JP6441930B2 - Data analysis apparatus, data analysis apparatus control method, and data analysis apparatus control program - Google Patents

Description

  The present invention relates to a data analysis apparatus or the like that can extract data related to a predetermined case from newly acquired data.

  Due to the background that corporate legal risks are increasing, such as companies pursuing violations of the antitrust law (antitrust law) due to suspicion of participating in the price cartel, and leaking confidential information by internal guides, There is a need for a system that detects such fraud. For example, Patent Document 1 discloses a document separation system that analyzes digitized document information collected for submission as evidence in a lawsuit and separates the information so that it can be easily used in a lawsuit. .

  On the other hand, techniques for recording behaviors related to computer use (such as which file has been accessed) have been proposed. For example, Patent Document 2 discloses a display method for detecting information file leakage useful for information leakage prevention measures by grasping the behavior of an information network user.

JP 2013-182338 A JP 2007-304943 A

  For example, the conventional system disclosed in Patent Literature 1 cannot detect a sign of fraudulent activity as described above from e-mail distributed daily on a network. Since the system analyzes related documents to be submitted in a lawsuit filed after the fraud has occurred, it assumes that all documents subject to the analysis exist in advance. is there.

  As a method of capturing an ongoing fraud, for example, a user behavior is recorded as in the display method disclosed in Patent Document 2, and a predetermined behavior (behavior defined as “fraud”) regarded as a problem is used. If it is discovered, there is a way to alert the administrator. However, in the above method, (a) when the predetermined behavior is discovered, it is often after the fraud has already occurred, and (b) a warning requirement is set in order to detect the fraud in advance. The more you loosen, the more frequently you will be alerted and the less effective monitoring will be.

  Further, the prior art disclosed in Patent Document 1 or 2 is only specialized for a specific fraud and is not general-purpose, and is not applicable to cases other than the specific fraud.

  The present invention has been made in view of the above problems, and its purpose is to extract data related to a predetermined case by analyzing current data based on the result of analyzing past data. Providing a simple data analysis device.

  In order to solve the above problems, a data analysis apparatus according to an aspect of the present invention is a data analysis apparatus capable of extracting data related to a predetermined case from newly acquired data, and is related to the predetermined case. When undecided data that has not been determined whether or not to be newly acquired is newly acquired by the user, it is determined whether or not the threshold that is the basis of the determination for the undecided data is related to the predetermined case. For the already-determined data, the threshold value specifying unit specified from the score calculated as an index indicating the strength of the relationship with the predetermined case, the threshold value specified by the threshold specifying unit, and the undetermined data are calculated. And a data setting unit that sets undecided data as data to be reported to the user according to the result of comparison with the score.

  Further, in the data analysis device according to one aspect of the present invention, the threshold value specifying unit uses, as a threshold value, a score that can exceed the target value set for the relevance ratio among the scores calculated for the already determined data. Can be identified.

  In addition, the data analysis device according to one aspect of the present invention compares the score calculated for undecided data with the threshold specified by the threshold specifying unit, thereby determining whether the score exceeds the threshold. The data setting unit can set undecided data as data to be reported to the user when it is determined by the excess determining unit that the data has been exceeded.

  Further, the data analysis apparatus according to one aspect of the present invention is based on the element evaluation unit that evaluates the data elements included in the already-determined data based on predetermined criteria, and the results evaluated by the element evaluation unit, A score calculation unit that calculates a score may be further included.

  In the data analysis apparatus according to one aspect of the present invention, the element evaluation unit determines a predetermined amount of transmission information indicating a dependency relationship between the data element and the result of the user's determination on the already determined data including the data element. As one of the criteria, the data element can be evaluated.

  In addition, the data analysis apparatus according to one aspect of the present invention provides a result of determination by a user whether or not the data set by the data setting unit is related to a predetermined case via the predetermined input unit. A result acquisition unit acquired from the user is further provided, and the element evaluation unit can evaluate each data element included in the data set by the data setting unit based on the result acquired by the result acquisition unit.

  The data analysis apparatus according to one aspect of the present invention further includes a storage unit that associates the data element evaluated by the element evaluation unit with the result of evaluation of the data element and stores the data element in a predetermined storage unit. You may prepare.

  Further, in the data analysis apparatus according to one aspect of the present invention, the undecided data includes unique data elements that can respectively identify a plurality of persons or organizations, and the data setting unit includes unique data elements from the undecided data, respectively. Extracting and visualizing the strength of connection between a plurality of persons or organizations by estimating the correspondence between the first unique data element and a second unique data element different from the first unique data element Can do.

  In addition, the data analysis apparatus according to one aspect of the present invention provides a result of a determination made by a user whether or not data extracted from a predetermined data group is related to a predetermined case via the predetermined input unit. A determination data acquisition unit that acquires the determination data by acquiring it from the user may be further provided.

  In addition, the data analysis apparatus according to one aspect of the present invention may further include a relationship adding unit that adds relationship information indicating that the data is related to a predetermined case to the data set by the data setting unit. .

  In the data analysis apparatus according to one embodiment of the present invention, the data may be a document digitized so that it can be processed by a computer, and the data element may be a keyword included in the document.

  In the data analysis device according to one embodiment of the present invention, the data may be voice that has been digitized so that it can be processed by a computer, and the data element may be partial voice included in the voice.

  In order to solve the above problem, a control method for a data analysis apparatus according to an aspect of the present invention is a control method for a data analysis apparatus capable of extracting data related to a predetermined case from newly acquired data. Whether or not the threshold value that is the basis of the judgment for the undecided data is related to the predetermined case when new undecided data that has not been determined whether or not it is related to the predetermined case For the already determined data determined by the user from the score respectively calculated as an index indicating the strength of the relationship with the predetermined case, the threshold specified in the threshold specifying step, and the undetermined data A data setting step for setting undecided data as data to be reported to the user according to the result of comparison with the score calculated for

  In order to solve the above problems, a control program for a data analysis apparatus according to an aspect of the present invention is a control program for a data analysis apparatus capable of extracting data related to a predetermined case from newly acquired data. When a new undetermined data that has not been determined whether or not it is related to a predetermined case is newly acquired in the computer, a threshold that is the basis of the determination for the undetermined data is related to the predetermined case. Threshold determination function that specifies from the score respectively calculated as an index indicating the strength of the relationship with the predetermined case, and the threshold determination function Data that sets the undetermined data as data to be reported to the user according to the result of comparing the threshold and the score calculated for the undetermined data. To realize the setting function.

  According to one aspect of the present invention, a data analysis device, a data analysis device control method, and a data analysis device control program newly acquire undecided data that has not been determined whether or not it relates to a predetermined case. If the judgment data for which the user has determined whether or not the threshold that is the basis of the judgment for the undecided data is related to the predetermined case, the strength of the relationship with the predetermined case The undetermined data is set as data to be reported to the user according to the result of comparing the threshold and the score calculated for the undetermined data.

  With the above configuration, the data analysis apparatus and the like have an effect that data related to a predetermined case can be extracted by analyzing current data based on the result of analyzing past data.

It is a block diagram which shows the principal part structure of the document analysis system which concerns on embodiment of this invention. It is a schematic diagram which shows an example of the said document analysis system roughly. It is a table | surface which shows the minimum score corresponding to the number of documents to which the review result was provided, (a) shows the case where a target precision is set to 100%, (b) sets the target precision to 90% Show the case. It is a flowchart which shows an example of the process which the said document analysis system performs.

  An embodiment of the present invention will be described with reference to FIGS.

[Outline of Document Analysis System 100]
The document analysis system (data analysis apparatus) 100 is an information processing system that can extract a document related to a predetermined case from a newly acquired document by analyzing a digital document. The document analysis system 100 only needs to include a computer that can execute the processing described below, and can be realized using, for example, a server device, a personal computer, a mainframe, a workstation, or other electronic devices.

  FIG. 2 is a schematic diagram schematically showing an example of the document analysis system 100. As shown in FIG. 2, the reviewer (user) determines whether or not the document is related to a predetermined case, and inputs the determined result (review result 5 a) to the document analysis system 100.

  Here, the “document” is data digitized so that it can be processed by a computer. For example, e-mail, technical documents, presentation materials, spreadsheet materials, financial statements, meeting materials, contracts, Organization charts, business plans, etc. can be widely included. FIG. 2 shows an example in which the document analysis system 100 takes in an electronic mail distributed on a network every day and analyzes a document included in the electronic mail.

  In addition, the “predetermined case” includes an incident that occurs when a general user who uses the document in the organization takes an illegal act and / or a preliminary act thereof, for example, an incident in which confidential information is leaked to the outside. This includes a wide range of incidents that the organization wants to prevent from occurring, such as incidents that have been negotiated with other companies, incidents that have been decorated with financial results, incidents that have been fictitiously charged to a trading company, and incidents that are unfavorable to other companies. However, the predetermined case is not limited to the above example, and may include a wide range of cases that can generate related data (eg, document, audio, video, etc.).

  Based on the review result 5a, the document analysis system 100 evaluates keywords (data elements) included in the document (determined data) based on predetermined criteria (for example, the amount of transmitted information). Then, the document analysis system 100 calculates, for each of the documents, a score indicating the strength of the relationship with the predetermined case based on the evaluated result, and determines the relevance rate (relevant to the predetermined case). The minimum score that can exceed the target value (target relevance ratio) set for the ratio of the recorded document to the document group including a predetermined number of documents is specified as the relevance threshold.

  That is, the document analysis system 100 sets the matching threshold based on the review result 5a given by the reviewer (the result of human judgment on past data), and the score exceeding the matching threshold is set. As a document that is highly likely to be related to the predetermined case, the list result 5b (information that can be listed and presented to the reviewer) can be returned to the reviewer. In other words, the document analysis system 100 can extract data related to a predetermined case by analyzing current data based on the result of analyzing past data. Thereby, the document analysis system 100 can detect, for example, a sign that an illegal act occurs.

[Configuration of Document Analysis System 100]
FIG. 1 is a block diagram showing a main configuration of the document analysis system 100. As shown in FIG. 1, the document analysis system 100 includes a control unit 10 (a data extraction unit 11, a result acquisition unit 12, an element evaluation unit 13, a score calculation unit 14, a score specification unit 15, an excess determination unit 16, a data setting. Unit 17, relationship giving unit 18, storage unit 19), receiving unit 20, input unit 40, display unit 50, and storage unit 30.

  The control unit 10 comprehensively controls various functions of the document analysis system 100. The control unit 10 includes a data extraction unit 11, a result acquisition unit 12, an element evaluation unit 13, a score calculation unit 14, a score specification unit 15, an excess determination unit 16, a data setting unit 17, a relationship assignment unit 18, and a storage unit 19. Including.

  The data extraction unit (determined data acquisition unit) 11 extracts a predetermined number of documents 1a to be determined by a reviewer from a predetermined document group (data group) as to whether or not it is related to a predetermined case. The document group may be data distributed on the network, or may be data stored in advance in the storage unit 30.

  The data extraction unit 11 can present the document 1a to the reviewer by outputting the extracted document 1a to the display unit 50. As a result, the reviewer can give, for example, the review result 5a indicating that the document 1a is “related to the predetermined case” or “not related to the predetermined case” to the document 1a. In addition, the data extraction unit 11 outputs the document 1 a to the result acquisition unit 12 and the element evaluation unit 13.

  When the document 1a is input from the data extraction unit 11, the result acquisition unit (determined data acquisition unit) 12 determines whether the reviewer determines whether or not the document 1a is related to a predetermined case (review result 5a). Is obtained via the input unit 40, and the review result 5a is output to the element evaluation unit 13 and the score identification unit 15.

  The element evaluation unit 13 evaluates each keyword (data element) included in the document 1a determined by the reviewer whether or not it is related to a predetermined case based on a predetermined criterion. The element evaluation unit 13 uses, for example, the amount of transmitted information representing the dependency relationship between the keyword and the result (review result 5a) determined by the reviewer for the document 1a including the keyword as one of the predetermined criteria. The keyword can be evaluated by calculating the weight of the keyword. Thereby, since the document analysis system 100 can accurately evaluate the keyword, it is possible to accurately extract data related to a predetermined case.

  Alternatively, the element evaluation unit 13 may evaluate the keyword by assigning a predetermined weight to the keyword. In this case, for example, the element evaluation unit 13 can assign a weight of “1” to the keyword.

  The “keyword” is a meaningful character string (morpheme). For example, a sentence “classify a document” includes keywords “document” and “classification”. The element evaluation unit 13 outputs the keyword information 5 c that is a pair of the keyword and the weight of the keyword to the score calculation unit 14 and the storage unit 19.

  Based on the result (keyword information 5c) evaluated by the element evaluation unit 13, the score calculation unit 14 calculates a score 5d indicating the strength of the relationship with the predetermined case for each document 1a, and the score 5d is calculated. The result is output to the score specifying unit 15. When the document 1b (data that has not been determined whether or not related to a predetermined case) is newly acquired from the receiving unit 20, the score calculation unit 14 calculates a score 5e for the document 1b, The score 5e is output to the excess determination unit 16.

  The score calculation unit 14 can calculate the score of the document by adding the weights of keywords appearing in the document. For example, as a result of the text “adjust price” included in the document, the keywords “price” and “adjustment” are evaluated by the element evaluation unit 13, respectively. As a result, “1.2” and “2. When the weight “2” is set, the score calculation unit 14 can calculate the score of the document as “3.4” (1.2 + 2.2).

  Specifically, the score calculation unit 14 generates a keyword vector indicating whether or not a predetermined keyword is included in the document. The keyword vector is a vector indicating whether or not a predetermined keyword associated with the element is included in the document when each element of the keyword vector takes a value of “0” or “1”. It is. For example, when the keyword “price” is included in the document, the score calculation unit 14 changes the element corresponding to the “price” of the keyword vector from “0” to “1”. Then, the score calculation unit 14 calculates the inner product of the keyword vector (vertical vector) and the weight vector (vertical vector using the weight for each keyword as an element) as in the following formula, thereby calculating the score of the document. S is calculated.

ここで、ｓはキーワードベクトルを表し、Ｗは重みベクトルを表す。なお、Ｔは行列・ベクトルを転置する（行と列とを入れ替える）ことを表す。

Here, s represents a keyword vector, and W represents a weight vector. T represents transposing a matrix / vector (replaces rows and columns).

  Or the score calculation part 14 may calculate the score S according to the following formula | equation.

ここで、ｍ_ｊは、ｊ番目のキーワードの出現頻度を表し、ｗ_ｉは、ｉ番目のキーワードの重みを表す。なお、スコア算出部１４は、文書１ａおよび／または文書１ｂに含まれる第１キーワードが評価された結果（第１キーワードの重み）と、当該文書１ａおよび／または文書１ｂに含まれる第２キーワードが評価された結果（第２キーワードの重み）とに基づいて、スコア５ｄおよび／またはスコア５ｅを算出してよい。また、スコア算出部１４は、文書１ａおよび／または文書１ｂにそれぞれ含まれるセンテンスごとに、スコア５ｄおよび／または５ｅを算出してよい（いずれも後で詳細に説明する）。

Here, m _j represents the appearance frequency of the j-th keyword, and w _i represents the weight of the i-th keyword. Note that the score calculation unit 14 determines that the first keyword included in the document 1a and / or the document 1b is evaluated (the weight of the first keyword) and the second keyword included in the document 1a and / or the document 1b. Based on the evaluated result (weight of the second keyword), the score 5d and / or the score 5e may be calculated. The score calculation unit 14 may calculate the scores 5d and / or 5e for each sentence included in the document 1a and / or the document 1b (both will be described in detail later).

  The score specifying unit (threshold specifying unit) 15 is configured to set a target value (target matching) indicating a ratio of the document 1a determined to be related to a predetermined case to a document group including a predetermined number of documents. The minimum score that can exceed (rate) is identified as the fitness threshold 6. Specifically, when the score 5d is input from the score calculation unit 14, the score specifying unit 15 sorts the scores 5d in descending order. Next, the score specifying unit 15 scans the review result 5a given to the document 1a in order from the document 1a having the maximum score 5d (score rank is first), and “relevant to a predetermined case”. The ratio of the number of documents to which the review result 5a is given to the number of documents that have been scanned at the present time (the relevance ratio) is sequentially calculated.

  For example, when the number of documents 1a to which the review result 5a is assigned is 100, when the scan is finished for the documents whose score ranks are 1st to 20th, the review result is “related to a predetermined case”. When the number of documents to which 5a is assigned is 18, the score specifying unit 15 calculates the relevance rate as 0.9 (18/20). Alternatively, when scanning of the documents whose score ranks are 1st to 40th is finished, when the number of documents to which the review result 5a “related to a predetermined case” is given is 35, the score specifying unit 15 calculates the precision as 0.875 (35/40).

  The score specifying unit 15 calculates all the relevance ratios for the document 1a, and specifies the minimum score that can exceed the target relevance ratio. Specifically, the score specifying unit 15 scans the relevance ratio calculated for the document 1a in order from the document 1a having the minimum score 5d (score rank is 100th), and the relevance ratio is the target. When the accuracy rate is exceeded, the score corresponding to the accuracy rate is output to the excess determination unit 16 and the storage unit 19 as the minimum score (matching threshold 6) that can maintain the target accuracy rate.

  Based on the result (keyword information 5c) evaluated by the element evaluation unit 13, the excess determination unit 16 determines that the score 5e calculated for the document 1b that has not yet been determined whether or not it is related to a predetermined case It is determined whether or not the threshold value 6 is exceeded, and the determined result (determination result 5f) is output to the data setting unit 17.

  The data setting unit 17 sets the document 1b as a document to be reported to the reviewer when the excess determination unit 16 determines that the number is exceeded. The data setting unit 17 sets the document 1b as a document to be reported to the reviewer, for example, by setting a flag for the document 1b that has exceeded the conformance threshold 6. The data setting unit 17 outputs setting information 5g that can specify the set document to the relationship adding unit 18.

  The relationship assigning unit 18 assigns relationship information (review result by the document analysis system 100) indicating that the document 1b is related to a predetermined case to the document 1b set by the data setting unit 17. The relationship assigning unit (display processing unit) 18 outputs the list result 5b to the display unit 50, whereby the document 1b set by the data setting unit 17 (the document determined by the document analysis system 100 to be related to a predetermined case). ) Can be displayed as a list.

  When the keyword information 5 c is input from the element evaluation unit 13, the storage unit 19 associates the keyword included in the keyword information 5 c with the result (weight) of evaluation of the keyword and stores it in the storage unit 30. To do. Thus, the document analysis system 100 can extract data related to a predetermined case by analyzing the current data based on the result of analyzing past data (weight as a result of evaluating the keyword). In addition, when the adaptation threshold 6 is input from the score specifying unit 15, the storage unit 19 stores the adaptation threshold 6 in the storage unit 30.

  The input unit (predetermined input unit) 40 receives an input (review result 5a) from a reviewer. FIG. 1 shows a configuration in which the document analysis system 100 includes an input unit 40 (for example, a configuration in which a keyboard, a mouse, and the like are connected as the input unit 40). The input unit 40 communicates with the document analysis system 100. It may be an external input device (for example, a client terminal) that is connected as possible.

  The receiving unit 20 receives the document 1a and / or the document 1b from the network via a communication network according to a predetermined communication method. It is only necessary that the receiving unit 20 has an essential function for realizing communication with an external device (for example, a terminal used by a general user), and a communication line, a communication method, a communication medium, and the like are not limited. The receiving unit 20 can be configured by a device such as an Ethernet (registered trademark) adapter, for example. The receiving unit 20 can use a communication method or a communication medium such as IEEE 802.11 wireless communication or Bluetooth (registered trademark).

  The display unit 50 is a device that displays an interface screen that can be operated by the reviewer. FIG. 1 shows a configuration in which the document analysis system 100 includes a display unit 50 (for example, a configuration in which a liquid crystal display or the like is connected as the display unit 50). The display unit 50 can communicate with the document analysis system 100. It may be an external display device (for example, a client terminal) connected to the.

  The storage unit (predetermined storage unit) 30 is a storage device composed of an arbitrary recording medium such as a hard disk, an SSD (silicon state drive), a semiconductor memory, a DVD, and the like, and includes documents 1a and keyword information 5c. A threshold 6 and / or a control program capable of controlling the document analysis system 100 are stored. 1 illustrates a configuration in which the document analysis system 100 includes the storage unit 30, the storage unit 30 may be an external storage device connected to be communicable with the document analysis system 100.

[Performance verification of document analysis system 100]
FIG. 3 is a table showing the minimum score corresponding to the number of documents to which the review result 5a is given. (A) shows a case where the target precision is 100%, and (b) is the target precision. Is 90%.

  As illustrated in FIG. 3A, when the number of documents determined by the reviewer as to whether or not it is related to a predetermined case is 100 (“sample number” in the table of FIG. 3 is “100”). ), The lowest rank that can achieve the target precision 100% is 11th, and the score corresponding to the rank (the minimum score that can achieve the precision 100%) is 0.110. is there. The document analysis system 100 sets the minimum score as a conformance threshold, and a document 1b having a score exceeding the conformance threshold is a document that can maintain a conformance rate of 100% (ie, related to a predetermined case). Document).

  In order to verify the validity of the conformance threshold, a document having a score exceeding the conformance threshold was extracted from 7994 documents to which the review result 5a was given by the reviewer. Note that the document is a special document prepared for verifying the performance of the document analysis system 100, and the document that is analyzed by the document analysis system 100 is only related to a predetermined case. Note that the document 1b has not yet been determined.

  As a result, 766 documents exceeded the conformance threshold, and a review result 5 a “related to a predetermined case” was given to 605 of these documents. That is, as long as the review result 5a is given to only 100 documents, the document analysis system 100 can perform a predetermined case with an accuracy (accuracy rate) of 79% (605/766 = 0.790) for about 8000 documents. It is proved quantitatively that documents related to can be extracted.

  As shown in FIG. 3A, the accuracy (accuracy rate) of the document analysis system 100 increases as the number of documents determined by the reviewer as to whether or not it is related to a predetermined case, and the target conformity is increased. It can be seen that the rate is approaching (see the column “Fit rate” included in “Total sample” in the figure). As shown in FIG. 3B, the above tendency is also established when the target precision is lowered to 90%.

  As described above, if the document analysis system 100 gives the result (review result 5a) determined by the reviewer to a part of an arbitrary document in which the total number of samples cannot be determined, the remaining most documents are processed. Can be sorted with high accuracy. That is, the document analysis system 100 can extract data related to a predetermined case by analyzing current data based on the result of analyzing past data. As a result, the document analysis system 100 can detect signs of increasing legal risks such as leaking confidential information to the outside or rigging other companies without manpower.

[Processing executed by the document analysis system 100]
FIG. 4 is a flowchart illustrating an example of processing executed by the document analysis system 100. In the following description, parenthesized “˜step” represents each step included in the control method of the data analysis apparatus.

  First, the data extraction unit 11 extracts a predetermined number of documents 1a to be judged by a reviewer from a predetermined document group as to whether or not they are related to a predetermined case (Step 1, hereinafter “Step” is “S”). Abbreviated). Next, the result acquisition unit 12 acquires a result (review result 5a) determined by the reviewer as to whether or not the document 1a is related to a predetermined case via the input unit 40 (S2). Next, the element evaluation unit 13 evaluates each keyword included in the document determined by the reviewer whether or not it is related to the predetermined case based on a predetermined criterion (S3). Then, the score calculation unit 14 calculates, for each document 1a, a score 5d indicating the strength of the relationship with the predetermined case based on the result (keyword information 5c) evaluated by the element evaluation unit 13 (S4). ), The score specifying unit 15 sets a target value (target relevance ratio) set for the relevance ratio indicating the ratio of the document 1a determined to be related to the predetermined case to the document group including the predetermined number of documents. Is specified as the matching threshold 6 (S5, threshold specifying step).

  Next, the score calculation unit 14 calculates a score 5e indicating the strength of the relationship with the predetermined case for each document 1b based on the result (keyword information 5c) evaluated by the element evaluation unit 13 ( S6). Based on the result (keyword information 5c) evaluated by the element evaluation unit 13, the excess determination unit 16 has a score 5e calculated for the document 1b that has not yet been determined whether or not it is related to the predetermined case. It is determined whether or not the conformance threshold 6 has been exceeded (S7), and if it is determined that it has been exceeded (YES in S7), the data setting unit 17 is a document that should report the document 1b to the reviewer. (S8, data setting step). Finally, the relationship assigning unit 18 assigns relationship information (review result by the document analysis system 100) indicating that the document 1b is related to a predetermined case to the document 1b set by the data setting unit 17 ( S9).

  Note that the above control method may optionally include not only the above-described processing described with reference to FIG. 4 but also processing executed in each unit included in the control unit 10.

[Score calculation based on co-occurrence]
As described above, the score calculation unit 14 can calculate the score based on the result of evaluating the first keyword included in the document and the result of evaluating the second keyword included in the document. That is, when the first keyword appears in the document, the score calculation unit 14 takes into account the frequency with which the second keyword appears in the document (that is, the correlation between the first keyword and the second keyword or co-occurrence). The document score.

  In this case, the score calculation unit 14 uses the correlation matrix (co-occurrence matrix) C that represents the correlation (co-occurrence) between the first keyword and the second keyword to express the following equation (instead of [Equation 1] above). Therefore, the score S can be calculated.

  The correlation matrix C is optimized in advance using a learning data set including a predetermined number of predetermined documents. For example, when a keyword “price” appears in a document, a value obtained by normalizing the number of occurrences of other keywords with respect to the keyword between 0 and 1 (that is, a maximum likelihood estimate) is the correlation matrix C. Stored in each element (therefore, the sum for each column of the correlation matrix C is 1).

  As described above, since the document analysis system 100 can calculate the score in consideration of the correlation between keywords, it can extract data related to a predetermined case with higher accuracy.

[Score calculation for each sentence]
As described above, the score calculation unit 14 can calculate a score for each sentence included in each document. In this case, the score calculation unit 14 generates a keyword vector indicating whether or not a predetermined keyword is included in the sentence included in the document for each sentence. And the score calculation part 14 calculates a score for every document according to the following formula.

ここで、ｓ_ｓは、ｓ番目のセンテンスに対応するキーワードベクトルである。なお、上記〔数４〕にしたがうスコアの算出においては、共起を考慮している（相関行列Ｃを用いている）ことに注意する。

Here, s _s is a keyword vector corresponding to the sth sentence. It should be noted that co-occurrence is taken into account (correlation matrix C is used) in calculating the score according to [Equation 4].

  TFnorm can be calculated as shown in [Formula 5] below.

ここで、上記〔数５〕において、ＴＦ_ｉはｉ番目のキーワードの出現頻度（Term Frequency）を表し、ｓ_ｊｉは上記ｉ番目のキーワードベクトルのｊ番目の要素を表し、ｃ_ｊｉは相関行列Ｃのｊ行ｉ列の要素を表す。

Here, in [Formula 5], TF _i represents the appearance frequency (Term Frequency) of the i-th keyword, s _ji represents the j-th element of the i-th keyword vector, and c _ji represents the correlation matrix C Of j rows and i columns.

  When the above [Equation 4] and [Equation 5] are put together, the score calculation unit 14 calculates the above score for each document by calculating the following [Equation 6].

ここで、上記〔数６〕において、ｗ_ｉは上記重みベクトルｗのｉ番目の要素である。

Here, in [Formula 6], w _i is the i-th element of the weight vector w.

  As described above, since the document analysis system 100 can calculate a score that correctly reflects the sentence meaning, it can extract data related to a predetermined case with higher accuracy.

[Phase analysis]
The document analysis system 100 can estimate a phase to which a predetermined case belongs and calculate a score according to the phase. Here, the “phase” is an index indicating each stage where the predetermined case progresses (classified according to the progress of the predetermined case).

  For example, the predetermined case is a fraud case called “collusion with another company”, and the document analysis system 100 determines whether or not an email distributed daily on the network is related to the fraud case. When the purpose is to detect a sign of the fraud case, the above phases include the “preparation phase for collecting information on competitors with other companies”, the “relationship building phase for building relationships with customers / competitors”, “ It may include a “competition phase” that presents a price to the customer, gets feedback, and communicates with the competitor regarding that feedback.

  In addition, time series information and generation process information are stored in the storage unit 30. Here, the “time-series information” is information indicating the temporal order of the phases, for example, a time development model indicating that the “competition phase” is reached from the “preparation phase” through the “relationship building phase”. It may be. The “generation process information” is information that models a process in which each keyword is generated in a certain phase, and may be, for example, a multinomial distribution model defined for each phase.

  The result acquisition unit 12 obtains the review result based on the result of the review by the reviewer as to whether or not the document 1a is related to the predetermined case and the result of determining which phase of the predetermined case the document 1a belongs to. The element evaluation unit 13 evaluates the keywords included in the document 1a for each phase (determines the weight of each keyword).

  When calculating the score 5e for the document 1b, the score calculation unit 14 estimates in which phase the document 1b is based on the generation process information. Specifically, the likelihood for each phase is calculated based on the generation process information, and the phase that maximizes the likelihood is estimated as the phase of the document 1b. And the score calculation part 14 calculates the score of the document 1b, respectively using the weight corresponding to the estimated phase. At this time, the score calculation unit 14 may use the correlation matrix C corresponding to the phase.

  The relationship assigning unit 18 can display the document 1b set by the data setting unit 17 in a listable manner and can display the estimated phase. At this time, the relationship assigning unit 18 can predict the possibility and timing of the estimated phase developing to the next phase based on the time series information, and can display the predicted result together.

  As described above, since the document analysis system 100 can accurately calculate a score according to a phase, it can extract data related to a predetermined case with higher accuracy.

[Recalculation of weights]
After the document 1b set by the data setting unit 17 (the document determined by the document analysis system 100 to be related to a predetermined case) is displayed in a listable manner by the relationship granting unit 18, the result acquisition unit 12 makes the determination Feedback can be received from the reviewer. That is, each reviewer can input whether the result determined by the document analysis system 100 is valid as the feedback. The element evaluation unit 13 can reevaluate each keyword based on the feedback.

  In other words, the element evaluation unit 13 can recalculate the weight based on the feedback newly obtained for the determination of the document analysis system 100. As a result, the document analysis system 100 can obtain a weight suitable for the document to be analyzed, and can accurately calculate a score based on the weight, so that data related to a predetermined case can be extracted with higher accuracy. .

[Display of person-organization correlation]
The data setting unit 17 extracts proper nouns appearing in the document 1b (for example, unique data elements such as a person's name, company name, place name, etc.), a predetermined proper noun (first unique data element) and others By estimating the correspondence with the proper noun (second unique data element), the strength of the connection between a plurality of persons or organizations can be visualized.

  For example, if an e-mail transmitted from person A to person B is analyzed as document 1b, and the document 1b contains a sentence "I will contact Mr. C", data setting unit 17 Extracts “person A”, “person B”, and “person C” and puts an arrow on each of the node indicating “person B” and the node indicating “person C” from the node indicating “person A”. The connected chart can be displayed. In this case, the data setting unit 17 may display the chart so that the thickness of the arrow varies depending on the strength of the correlation between the person and the organization.

  As described above, since the document analysis system 100 can easily understand and display the correlation between persons and organizations based on the result of analyzing the document, the subject that causes the fraud case can be identified without omission.

[Configuration in which server device provides part or all of functions]
The configuration (stand-alone configuration) in which the control program (data analysis device control program) capable of providing a function of analyzing data is mainly executed in the document analysis system 100 (data analysis device) has been described above. On the other hand, a configuration (cloud configuration) in which part or all of the control program is executed in the server device and the result of the executed processing is returned to the document analysis system 100 (user terminal) may be employed. That is, the data analysis device of the present invention can function as a server device that is communicably connected to a user terminal via a network. Thereby, the server device has the same effect as the document analysis system 100 when the document analysis system 100 provides a function.

[Example of software implementation]
The control block (particularly, the control unit 10) of the document analysis system 100 may be realized by a logic circuit (hardware) formed in an integrated circuit (IC chip) or the like, or using a CPU (Central Processing Unit). It may be realized by software. In the latter case, the document analysis system 100 includes a CPU that executes instructions of a control program that is software that implements each function, and a ROM (Read Only) in which the control program and various data are recorded so as to be readable by a computer (or CPU). Memory) or a storage device (these are referred to as “recording media”), a RAM (Random Access Memory) that expands the control program, and the like. The computer (or CPU) reads the control program from the recording medium and executes it, thereby achieving the object of the present invention. As the recording medium, a “non-temporary tangible medium” such as a tape, a disk, a card, a semiconductor memory, a programmable logic circuit, or the like can be used. The control program may be supplied to the computer via any transmission medium (such as a communication network or a broadcast wave) that can transmit the control program. The present invention can also be realized in the form of a data signal embedded in a carrier wave, in which the control program is embodied by electronic transmission.

  The above control program is implemented using, for example, a script language such as Python, ActionScript, JavaScript (registered trademark), an object-oriented programming language such as Objective-C, Java (registered trademark), or a markup language such as HTML5. it can. In addition, an information processing apparatus (for example, the document analysis system 100) including each unit that implements each function implemented by the control program, and a server apparatus that includes each unit that implements the remaining functions different from the respective functions, An analysis system including the above also falls within the scope of the present invention.

[Example applied to data other than documents]
Although the document analysis system 100 for analyzing a document has been described as an embodiment for implementing the data analysis apparatus of the present invention, the data analysis apparatus can also analyze data other than documents.

  For example, the data analysis apparatus of the present invention can be implemented in the form of a voice analysis system that analyzes voice. In this case, the voice analysis system (1) recognizes the voice, converts the content of the conversation included in the voice into characters (document data), and processes the document data in the same manner as the document analysis system 100. (2) The audio data may be processed as it is.

  In the case of (1), the speech analysis system converts speech data into document data by using an arbitrary speech recognition algorithm (for example, a recognition method using a hidden Markov model). Thus, the same processing as the processing executed by the document analysis system 100 is executed. Thereby, the voice analysis system has the same effect as the document analysis system 100.

  In the case of (2), the voice analysis system can distinguish whether or not the voice data is related to a predetermined case by extracting partial voices included in the voice data. For example, when voice data “adjust price” is obtained, the voice analysis system extracts partial voices “price” and “adjustment” from the voice data, and based on the result of evaluating the partial voice, Relevance information can be given to unsorted audio data. In this case, the speech analysis system can classify speech data using a time series data classification algorithm (for example, a hidden Markov model, a Kalman filter, a neural network, etc.). Thereby, the voice analysis system has the same effect as the document analysis system 100.

  Alternatively, the data analysis apparatus of the present invention can be implemented in the form of a video analysis system that analyzes video (moving images). In this case, the video analysis system can identify a person included in the frame image by extracting a frame image included in the video data and using an arbitrary face recognition technique. In addition, the video analysis system uses an arbitrary motion recognition technique (for example, a pattern matching technique may be applied), thereby enabling a partial video (all frame images included in the video to be included) included in the video data. The motion (motion) of the person can be extracted from the video including a part of the video. The video analysis system can sort the video data based on the person and / or motion. Thus, the video analysis system has the same effect as the document analysis system 100.

  That is, the data analysis apparatus of the present invention can analyze digital data (document, audio, video, etc.) in which information is developed in time series. Accordingly, the data analysis apparatus can extract data related to a predetermined case by analyzing current data based on the result of analyzing past data (document, audio, video, etc.) (for example, illegal It is possible to detect a sign that an action will occur.

[Additional Notes]
The present invention is not limited to the above-described embodiments, and various modifications can be made within the scope of the claims, and the technical means disclosed in different embodiments can be appropriately combined. Embodiments to be made are also included in the technical scope of the present invention. Furthermore, a new technical feature can be formed by combining the technical means disclosed in each embodiment.

  For example, the present invention can be expressed as follows. That is, based on the result evaluated by the element evaluation unit and the element evaluation unit that evaluates each data element included in the data determined by the user as to whether or not it is related to the predetermined case based on a predetermined standard A score calculation unit for calculating a score indicating the strength of the relationship with a predetermined case for each data, and a ratio of data determined to be related to the predetermined case to a data group including a predetermined number of data It is related to a given case based on the score evaluation unit that specifies the minimum score that can exceed the target value set for the accuracy rate shown as the adaptation threshold and the result evaluated by the element evaluation unit Exceeded by the excess determination unit that determines whether the score calculated for data that has not been determined yet exceeds the conformance threshold and the excess determination unit If it is determined that the data analyzer and a data setting unit for setting the data to be reported the data to the user.

  Alternatively, the present invention can be expressed as follows. That is, as an index indicating the strength of the relationship with a predetermined case, out of the scores calculated for the already-determined data determined by the user whether or not it is related to the predetermined case, When new data is obtained for the score identification unit that identifies the minimum score that can exceed the set target value, and whether or not it is determined whether or not it is related to a predetermined case, When it is determined that the calculated score exceeds the minimum score specified by the score specifying unit, and the excess determining unit determines that the calculated score exceeds the minimum score, A data analysis device including a data setting unit that sets data to be reported to the computer.

  Alternatively, the present invention can be expressed as follows. That is, when undecided data that has not been determined whether or not it is related to a predetermined case is newly acquired, the basic information that is the basis of the determination for the undecided data is related to the predetermined case An information specifying unit that specifies whether or not the data has already been determined by the user, and a data setting unit that sets undetermined data as data to be reported to the user based on the basic information specified by the information specifying unit; Data analysis device equipped with.

  In the data analysis apparatus, the score calculation unit determines whether or not the score indicating the strength of the relationship with the predetermined case is related to the predetermined case based on the result evaluated by the element evaluation unit. Is calculated for each sentence included in each document determined by the user, and the excess determination unit has not yet been determined whether it is related to a predetermined case based on the result evaluated by the element evaluation unit It is possible to determine whether or not the score calculated for each sentence included in each exceeds a fitness threshold.

  The present invention can be widely applied to personal computers, server devices, mainframes, workstations, and other electronic devices.

  1a: Document (determined data), 1b: Document (undetermined data), 5a: Review result (result determined by the user), 5d: Score, 5e: Score, 6: Relevance threshold, 11: Data extraction Part (predetermined data acquisition part), 12: result acquisition part (predetermined data acquisition part), 13: element evaluation part, 14: score calculation part, 15: score specification part (threshold specification part), 16: excess determination part , 17: data setting unit, 18: relationship adding unit, 19: storage unit, 30: storage unit (predetermined storage unit), 40: input unit (predetermined input unit), 100: document analysis system (data analysis device)

Claims (11)

An already-determined data acquisition unit that acquires a plurality of already-determined data determined by the user as to whether or not it is related to a predetermined case;
For each of the plurality of previously determined data, the data elements included in the already determined data, commentary worthy elements evaluated in accordance with the dependencies of the results the user decides on the data elements and the already determined data And
Based on the evaluated data element, a score calculation unit that calculates a score indicating the strength of the relationship between each of the plurality of already determined data and the predetermined case;
An undetermined data acquisition unit that newly acquires undetermined data that has not been determined by the user as to whether or not it relates to the predetermined case;
As an index for determining whether or not the undecided data is related to the predetermined case, a threshold specifying unit that specifies a threshold provided for the calculated score;
The threshold specifying unit, among the scores calculated for said plurality of previously determined data, already determined data determined to be related to the predetermined incidents, conform showing the percentage of data group including a predetermined number of data The minimum score that can exceed the target set for the rate is identified as the threshold ,
The score calculation unit uses the evaluation result of the data element corresponding to the data element included in the acquired undetermined data among the evaluation results of the evaluated data element, and the acquired undetermined data and calculating a score indicating the previous SL relationship in the strength of the predetermined cases,
An excess determination unit that determines whether or not a score calculated for the acquired undetermined data exceeds the threshold;
A data analysis apparatus comprising: a data setting unit configured to set undecided data having a score determined to exceed the data to be reported to the user. The element evaluation unit represents , for each of the plurality of already-determined data, a data element included in the already-determined data and a dependency relationship between the data element and a result of the determination made by the user with respect to the already-determined data data analyzer according to transmission information amount to claim 1, wherein the evaluation child by calculating the weight of the data element. Whether or not the data set by the data setting unit is related to the predetermined case, further comprising a result acquisition unit for acquiring a result determined by the user from the user via a predetermined input unit;
The said element evaluation part evaluates each data element contained in the data set by the said data setting part based on the result acquired by the said result acquisition part, respectively. Data analysis device.   4. The storage device according to claim 1, further comprising a storage unit that associates the data element evaluated by the element evaluation unit with a result of evaluating the data element and stores the data element in a predetermined storage unit. The data analysis device according to any one of the above. The undetermined data includes unique data elements that can respectively identify a plurality of persons or organizations,
The data setting unit extracts each of the unique data elements from the undetermined data, and estimates a correspondence relationship between the first unique data element and a second unique data element different from the first unique data element. The data analysis apparatus according to claim 1, wherein the strength of connection between the plurality of persons or organizations is visualized.   The data analysis apparatus according to claim 1, wherein the determination data acquisition unit acquires the plurality of determination data via a predetermined input unit.   The relation setting section for adding relation information indicating that the data is related to the predetermined case to the data set by the data setting section. The data analysis device according to one item. The data is a document digitized so that it can be processed by a computer,
The data analysis apparatus according to claim 1, wherein the data element is a keyword included in the document. The data is voice that has been digitized so that it can be processed by a computer,
The data analysis apparatus according to claim 1, wherein the data element is a partial sound included in the sound. A method for controlling a data analyzer,
Previously determining the data acquisition unit, and the previously determined data acquisition step whether related to cases of Jo Tokoro to obtain multiple previously determined data determined by the user,
Element evaluation unit, for each of the previous SL plurality of previously determined data, according to the data elements included in the already determined data, the dependency on the result of the user decides on the data elements and the already determined data and elements evaluation step that deserves Review Te,
Score calculating unit, based on the previous SL evaluated data element, and score calculation step of calculating the plurality of the respective pre-determined data a score indicating the relationship between the intensity of the predetermined cases, respectively,
Undetermined data acquisition unit, the undetermined data acquisition step whether related to previous SL predetermined incidents newly acquires undetermined data not determined by the user,
Threshold identifying unit as an index for before Symbol undetermined data to determine whether the relationship between the predetermined cases, a threshold specifying step of specifying a threshold provided for the calculated score,
In the threshold identifying step, the threshold specifying unit, among the score calculated for the previous SL plurality of previously determined data, already determined data determined to be related to the predetermined incidents, data group including a predetermined number of data specified as the threshold score of possible exceeding the minimum target value set for conformance ratio indicating the percentage of,
In the score calculating step, the score calculating unit uses the evaluation result of the data element corresponding to the data element included in the acquired undecided data among the evaluation results of the evaluated data element, calculating a score indicating the relationship between the intensity of the undetermined data and the previous SL predetermined cases that are,
Excess determination unit, the score calculated for undetermined data pre SL acquired, the excess judgment step of judging whether or not exceeded the threshold,
The method of data analysis device includes a data setting unit, the undetermined data having the determined scores to be pre Symbol exceeded, the data setting step of setting as the data to be reported to the user. A control program for a data analyzer,
On the computer,
An already-determined data acquisition function for acquiring a plurality of already-determined data determined by the user as to whether or not it is related to a predetermined case;
For each of the plurality of previously determined data, the data elements included in the already determined data, commentary worthy elements evaluated in accordance with the dependencies of the results the user decides on the data elements and the already determined data Function and
A score calculation function for calculating a score indicating the strength of the relationship between each of the plurality of already determined data and the predetermined case, based on the evaluated data element;
An undetermined data acquisition function for newly acquiring undetermined data that has not been determined by the user as to whether or not it relates to the predetermined case;
A threshold specifying function for specifying a threshold provided for the calculated score as an index for determining whether the undecided data is related to the predetermined case;
The threshold specific function, of the scores calculated for said plurality of previously determined data, already determined data determined to be related to the predetermined incidents, conform showing the percentage of data group including a predetermined number of data The minimum score that can exceed the target set for the rate is identified as the threshold,
The score calculation function uses the evaluation result of the data element corresponding to the data element included in the acquired undetermined data among the evaluation results of the evaluated data element, and the acquired undetermined data and calculating a score indicating the previous SL relationship in the strength of the predetermined cases,
An excess determination function for determining whether a score calculated for the acquired undetermined data exceeds the threshold; and
A control program for a data analysis device that realizes a data setting function for setting undetermined data having a score determined to be in excess as data to be reported to the user.

Images