JP2016012195A - Factor estimation device, program, and factor estimation method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable automatic estimation of candidates of factors relating to a request or a dissatisfaction, and also estimate a factor relating to a human relationship or a personal life event.SOLUTION: The factor estimation device is provided with: a subjective view detection unit 20 that performs document classification processing on inputted text data and gives, to classified each word group, a positive label indicating that the word group indicates a positive subjective view, or a negative label indicating that the word group indicates a negative subjective view; and a factor detection unit 30 that extracts, from a sentence, co-occurrence words that co-occur with the word group to which the positive label has been given or the word group to which the negative label has been given, and from a probability table constituted of the word groups that indicate the positive or negative subjective view, the word groups that indicate expression factors of the word groups indicating the subjective view, and numerical values that indicate expression probabilities of the word groups indicating the expression factors, extracts the word groups that include the extracted co-occurrence words, the word groups indicating the expression factors, the extracted word group that indicates the expression factors being selected in an order from the word group with higher expression probability and set as factor word groups.

Description

  The present invention relates to a technique for estimating a cause in which at least one of a word group showing a positive subjectivity or a word group showing a negative subjectivity appears in a sentence.

  2. Description of the Related Art Conventionally, purchasers and users of products can transmit opinions and impressions about products using the Internet. In Japanese Patent Laid-Open No. 2004-228867, there is a technique in which words relating to system operation are extracted from a document group showing product evaluation using a dictionary, and then those that are determined as “negative” are extracted as dissatisfied sentences by evaluation polarity classification (positive negative). It is disclosed. In Non-Patent Document 1, a product's popular opinion and dissatisfaction opinion is taken out from a review article on the Web (World Wide Web), and the product is arranged in a space summarized on one evaluation axis by principal component analysis. Discloses a system for supporting a consumer's purchase decision.

JP 2013-168043 A

DEIM Forum 2012 A9-2 Product Recommendation System Considering Dissatisfaction Opinions and Popular Opinions that Users Focus on

  However, in the technique described in Patent Document 1, the target for extracting requests and dissatisfaction is limited to matters related to the operation and malfunction of objects such as systems and products, and the target range is narrow. Furthermore, it does not mention the cause of demands and dissatisfaction. The technique described in Non-Patent Document 1 is intended for product review articles, and is intended to make it easier for a user to make a judgment when purchasing a product by collecting a plurality of popular reviews and dissatisfactions. There is no mention of the cause of its manifestation. Moreover, since both the patent document 1 and the non-patent document 1 are limited to matters relating to marketing such as products and products, requests and dissatisfactions regarding human relations and human life events are not targeted, The cause of human emotion cannot be estimated.

  The present invention has been made in view of such circumstances, and can automatically estimate a cause candidate corresponding to a request or dissatisfaction, and estimate a cause related to a human relationship or a human life event. It is an object to provide a cause estimation device, a program, and a cause estimation method.

  (1) In order to achieve the above object, the present invention takes the following measures. That is, the cause estimation apparatus of the present invention is a cause estimation apparatus that estimates a cause in which at least one of a word group indicating a positive subjectivity or a word group indicating a negative subjectivity appears in a sentence, and is input. The document classification processing is performed on the text data, and the positive label indicating the positive subjectivity for each classified word group or the word group indicating the negative subjectivity is indicated. A subjective detection unit for assigning a negative label, and extracting a co-occurrence word co-occurring with the word group with the positive label or the word group with the negative label from the sentence, An expression cause including the extracted co-occurrence word from a probability table composed of a word group indicating an expression, a word group indicating an expression cause of the word group indicating the subjectivity, and a numerical value indicating an expression probability of the word group indicating the expression Extract words that indicate The term group indicates expression causes the extracted, causing detector caused word group and elected expression probability descending order, characterized in that it comprises a.

  In this way, a document classification process is performed on the input text data, and a positive label indicating a positive subjectivity or a negative subjectivity word for each classified word group. A negative label indicating that it is a group, and a co-occurrence word that co-occurs with the word group to which the positive label is given or the word group to which the negative label is given is extracted from the sentence, positive or negative Including the extracted co-occurrence words from a probability table composed of a word group indicating the subjectivity, a word group indicating the cause of the expression of the word group indicating the subjectivity, and a numerical value indicating the expression probability of the word group indicating the cause of the expression The word group indicating the cause of expression is extracted, and the word group indicating the extracted expression cause is selected in descending order of the probability of expression as the cause word group, so the expression cause of the word group indicating positive or negative subjectivity Can be estimated automatically It made. In addition, since text is a target, it is not limited to products and systems, and it is possible to estimate the cause of expression of a word group indicating positive or negative subjectivity regarding human relationships or human life events.

  (2) Moreover, the cause estimation apparatus of the present invention is a word group indicating a positive label or a negative subjectivity indicating that it is a word group collected in advance and classified by a document classification process and indicating a positive subjectivity. It further includes a preprocessing unit that stores a word group to which a negative label indicating that is stored in a learning database and creates the probability table based on the word group stored in the learning database.

  In this way, collected in advance and classified by the document classification process, a positive label indicating that it is a word group indicating positive subjectivity or a negative label indicating that it is a word group indicating negative subjectivity is given. The word group is stored in the learning database, and the probability table is created based on the word group stored in the learning database, so that the cause of the expression of the word group showing positive or negative subjectivity is automatically estimated. It becomes possible to do.

  (3) Moreover, in the cause estimation apparatus of the present invention, the cause detection unit is given an affirmative label among the respective cause word groups when the plurality of selected cause word groups have the same occurrence probability. The causal word group having the smallest phrase interval is selected from the word group to which the negative label is assigned.

  Thus, when the expression probability of the selected plurality of causal word groups is the same, among the respective causal word groups, the phrase interval from the word group given a positive label or the word group given a negative label Since the causal word group having the smallest is selected, even if the expression probability is the same, the priority order can be given.

  (4) The program of the present invention is a program for estimating the cause of the occurrence of at least one of a word group showing positive subjectivity or a word group showing negative subjectivity in a sentence, and is input Document classification processing is performed on the text data, and a negative label indicating that it is a word group indicating a positive subjectivity or a word group indicating a negative subjectivity for each classified word group A process of assigning a label, a process of extracting a co-occurrence word co-occurring with a word group to which the positive label is given or a word group to which the negative label is given from the sentence, and a positive or negative subjectivity. An expression cause including the extracted co-occurrence word from a probability table composed of a word group indicating the expression cause of the word group indicating the subjectivity and a numerical value indicating the expression probability of the word group indicating the expression cause Extract word groups A processing that, the word group showing expression causes the extracted, the processing caused word group and elected expression probability descending order, a series of processing, and characterized by causing a computer to execute.

  In this way, a document classification process is performed on the input text data, and a positive label indicating a positive subjectivity or a negative subjectivity word for each classified word group. A negative label indicating that it is a group, and a co-occurrence word that co-occurs with the word group to which the positive label is given or the word group to which the negative label is given is extracted from the sentence, positive or negative Including the extracted co-occurrence words from a probability table composed of a word group indicating the subjectivity, a word group indicating the cause of the expression of the word group indicating the subjectivity, and a numerical value indicating the expression probability of the word group indicating the cause of the expression The word group indicating the cause of expression is extracted, and the word group indicating the extracted expression cause is selected in descending order of the probability of expression as the cause word group, so the expression cause of the word group indicating positive or negative subjectivity Can be estimated automatically It made. In addition, since text is a target, it is not limited to products and systems, and it is possible to estimate the cause of expression of a word group indicating positive or negative subjectivity regarding human relationships or human life events.

  (5) Moreover, the cause estimation method of the present invention is a cause estimation method for estimating a cause in which at least one of a word group indicating a positive subjectivity or a word group indicating a negative subjectivity appears in a sentence. The subjectivity detection unit performs document classification processing on the input text data, and displays an affirmative label or negative subjectivity indicating that it is a word group indicating a positive subjectivity for each classified word group. A co-occurrence word that co-occurs with the step of assigning a negative label indicating that it is a word group to be indicated, and the word group to which the positive label is assigned from the sentence or the word group to which the negative label is assigned in the cause detection unit From a probability table composed of a word group indicating positive or negative subjectivity, a word group indicating the cause of expression of the word group indicating the subjectivity, and a numerical value indicating the probability of expression of the word group indicating the cause of expression The extraction Extracting a word group indicating the cause of expression including the co-occurrence word and selecting the word group indicating the extracted expression cause in descending order of the probability of expression to be a cause word group, To do.

  In this way, a document classification process is performed on the input text data, and a positive label indicating a positive subjectivity or a negative subjectivity word for each classified word group. A negative label indicating that it is a group, and a co-occurrence word that co-occurs with the word group to which the positive label is given or the word group to which the negative label is given is extracted from the sentence, positive or negative Including the extracted co-occurrence words from a probability table composed of a word group indicating the subjectivity, a word group indicating the cause of the expression of the word group indicating the subjectivity, and a numerical value indicating the expression probability of the word group indicating the cause of the expression The word group indicating the cause of expression is extracted, and the word group indicating the extracted expression cause is selected in descending order of the probability of expression as the cause word group, so the expression cause of the word group indicating positive or negative subjectivity Can be estimated automatically It made. In addition, since text is a target, it is not limited to products and systems, and it is possible to estimate the cause of expression of a word group indicating positive or negative subjectivity regarding human relationships or human life events.

  According to the present invention, it is possible to automatically estimate the cause of expression of a word group showing a positive or negative subjectivity. In addition, since text is a target, it is not limited to products and systems, and it is possible to estimate the cause of expression of a word group indicating positive or negative subjectivity regarding human relationships or human life events.

It is a block diagram which shows schematic structure of the cause estimation apparatus which concerns on this embodiment. It is a flowchart which shows operation | movement of a pre-processing part. It is a flowchart which shows operation | movement of a co-occurrence word extraction part. It is a flowchart which shows operation | movement of a cause word candidate extraction part. It is a flowchart which shows operation | movement of a cause detection part. It is a flowchart which shows operation | movement of a cause detection part.

  The inventors of the present invention pay attention to the fact that the cause of the request or dissatisfaction could not be estimated in the existing technology, and the word group indicating the subjectivity such as the request or dissatisfaction, the word group indicating the subjectivity It is possible to automatically estimate the cause of the manifestation of subjectivity such as desire or dissatisfaction by using a probability table composed of a word group indicating the cause of the occurrence of the disease and a numerical value indicating the expression probability of the word group indicating the cause of the expression. The inventors have found that this is possible and have come to the present invention.

  That is, the cause estimation apparatus of the present invention is a cause estimation apparatus that estimates a cause in which at least one of a word group indicating a positive subjectivity or a word group indicating a negative subjectivity appears in a sentence, and is input. The document classification processing is performed on the text data, and the positive label indicating the positive subjectivity for each classified word group or the word group indicating the negative subjectivity is indicated. A subjective detection unit for assigning a negative label, and extracting a co-occurrence word co-occurring with the word group with the positive label or the word group with the negative label from the sentence, An expression cause including the extracted co-occurrence word from a probability table composed of a word group indicating an expression, a word group indicating an expression cause of the word group indicating the subjectivity, and a numerical value indicating an expression probability of the word group indicating the expression Extract words that indicate The term group indicates expression causes the extracted, causing detector caused word group and elected expression probability descending order, characterized in that it comprises a.

  With this configuration, the present inventors have been able to automatically estimate the cause of expression of a word group showing a positive or negative subjectivity. In addition, since the subject is a sentence, the present invention is not limited to products and systems, and it is possible to estimate the cause of expression of a word group indicating positive or negative subjectivity regarding human relations or human life events. Embodiments of the present invention will be specifically described below with reference to the drawings.

  FIG. 1 is a block diagram illustrating a schematic configuration of the cause estimation apparatus according to the present embodiment. The cause estimation apparatus includes a preprocessing unit 10, a subjective detection unit 20, and a cause detection unit 30. The pre-processing unit 10 includes a learning database 1 and a table creation unit 3. The learning database 1 is given in advance as a label, which is collected in advance and includes a request D that is a group of words indicating positive subjectivity, a dissatisfaction F that is a group of words indicating negative subjectivity, and a cause C in which at least one of them is expressed. Stored text. As the learning data, for example, various text data such as a review sentence for a product or the like, a SNS (Social Networking Service) post and its comment can be used. The table creation unit 3 creates a cause sentence probability table from a data set to which the learning data request D, dissatisfaction F, and cause C labels are assigned. The data set has, for example, the contents shown in the following table [learning data example].

Probability table table creating unit 3 creates, as the following table [probability table Example 1], refers to something probability that demand D _N dissatisfaction F _M and cause C _L and may occur that is a plurality of sets stored .

  There is no need to have three elements (desired / dissatisfied / cause) in the data, but the set data of the desired and the cause, dissatisfied / caused Set data can also be included.

  Requests, complaints, and causes can include words, phrases, and sentences. For example, requests and dissatisfactions can be created for each word, and the cause can be created by sentences. The combination of words, phrases, and sentences can be freely set.

  As shown in [Probability Table Example 1] above, the probability values in the probability table are C1 “window” as the cause for 100 content sentences “desiring wind” in request D and “hot” in dissatisfaction F. 35 cases of “closed”, 5 cases of C2 “no window in room”, 50 cases of C3 “air conditioner turned off”, 10 cases of C4 “air conditioner failed” The respective probabilities are calculated as P1 = 0.35, P2 = 0.05, P3 = 0.50, and P4 = 0.10.

  FIG. 2 is a flowchart showing the operation of the preprocessing unit. First, the learning database 1 is referred to (step S1), and a probability value is calculated (step S2). And the probability table of a cause sentence is created from the data set to which the desire D of learning data, dissatisfaction F, and the cause C label were provided (step S3, step S4).

  Next, in FIG. 1, the subjective detection unit 20 includes a demand determination unit 5 and a dissatisfaction determination unit 7. The subjectivity detection unit 20 inputs the text group 4 (text sentence). Note that the sentence may be a sentence such as a paragraph or a punctuation break as appropriate. The input text can be a communication text such as a posted text on the Internet or a series of mail text such as SNS or blog. The following table shows [Input text example of clause sentence].

  The request determination unit 5 classifies the document by a text search method or a vector space model, detects the request D, and labels it. The following table shows examples of labels given.

  For example, using the following related techniques 1 and 2, a desired feature space and a teacher vector are created from desired data included in the learning data. A feature vector of an input text sentence can be created based on the desired feature space, and a desired label can be assigned based on the similarity to the teacher vector. It is also possible to create and use a request classifier from the request data included in the learning data using the related technique 3.

[Related Technology 1] Bag-of-words model: http://en.wikipedia.org/wiki/Bag-of-words_model
[Related Technology 2] Mecab: http://mecab.googlecode.com/svn/trunk/mecab/doc/index.html
[Related Technology 3] SVMLIGHT: http://svmlight.joachims.org/
The dissatisfaction determination unit 7 detects dissatisfaction F by analyzing the emotional expression and labels it. The analysis of the emotional expression is realized by using a prepared dictionary, positive / negative determination, etc., as shown in the above related technologies and other conventional technologies. The following table shows examples of labels given.

  In FIG. 1, the cause detection unit 30 includes a co-occurrence word extraction unit 9, a cause word candidate extraction unit 11, and a cause determination unit 13. The co-occurrence word extraction unit 9 reads a sentence (request sentence) labeled as “request D” by the request determination unit 5 and a sentence (dissatisfaction sentence) labeled “dissatisfaction F” by the dissatisfaction determination unit 7. , And extract a combination of co-occurring words. As a method of associating, for example, it is possible to combine the most recent requested text starting from the detected dissatisfied text. For example, in the above table [labeling example 2], t5 F “hot” and tn D “want a wind” are associated with each other, “room temperature, humidity, air conditioner, patience” etc. co-occurring with these two Are extracted.

  FIG. 3 is a flowchart showing the operation of the co-occurrence word extraction unit 9. First, a labeled sentence group is acquired (step S101), and the presence or absence of a desired label or a dissatisfied label is determined (step S102). In step S102, when only one of the desired label or the dissatisfied label exists, the process proceeds to step S104. When both the desired label or the dissatisfied label exists, association between the desired / unsatisfied is performed (step S103). Next, a co-occurrence word extraction process is performed (step S104), and a co-occurrence word is acquired (step S105).

  The causal word candidate extraction unit 11 refers to the probability table created by the table creation unit 3 based on the co-occurrence words extracted by the co-occurrence word extraction unit 9, and extracts all corresponding cause candidates. For example, if there is a co-occurrence word such as “room temperature, humidity, air conditioner, patience” as described above, in the above table [probability table example 1], “no air conditioner” on the third line and “ The air conditioner is broken. " The reference destination of the probability table can be appropriately changed according to the output result of the subjective detection unit 20 for the input text group 4.

  For example, if the subjective detection unit 20 makes a request or dissatisfaction, the table DFC including three elements such as the above table [probability table example 1] is referred to with priority, and the other [probability table example 2]. ] And [Probability Table Example 3] can be referred to as the second and third candidates. At this time, the second candidate may be the table DC or the table FC. In the case of dissatisfaction only, a table FC including two elements of dissatisfaction and cause such as [Probability table example 3] is referred to preferentially, and three elements such as [Probability table example 1] are ignored in a way that the request is ignored. It is also possible to refer to a table including as a second candidate. The following table shows an example of priorities of reference destinations.

Ｎ_Ｄ：主観検出部により検出されたＤの数
Ｄ_Ｆ：主観検出部により検出されたＦの数

N _D : Number of D detected by the subjective detection unit D _F : Number of _F detected by the subjective detection unit

  FIG. 4 is a flowchart showing the operation of the causal word candidate extraction unit 11. First, a labeled sentence group and co-occurrence words are input (step S201). Next, it is determined whether or not there are desired labels and dissatisfied labels (step S202). In step S202, if both the desired label and the dissatisfied label exist, the table DFC is referred to (step S203), and the presence / absence of the second candidate is determined (step S204). In step S204, if there is no second candidate, the process proceeds to step S208. On the other hand, if there is a second candidate, the table DC is referred to (step S205), and the presence / absence of the third candidate is determined (step S206). In step S206, when there is no third candidate, the process proceeds to step S208. On the other hand, when there is a third candidate, the table FC is referred to (step S207), and a cause word candidate is extracted (step S208).

  Next, in step S202, when there is only a dissatisfied label, the table FC is referred to (step S209), and the presence / absence of the second candidate is determined (step S210). If there is no second candidate, the process proceeds to step S208. If there is a second candidate, the table DFC is referenced (step S211), and the process proceeds to step S208.

  Next, in step S202, when there is only a desired label, the table DC is referred to (step S212), and the presence / absence of the second candidate is determined (step S213). If there is no second candidate, the process proceeds to step S208. If there is a second candidate, the table DFC is referenced (step S214), and the process proceeds to step S208.

  The cause determination unit 13 determines the cause 15 from the cause word candidates extracted by the cause word candidate extraction unit 11. The cause determination method refers to a cause word candidate table and determines the cause using a probability value. For example, it is possible to display a plurality of upper ranks or to display a higher one than a preset threshold value. For example, in the case of the following table [table example of cause word candidates], C1, C2, and C3 are preferentially selected using probability values.

  In the cause word candidate table, when there are those with the same probability, the priority in the cause determination is that the demand D detected by the request determination unit 5 and the dissatisfion F detected by the dissatisfaction determination unit 7 among the input sentences. On the other hand, processing such as giving priority to the cause C that minimizes the phrase interval can be applied. For example, in the above table [cause word candidate table example], C1, C2, and C3 have the same probability, but from the next table [table example for determining the phrase interval], C1 is t3 and C2 is t7. , C3 is located at t50, but C1 and C2 having a small phrase interval are preferentially selected with respect to D of target t100 and F of t5.

  FIG. 5 is a flowchart showing the operation of the cause detection unit 30. Here, a case where a plurality of causes are specified will be described. First, a target probability table is acquired (step S301), and then it is determined whether the probability value in the probability table is equal to or greater than a threshold value (step S302). If the probability value is not equal to or greater than the threshold value, the process ends. If the probability value is equal to or greater than the threshold value, the cause having the probability value is specified (step S303), and the process ends.

  FIG. 6 is a flowchart showing the operation of the cause detection unit. Here, an example is shown in which the cause is narrowed down to one when the same probability exists. First, a target probability table is acquired (step S401), and then it is determined whether the probability value in the probability table is equal to or greater than a threshold value (step S402). If the probability value is not greater than or equal to the threshold value, the process ends. If the probability value is greater than or equal to the threshold value, it is determined whether or not the same probability exists (step S403). In step S403, if the same probability does not exist, the probability values are compared (step S405), the larger one is extracted, the cause is identified (step S406), and the process ends. In step S403, if the same probability exists, the sentence intervals are compared (step S404), the cause of the minimum sentence interval is specified (step S406), and the process ends.

  In addition, when the thing of the same space | interval exists, priority before C of D and F can also be given priority. This is due to the fact that sentences are generally written in the order of causes and results. For example, in the above [Example of Table for Determining Phrase Interval], C1 is selected with priority given to the positional relationship between C1 and C2. In addition, a cause label can be given to a sentence including a cause word extracted as a cause word candidate.

  As described above, according to the present embodiment, it is possible to automatically estimate the cause of expression of a word group indicating a request or dissatisfaction. In addition, since text is a target, it is not limited to products and systems, and it is possible to estimate the cause of expression of a word group indicating positive or negative subjectivity regarding human relationships or human life events.

DESCRIPTION OF SYMBOLS 1 Learning database 3 Table preparation part 4 Text group 5 Desire determination part 7 Dissatisfaction determination part 9 Co-occurrence word extraction part 10 Preprocessing part 11 Cause word candidate extraction part 13 Cause determination part 15 Cause 20 Subjective detection part 30 Cause detection part

Claims (5)

A cause estimation device for estimating the cause of at least one of a word group indicating a positive subjectivity or a word group indicating a negative subjectivity in a sentence,
A document classification process is performed on the input text data, and a positive label indicating that the classified word group is a positive subjective word group or a negative subjective word group. A subjective detection unit that gives a negative label indicating
Extracting co-occurrence words co-occurring with the group of words given the positive label or the group of negative labels from the sentence, the group of words showing positive or negative subjectivity, the word showing the subjectivity Extracting a word group indicating an expression cause including the extracted co-occurrence word from a probability table composed of a word group indicating the expression cause of the group and a numerical value indicating an expression probability of the word group indicating the expression cause; A cause detection unit comprising: a cause detection unit that selects a word group indicating the cause of occurrence as the cause word group in descending order of expression probability.   Learn words that have been collected in advance and classified by the document classification process and given a positive label indicating that it is a word group indicating positive subjectivity or a negative label indicating that it is a word group indicating negative subjectivity. The cause estimation apparatus according to claim 1, further comprising a pre-processing unit that stores the probability table based on a word group stored in a database and stored in the learning database.   When the cause detection unit has the same probability of expression of the selected plurality of causal word groups, among the causal word groups, from the word group to which a positive label is given or from the word group to which a negative label is given 3. The cause estimation apparatus according to claim 1, wherein a cause word group having the smallest phrase interval is selected. A program that estimates the cause of at least one of a word group showing positive subjectivity or a word group showing negative subjectivity in a sentence,
A document classification process is performed on the input text data, and a positive label indicating that the classified word group is a positive subjective word group or a negative subjective word group. A process of assigning a negative label indicating
A process of extracting a co-occurrence word that co-occurs with the group of words given the positive label or the group of words given the negative label from the sentence;
Extracted from a probability table composed of a word group indicating a positive or negative subjectivity, a word group indicating the cause of expression of the word group indicating the subjectivity, and a numerical value indicating the expression probability of the word group indicating the expression cause Processing to extract a word group indicating the cause of expression including co-occurrence words;
A program that causes a computer to execute a series of processes of selecting a word group indicating the extracted expression cause in descending order of expression probability and making it a cause word group. A cause estimation method for estimating a cause of occurrence of at least one of a word group indicating a positive subjectivity or a word group indicating a negative subjectivity in a sentence,
In the subjectivity detection unit, document classification processing is performed on the input text data, and a positive label indicating a positive subjectivity or a negative subjectivity is indicated for each classified word group. Assigning a negative label indicating that the word group,
In the cause detection unit, extract a co-occurrence word co-occurring with the word group given the positive label or the word group given the negative label from the sentence, a word group showing positive or negative subjectivity, A word group indicating an expression cause including the extracted co-occurrence word from a probability table composed of a word group indicating an expression cause of the word group indicating the subjectivity and a numerical value indicating an expression probability of the word group indicating the expression cause Extracting the word group indicating the extracted cause of occurrence and selecting the word group in descending order of the probability of occurrence as the cause word group.

Images