the corresponding word list W (W1, W2,..., By the text / time series data correspondence unit 231 and the time series data word acquisition unit 233 is stored. Wn) is acquired (S213). For example, for the time-series data D1, the text / time-series data correspondence unit 231 refers to the correspondence table storage unit 295 to associate the “exchange rate” (W1) with the “exchange rate” (W2). The series data word acquisition unit 233 acquires a word list W composed of “exchange rate” (W1) and “exchange rate” (W2).

  Further, a co-occurrence score is obtained for the pair of the word of the time series data and the user input word using the word list of each time series data acquired in step S213 and the user input word (S215). That is, in step S215, the degree of relationship between the time-series data and the user input word is obtained. Specifically, first, the frequency of occurrence of words corresponding to time series data and user input words in the text data is obtained from the matrix 125 shown in FIG. 2, and the words Wi (i) constituting the word list of the time series data are obtained. = 1, 2,..., N) and a co-occurrence score between user input words Uj (j = 1, 2,..., M). The calculation of the co-occurrence score can be performed using Formula 1 as in Step S109 in the first embodiment. When the co-occurrence score is calculated according to Equation 1, the co-occurrence score is 0 if either the word Wi of the time series data or the user input word Uj is not registered in the matrix 125. In this way, a list of co-occurrence scores with the word list of the time-series data can be obtained for each user input word.

For example, for the user input word U1, the co-occurrence score with the word Wi in the time series data
Score (W1, U1), Score (W2, U1), ..., Score (Wn, U1)
Get. Of the co-occurrence score list composed of these co-occurrence scores, the co-occurrence score having the maximum absolute value is set as a score Score (Dx, Uy) for the user input word Uy of the time series data Dx. For example, the score for the user input word U1 in the time series data D1 is represented as Score (D1, U1), and the value of Score (D1, U1) is 0 in the co-occurrence score list 241 for the user input word U1 in FIG. .7. Thus, the co-occurrence score of each time series data D1, D2,... For each user input word.
Score (D1, U1), Score (D2, U1), ..., Score (D1, U1)
Score (D1, U2), Score (D2, U2), ..., Score (D1, U2)
Score (D1, U3), Score (D2, U3), ..., Score (D1, U3)
...
Get.

  Here, in FIG. 10, the co-occurrence score list 241 for the user input word U1, the co-occurrence score list 242 for the user input word U2, and the co-occurrence score list 243 for the user input word U3 have a large co-occurrence score. They are shown sorted in order. Note that the co-occurrence score list obtained in step S215 may be stored in a storage unit (not shown). The score shown in the co-occurrence score list can be considered as a measure indicating how much each time-series data is likely to be related to each user input word.

  Thereafter, a correlation score is calculated for each pair (Di, Dj) of time series data D1, D2,..., Dl (S217). In the following, processing is performed on the time series data D1, D2,..., D1, but all the time series data included in the time series data storage unit 295 may be targeted as this time series data. Thus, the calculation amount of the subsequent processing may be saved by limiting to a part thereof.

That is, as described above, the co-occurrence score of each time series data D1, D2,... For each user input word.
Score (D1, U1), Score (D2, U1), ..., Score (D1, U1)
Score (D1, U2), Score (D2, U2), ..., Score (D1, U2)
Score (D1, U3), Score (D2, U3), ..., Score (D1, U3)
...
For example, for the time series data D1, the co-occurrence score for each user input word
Score (D1, U1), Score (D1, U2), Score (D1, U3), ...
Exists. Here, for example, when the maximum absolute value of these scores is smaller than a certain threshold value, the time series data D1 may be excluded from the set of time series data to be processed later. The reason is that the time-series data D1 is considered to be poorly related to any user input word. Similarly, for the other time series data D2, D3,..., The maximum absolute value of the score is compared with the threshold value, and it is determined whether or not to include in the time series data to be processed thereafter. To do.

Here, the threshold value can be arbitrarily set, but may be a predetermined value based on experience, and the co-occurrence score of each time series data D1, D2,...
Score (D1, U1), Score (D2, U1), ..., Score (D1, U1)
Score (D1, U2), Score (D2, U2), ..., Score (D1, U2)
Score (D1, U3), Score (D2, U3), ..., Score (D1, U3)
...
An average value Aver of the absolute values may be obtained, and the value Aver or a value based on the average Aver may be set as a threshold value.

In step S217, first, a word pair (Uk, Ul) is created by combining two of the user input words. Next, the correlation scores of the word pairs with the time series data Di and Dj of the user input word Uk are respectively Score (Di, Uk) and Score (Dj, Uk), and the time series data Di, Dj of the user input word Ul are The correlation scores of Score (Di, Ul) and Score (Dj, Ul) are
Score (Di, Uk) x Score (Dj, Ul)
Score (Dj, Uk) x Score (Di, Ul)
Is calculated. Of these, the larger value is used as the correlation score of the time-series data Di and Dj for the word pair (Uk, Ul) of the user input word. Similarly, correlation scores are calculated for word pairs of all possible user input words, and the maximum value of the calculated correlation scores is defined as the correlation score Score (Di, Dj) of (Di, Dj). . The correlation score Score (Di, Dj) can be considered as an estimate from the co-occurrence relationship based on user input words and text data to what extent the pair of time-series data is likely to be related.

For example, as shown in FIG. 10, it is assumed that three user input words U1, U2, and U3 are input from the user. When the correlation score between the time series data D3 and D5 is calculated in the word pair of the user input words U1 and U2,
Score (D3, U1) × Score (D5, U2) = 0.9 × 0.7 = 0.63
Score (D5, U1) × Score (D3, U2) = 0.5 × 0.6 = 0.30
Thus, the correlation score of D3 and D5 for the word pair (U1, U2) is 0.63. Similarly, the correlation score of the time series data D3 and D5 in the word pair of the user input words U2 and U3 is 0.35, and the correlation score of the time series data D3 and D5 in the word pair of the user input words U3 and U1 is 0. .27. Accordingly, the correlation score Score (Di, Dj) is determined to be 0.63 which is the maximum value among the calculated correlation scores.

  Next, based on the correlation score of the time series data pair calculated in step S217, the priority order for calculating the correlation coefficient for each pair of time series data is determined (S219). The priority is set in descending order of the correlation score calculated in step S217. Here, after creating a time-series data pair list in which time-series data is arranged in order of priority, the following processing (S221 to S229) may be performed in the order of description of the pair list.

  Further, based on the priority determined in step S219, correlation coefficients of time-series data pairs are calculated in descending order of priority (S221). The correlation coefficient calculation unit 270 that has received the time series data can calculate the correlation coefficient from Equation 2 described above, as in the first embodiment. A detailed description of the calculation of the correlation coefficient is omitted because it is the same as that of the first embodiment.

  When the correlation coefficient calculation unit 270 calculates the correlation coefficient, the correlation determination unit 280 checks whether the absolute value of the correlation coefficient is greater than or equal to a predetermined value (S223). The correlation coefficient takes a value of −1 ≦ (correlation coefficient) ≦ 1, and the larger the absolute value, the stronger the correlation. Therefore, a pair of time series data having a correlation coefficient whose absolute value is larger than a predetermined value is output as new knowledge (S225). On the other hand, if the absolute value of the correlation coefficient is smaller than the predetermined value, the new knowledge is not output and the process proceeds to step S227. Here, the predetermined value can be arbitrarily determined, but can be set to a value considered to have a strong correlation, for example, 0.7.

  When the correlation for one time-series data pair has been obtained, it is confirmed whether the processing for obtaining the correlation has been completed for all pairs of time-series data for which correlation coefficients are to be calculated (S227). The pair of time-series data for which the correlation coefficient is to be calculated may be, for example, a pair of all time-series data, or a pair of time-series data in the upper half having a high priority. The number of pairs of time series data for calculating the correlation coefficient can be determined in consideration of, for example, a time that can be allocated to the knowledge extraction process. Here, in the knowledge extraction device 200 of the present embodiment, a pair of time series data that seems to have a high correlation is selected with a priority using user input words and text data. Therefore, even when the number of pairs of time series data for calculating the correlation coefficient is limited, there is a high possibility of finding a pair of time series data having a high correlation.

  If it is determined in step S227 that the process for obtaining the correlation has been completed, the entire process is terminated. On the other hand, if there are still time-series data pairs for which correlation coefficients are to be calculated, the process proceeds to step S229, and the next highest-priority time-series data pair is read from, for example, the time-series data pair list ( S229), the process is repeated from step S221.

  Note that when calculating the correlation coefficient in step S221, the delay of the time-series data can be taken into consideration as in the first embodiment. When the delay L is taken into account, the correlation coefficient can be calculated from Equation 3 described above. By taking into account the delay of time-series data, it is possible to find a pair of time-series data in which a strong correlation appears when a delay occurs, and it is possible to extract highly correlated knowledge with higher accuracy .

  The knowledge extraction method according to the second embodiment of the present invention has been described above. According to the present embodiment, by inputting words that the user thinks are related to each other, a pair of time-series data that is considered to have a high correlation is determined with priority. Then, the correlation is obtained by calculating the correlation coefficient of the time series data in descending order of priority. Thereby, since text data, a user input word, and time series data are integrated and analyzed, extraction of knowledge with higher accuracy can be expected. In the present embodiment, the user input word greatly affects the determination of the priority order for calculating the correlation coefficient of the time-series data pair, and the text data is used as an auxiliary to the user input word. The user can limit the data for which the co-occurrence score is calculated with a simple input, and is faster than if the correlation coefficient was calculated for every pair of time series data blindly. There is an effect that new knowledge can be extracted by finding the correlation of time series data.

  As mentioned above, although preferred embodiment of this invention was described referring an accompanying drawing, it cannot be overemphasized that this invention is not limited to the example which concerns. It will be apparent to those skilled in the art that various changes and modifications can be made within the scope of the claims, and these are naturally within the technical scope of the present invention. Understood.

  For example, in the first embodiment, the presence / absence of time-series data and the difference between the time-series data IDs are checked as a check before calculating the correlation coefficient, but the present invention is not limited to such an example. For example, the correlation coefficient may be calculated only when the absolute value of the co-occurrence score of the word constituting the word pair is equal to or greater than a predetermined threshold. Here, the predetermined threshold value can be an arbitrary value, and for example, an average value of absolute values of all co-occurrence scores calculated in step S109 can be set as the threshold value. Similarly, also in the second embodiment, as a check before calculating the correlation coefficient of the time series data pair, only when the absolute value of the correlation score of the time series data pair is equal to or greater than a predetermined threshold value. The number of relationships may be calculated. Here, the predetermined threshold value can be an arbitrary value. For example, the average value of the correlation scores of the time series data pairs calculated in step S217 can be used as the threshold value.

  In the above embodiment, the time acquisition data intervals of the time series data corresponding to the time series data IDs “D1” and “D2” have been described as being the same. However, the present invention is not limited to this example. Different time-series data such as data acquisition time intervals may be stored in the time-series data storage units 170 and 293. However, in order to calculate the correlation coefficient, it is necessary to match the data acquisition time intervals to ensure consistency. For this reason, for example, data is created by matching time from time series data with a short data acquisition time interval to time series data with a long data acquisition time interval, or by interpolating time series data with a long data acquisition time interval. It is necessary to match the time series data with a short acquisition time interval.

  Further, in the above embodiment, the time series data is stored in advance in the time series data storage units 170 and 293. However, the present invention is not limited to such an example, and a time series data input unit is provided in the same manner as text data. You may make it input from the outside.

It is a block diagram which shows schematic structure of the knowledge extraction apparatus concerning the 1st Embodiment of this invention. An example of the table | surface showing the appearance frequency of the word which comprises the document concerning the embodiment is shown. An example of a structure of the time series data storage part concerning the embodiment is shown. An example of the structure of the correspondence table memory | storage part which memorize | stores the correspondence table which matched the word and time series data concerning the embodiment is shown. It is a flowchart which shows the knowledge extraction method concerning the embodiment. It is explanatory drawing for demonstrating the property of a co-occurrence score. It is a block diagram which shows schematic structure of the knowledge extraction apparatus concerning the 2nd Embodiment of this invention. It is a flowchart which shows the knowledge extraction method concerning the embodiment. It is explanatory drawing about the word input into the model input part concerning the embodiment. It is explanatory drawing about the process of the time series data correlation score calculation part concerning the embodiment.

Explanation of symbols

100, 200 Knowledge extraction device 110, 211 Text input unit 120, 220 Text analysis unit 122, 221 Text division unit 124, 223 Appearance frequency measurement unit 126 Word pair creation unit 128 Co-occurrence score calculation unit 130, 260 Priority order determination unit 140 270 Correlation coefficient calculation unit 150, 231 Text / time series data correspondence unit 160, 280 Correlation determination unit 170, 293 Time series data storage unit 180, 295 Correspondence table storage unit 205 Input device 213 Model input unit 233 Time series data word Acquisition unit 240 Time series data co-occurrence score calculation unit 250 Time series data correlation score calculation unit 291 Text data storage unit

Claims (28)

A knowledge extraction device that extracts knowledge using text data and time series data,
A text division unit for dividing text data into words;
A word pair creation unit that creates a word pair by combining the words of the text data;
A co-occurrence score calculating unit that calculates a co-occurrence score indicating a degree of appearance of constituent words constituting the word pair in a plurality of the text data;
A priority determining unit that determines the priority of word pairs based on the co-occurrence score;
A text / time-series data corresponding unit for acquiring time-series data corresponding to the constituent words of the word pair;
A correlation coefficient calculating unit that calculates the correlation coefficient of the word pair using time-series data associated with the constituent words of the word pair according to the determined priority;
A knowledge extraction device comprising:   The knowledge extraction device according to claim 1, wherein the priority order determination unit assigns priorities in descending order of the absolute value of the co-occurrence score. A time-series data storage unit that stores time-series data and a time-series data ID for specifying the time-series data in association with each other;
A correspondence table storage unit that associates and stores words and time-series data IDs;
With
The correlation coefficient calculation unit does not calculate a correlation coefficient of the word pair when time series data ID corresponding to the constituent words of the word pair is not stored in the correspondence table storage unit. The knowledge extraction apparatus according to 1 or 2.   The correlation coefficient calculation unit does not calculate the correlation coefficient of the word pair when the time-series data ID corresponding to the constituent words of one word pair is the same. The knowledge extraction device described in Crab. The correlation coefficient calculation unit
A plurality of correlation coefficients are calculated for one word pair by sequentially generating a predetermined delay between the time series data and calculating a correlation coefficient for each delay,
The knowledge extraction device according to any one of claims 1 to 4, wherein the correlation coefficient having the maximum absolute value among the plurality of correlation coefficients is used as the correlation coefficient of the word pair. A correlation determining unit that determines whether or not the constituent words of the word pair have a correlation based on the correlation coefficient of the word pair;
The knowledge extraction apparatus according to claim 1, wherein the correlation determination unit determines that there is a correlation when an absolute value of a correlation coefficient is equal to or greater than a predetermined value. A computer program that causes a computer to function as a knowledge extraction device that extracts knowledge using text data and time-series data,
Text dividing means for dividing the text data into words;
Word pair creation means for creating a word pair by combining the words of the text data;
Co-occurrence score calculating means for calculating a co-occurrence score indicating the degree of appearance of constituent words constituting the word pair in a plurality of the text data;
Priority determining means for determining the priority of word pairs based on the co-occurrence score;
Text / time-series data correspondence means for obtaining time-series data corresponding to the constituent words of the word pair;
A computer that functions as a correlation coefficient calculation unit that calculates a correlation coefficient of the word pair using time-series data associated with the constituent words of the word pair according to the determined priority order. program. A knowledge extraction method for extracting knowledge using text data and time series data,
A text splitting step for splitting the text data into words;
A word pair creation step of creating a word pair by combining the divided words;
A co-occurrence score calculating step for calculating a co-occurrence score indicating a degree of appearance of constituent words constituting the created word pair in a plurality of the text data; and
A priority determining step for determining the priority of word pairs based on the co-occurrence score;
Text / time-series data correspondence step for obtaining time-series data corresponding to the constituent words of the word pair;
A correlation coefficient calculating step of calculating a correlation coefficient of the word pair using time-series data associated with the constituent words of the word pair according to the determined priority;
A knowledge extracting method characterized by comprising:   9. The knowledge extracting method according to claim 8, wherein the priority order determining step assigns priorities in descending order of the absolute value of the co-occurrence score.   In the correlation coefficient calculating step, a time series data ID for specifying time series data corresponding to the constituent words of the word pair is stored in a correspondence table storage unit that stores the word and the time series data ID in association with each other. The knowledge extraction method according to claim 8 or 9, wherein if there is no word, the correlation coefficient of the word pair is not calculated.   The correlation coefficient of the word pair is not calculated when the time-series data corresponding to the constituent words of one word pair is the same in the correlation coefficient calculating step. The knowledge extraction method described in Crab. The correlation coefficient calculating step includes:
Calculating a plurality of correlation coefficients for one word pair by sequentially generating a predetermined delay between the time series data and calculating a correlation coefficient for each delay;
Of the plurality of calculated correlation coefficients, the correlation coefficient having the maximum absolute value is set as the correlation coefficient of the word pair;
The knowledge extraction method according to claim 8, wherein the knowledge extraction method comprises:   The correlation determination step of determining whether or not the constituent words of the word pair have a correlation based on the correlation coefficient of the word pair is further included. Knowledge extraction method. A knowledge extraction device that extracts knowledge using a plurality of user input models that are text data, time series data, and mutually correlated information,
A text division unit for dividing text data into constituent words;
An appearance frequency measuring unit that measures the appearance frequency of the constituent words of the text data;
A model input unit to which at least two or more user input models are input from an input device for inputting a user input model;
Time-series data co-occurrence score calculation that calculates a co-occurrence score indicating the degree of appearance of both the user input model and the time-series data for each user input model based on the appearance frequency of the constituent words of the text data And
A time-series data correlation score calculation unit that calculates a correlation score indicating a correlation strength of each time-series data pair created by combining any two of the time-series data based on the calculated co-occurrence score; ,
A priority order determining unit for determining a priority order for calculating a correlation coefficient of the time series data pair based on a correlation score of the time series data pair;
A correlation coefficient calculating unit for calculating a correlation coefficient of the time series data pair according to the determined priority;
A knowledge extraction device comprising: The time series data co-occurrence score calculation unit
The appearance frequency of the constituent words of the text data that is the same as the user input model in the text data, and the appearance frequency of the constituent words of the text data that are the same as the words corresponding to the time-series data The knowledge extraction apparatus according to claim 14, wherein a co-occurrence score between the user input model and the time-series data is calculated using. A time-series data storage unit that stores time-series data and a time-series data ID for specifying the time-series data in association with each other;
A correspondence table storage unit for storing words and time-series data IDs in association with each other;
For the time series data ID stored in the time series data storage unit, a time series for acquiring a word list composed of one or more words stored in the corresponding notation part corresponding to the same time series data ID A data word list acquisition unit;
Further comprising
The time series data co-occurrence score calculation unit
Calculating a co-occurrence score of each word constituting the word list of the time-series data and the user input model,
16. The knowledge extracting apparatus according to claim 14, wherein a co-occurrence score having the maximum absolute value among the calculated co-occurrence scores is set as a co-occurrence score between the time series data and the model. The time series data correlation score calculation unit includes:
Combining the user input models, creating a user input model pair consisting of a first user input model and a second user input model which are arbitrary two user input models;
A co-occurrence score between the first user input model and the first time-series data for the time-series data pair consisting of the first time-series data and the second time-series data, which is a correlation score calculation target, , A multiplication value of the co-occurrence score of the second user input model and the second time series data, a co-occurrence score of the second user input model and the first time series data, and Calculating a multiplication value of the first user input model and the co-occurrence score of the second time-series data;
The largest value among the multiplication values calculated for the user input pair consisting of all two arbitrary user input models assumed is the correlation score of the time-series data pair that is the correlation score calculation target. The knowledge extraction device according to any one of claims 14 to 16.   The knowledge extracting device according to any one of claims 14 to 17, wherein the priority determining unit assigns priorities in descending order of correlation scores of the time series data pairs. The correlation coefficient calculation unit
A plurality of correlation coefficients are calculated for one time-series data pair by sequentially generating a predetermined delay between the time-series data and calculating a correlation coefficient for each delay,
The knowledge extraction device according to any one of claims 14 to 18, wherein a correlation coefficient having a maximum absolute value among the plurality of correlation coefficients is set as a correlation coefficient of the time series data pair. A correlation determining unit that determines whether or not there is a correlation between the time-series data constituting the time-series data pair based on the correlation coefficient of the time-series data pair;
The knowledge extraction device according to any one of claims 14 to 19, wherein the correlation determination unit determines that there is a correlation when an absolute value of a correlation coefficient is equal to or greater than a predetermined value. A computer program for causing a computer to function as a knowledge extraction device that extracts knowledge using a plurality of user input models that are text data, time series data and mutually correlated information,
Text dividing means for dividing text data into constituent words;
Appearance frequency measuring means for measuring the appearance frequency of the constituent words of the text data;
Model input means for inputting at least two or more user input models from an input device for inputting a user input model;
Time-series data co-occurrence score calculation that calculates a co-occurrence score indicating the degree of appearance of both the user input model and the time-series data for each user input model based on the appearance frequency of the constituent words of the text data Means,
Time-series data correlation score calculating means for calculating a correlation score indicating the strength of correlation of each time-series data pair created by combining any two of the time-series data based on the calculated co-occurrence score; ,
Priority order determining means for determining a priority order for calculating a correlation coefficient of the time series data pair based on the correlation score of the time series data pair;
A computer program that functions as a correlation coefficient calculation unit that calculates a correlation coefficient of the time-series data pair according to the determined priority order. A knowledge extraction method for extracting knowledge using a plurality of user input models which are text data, time series data and mutually correlated information,
A text splitting step for splitting the text data into constituent words;
An appearance frequency measuring step for measuring an appearance frequency of the constituent words of the text data;
A model input step in which at least two or more user input models are input from an input device for inputting a user input model;
Time-series data co-occurrence score calculation that calculates a co-occurrence score indicating the degree of appearance of both the user input model and the time-series data for each user input model based on the appearance frequency of the constituent words of the text data Steps,
A time-series data correlation score calculating step for calculating a correlation score indicating a correlation strength of each time-series data pair created by combining any two of the time-series data based on the calculated co-occurrence score; ,
A priority determining step for determining a priority for calculating a correlation coefficient of the time series data pair based on the correlation score of the time series data pair;
A correlation coefficient calculating step of calculating a correlation coefficient of the time series data pair according to the determined priority;
A knowledge extracting method characterized by comprising: The time series data co-occurrence score calculating step includes:
The appearance frequency of the constituent words of the text data that is the same as the user input model in the text data, and the appearance frequency of the constituent words of the text data that are the same as the words corresponding to the time-series data The knowledge extraction method according to claim 22, wherein a co-occurrence score between the user input model and the time-series data is calculated using For the time-series data ID of the time-series data storage unit stored in association with the time-series data in order to identify the time-series data, the same notation from the corresponding notation 100 million parts that stores the word and the time-series data ID in association with each other. A time series data word list obtaining step of obtaining a word list by obtaining one or more words corresponding to the time series data ID;
The time series data co-occurrence score calculating step includes:
Calculating a co-occurrence score of each word constituting the word list of the time-series data and the user input model,
The knowledge extraction method according to claim 22 or 23, wherein a co-occurrence score having the maximum absolute value among the calculated co-occurrence scores is used as a co-occurrence score between the time series data and the model. The time series data correlation score calculating step includes:
A user input model pair creating step of creating a user input model pair composed of a first user input model and a second user input model which are arbitrary two user input models by combining the user input models;
A co-occurrence score between the first user input model and the first time-series data for the time-series data pair consisting of the first time-series data and the second time-series data, which is a correlation score calculation target, , A multiplication value of the co-occurrence score of the second user input model and the second time series data, a co-occurrence score of the second user input model and the first time series data, and A multiplication step of calculating a multiplication value of the first user input model and the co-occurrence score of the second time-series data;
Correlation score determination using a maximum value among the multiplication values calculated for the user input pair consisting of all two arbitrary user input models assumed as a correlation score of the time series data pair that is the correlation score calculation target Steps,
The knowledge extracting method according to any one of claims 22 to 24, characterized by comprising:   The knowledge extraction method according to any one of claims 22 to 25, wherein the priority order determining step assigns priorities in descending order of correlation scores of the time series data pairs. The correlation coefficient calculating step includes:
Calculating a plurality of correlation coefficients for one time-series data pair by sequentially generating a predetermined delay between the time-series data and calculating a correlation coefficient for each delay;
Of the plurality of calculated correlation coefficients, the correlation coefficient having the maximum absolute value is set as the correlation coefficient of the time-series data pair; and
The knowledge extraction method according to any one of claims 22 to 26, characterized by comprising:   The correlation determination step of determining whether or not there is a correlation between time-series data constituting the time-series data pair based on a correlation coefficient of the time-series data pair. The knowledge extraction method according to any one of -27.

Images