JP2012256283A - Sensitivity analysis system and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To present a study sentence which expresses the described content of each sentence at the point of time in a straight-forward manner when an analysis is made from a time sequence point of view for the aggregate of sentences whose number is exponentially increasing and decreasing.SOLUTION: A sensibility analysis system 10 includes: a negative/positive determination part 14 which gives a negative/positive polarity tag to a sensibility term and gives a negative expression tag to a negative expression of each sentence and gives a negative/positive determination tag to the sensibility term and the negative expression based on the type of polarity tag and the presence or absence of the negative expression tag; an emotion extraction part 18 for associating an emotion type tag with an emotion expression of each sentence; a study rule storage part 24 for storing a plurality of study rules; and an analysis result estimation part 21 for generating a time sequence change graph by adding up the number of sentences every day based on date information of each sentence, applying each study rule to an aggregate of sentences within a prescribed period of interest including change points on the graph and the aggregate of the sentences within all the periods, extracting a template on the study rule which matches only the aggregate of the sentences within the period of interest as the study sentence relating to the change point and outputting it together with the time sequence change graph.

Description

  The present invention relates to a sensitivity analysis system and program, and in particular, automatically evaluates end users' subjective evaluations and opinions regarding specific products and services from electronic document data such as article data and questionnaire article data on an electronic bulletin board. Related to collective technology.

  Each company's subjective evaluation (impressions) on specific products and services (hereinafter “products, etc.”) is an extremely important guideline for improving current products and developing next-generation products. The results of questionnaires collected from end users are analyzed from various viewpoints, or an electronic bulletin board on the Internet is accessed to check the evaluation of their products.

Also, in order to improve the efficiency of such manual analysis work, there is a technique for extracting end-user evaluations for specific products by automatically analyzing digitized text (questionnaire information) using a computer. It has already been proposed.
For example, in Patent Document 1, after performing morphological processing and syntax analysis processing on a sentence including emotional expressions for a predetermined object, refer to the effect term dictionary in which a large number of emotional expressions are registered, There is disclosed a technique for extracting emotional expressions from sentences, totaling attributes (negative / positive, etc.) of each emotional expression, and outputting the results to the outside.
As a result, regarding the evaluation target “lipstick”, it is possible to present the distribution of negative evaluations and positive evaluations for each of a plurality of evaluation axes such as color, odor, comfort, and package.
JP2003-248861

  As described above, by using the conventional analysis technique, it is possible to automatically aggregate alternative evaluations such as end user's likes and dislikes and good or bad for the products to be evaluated.

Since time information such as writing date / time and update date / time is usually added to digitized document data, the number of document data collected and accumulated over a certain period It is expected that useful knowledge can be obtained by time-series analysis of the increase and decrease. For example, when there is a point in time when the number of posts on a bulletin board about a specific product suddenly increases, it is possible to understand what happened by grasping the characteristics of the description content before and after that It becomes.
However, the conventional analysis technology lacks the function of analyzing a set of documents from a time-series perspective, so even though we have a set of documents with time information, we have a temporal factor. There was a problem that only static analysis results were available.

  This invention has been devised in order to solve such a conventional problem, and analyzes a set of documents with temporal information from a time-series viewpoint, and the number of cases is rapidly increased. The purpose is to realize a technology that enables presentation of a comment (comment) that expresses the description content at the time of increase or decrease.

In order to achieve the above object, a sensitivity analysis system according to claim 1 stores a combination of a sensitivity term representing an affirmative / negative value judgment for an object and an attribute indicating whether it is affirmative / negative. Sensitivity dictionary storage means, morpheme processing is performed on a document file, a plurality of text sentences included in the document file are decomposed into morpheme units, and morphemes or combinations of morphemes included in each sentence It is determined whether or not the expression corresponds to the sensitivity term stored in the sensitivity dictionary storage means, and if so, indicates either the positive or negative attribute combined with the expression Consists of means for associating attribute tags, negative expression storage means for storing negative expressions used in constructing negative sentences, and morphemes or combinations of morphemes included in each sentence Determining whether or not the current expression corresponds to the negative expression stored in the negative expression storage means, and, if applicable, means for associating a negative expression tag indicating that the expression is a negative expression with each expression; Based on the combination pattern of the type of attribute tag associated with the sentence and the presence or absence of a negative expression tag, it is determined whether the sentence is a negative evaluation sentence or a positive evaluation sentence. At the same time, a means for assigning a positive determination tag to a positive evaluation sentence and a particle for constructing the main case in each sentence are searched, and the subject recognition that recognizes the independent word attached to the particle as the subject. Meaning storage means for storing a combination of a means, an intention expression expressing the intention of the speaker, and a type of the intention, and an intention type meaning “dissatisfaction” with respect to the sentence with the negative determination tag When you associate tags The sentence with the positive determination tag is associated with an intention type tag that means “popular”, and an expression composed of a morpheme or a combination of morphemes included in each sentence is stored in the intention expression dictionary storage unit. Means for associating an intention type tag indicating a corresponding type with the expression, a character string indicating a plurality of types of evaluation axes, and each evaluation axis. The evaluation axis similar expression storage means for storing the correspondence relationship with the similar expressions included in the text, and whether the subject recognized in each sentence corresponds to the similar expression stored in the evaluation axis similar expression storage means And, if applicable, means for associating a corresponding evaluation axis with a sentence, and for a set of sentences within a predetermined range, the sentence at predetermined time intervals based on time information associated with each sentence. The number of cases Means for generating time-series information having a correspondence relationship between the time interval and the number of cases, means for generating a time-series change graph based on the time-series information, and the time-series change graph based on the time-series information Consideration that stores multiple consideration rules consisting of combinations of means for identifying the above change points, application conditions defined by the number and ratio of sentences with each intention type tag, and a template for the consideration sentence Each of the above consideration rules is applied to the rule storage means and the set of sentences within the predetermined period of interest including the change point to determine whether or not the application condition is met, and the change point is included. There is a means for applying each of the above consideration rules to a set of sentences within the entire period and determining whether or not the application condition is matched, and a consideration rule that only matches the set of sentences within the above period of interest. Place Is characterized in that it includes a consideration sentence generating means for extracting the contents described in the consideration rule template as a consideration sentence regarding the change point, and a means for outputting the time series change graph and the consideration sentence. .
As the “attribute”, for example, the polarity of “+” or “−” is applicable, but “positive” or “negative”, “P”, “N”, or the like may be used (the same applies hereinafter).
The “intention type tag meaning“ dissatisfied ”” includes “disappointment”, “negative”, “complaint”, etc. having negative meanings in addition to “dissatisfied”.
The “intention type tag meaning“ favored ”” corresponds to “satisfied”, “positive”, “affirmation”, etc. having a positive meaning in addition to “popular”.
In addition to displaying on the display, the above "output" means printing out through a printer, generating a file in a predetermined format and storing it in a storage means, generating a screen (such as an html file), and a client It is a concept including transmitting to a terminal.

  The sensibility analysis system according to claim 2 is the system according to claim 1, and further, an opinion insertion column is provided in advance in the template, and each sentence is divided into a common evaluation axis, intention type and Classifying into similar sentence groups that are groups of similar sentences with sensibility terms, and providing means for recognizing a similar sentence with the smallest number of characters in each similar sentence group as a representative opinion. Among the opinions, the representative opinion having the largest number of similar sentences included in the corresponding similar sentence group is inserted into the above-described opinion insertion column.

  The sensibility analysis system according to claim 3 is the system according to claim 1 or 2, further comprising a ratio insertion field for inserting a ratio of a sentence to which a specific intention type tag has been assigned in advance in the template. Provided, and includes means for calculating a ratio of a sentence to which the intention type tag is assigned, and the consideration sentence generation means inserts the ratio into a ratio insertion field of the template.

  The sensibility analysis system according to claim 4 is the system according to claims 1 to 3, and further includes a time series change graph for a set of sentences associated with a specified evaluation axis or intention type tag. Generation processing, change point identification processing, and consideration sentence generation processing are executed.

  The sensibility analysis system according to claim 5 is the system according to claims 1 to 4, and further, for the set of sentences belonging to the specified period, the time-series change graph generation process, the change point Specific processing and generation processing of a consideration sentence are executed.

  The sensibility analysis system according to claim 6 is the system according to claims 1 to 5, and further applies one time series analysis model specified among a plurality of time series analysis models prepared in advance. By doing so, the above-mentioned change point specifying process is executed.

  The sensibility analysis system according to claim 7 is the system according to claims 1 to 6, and further, a priority is set for each of the consideration rules, and the consideration sentence generation unit When there are a plurality of consideration rules that match only a set of sentences within a period, a consideration sentence is generated based on a template related to a consideration rule with a high priority.

  The sensitivity analysis program according to claim 8 stores a combination of a sensitivity term representing an affirmative / negative value judgment for an object and an attribute indicating whether the computer is affirmative / negative. Means for performing morpheme processing on a document file and decomposing a plurality of text sentences included in the document file into morpheme units; an expression comprising a morpheme or a combination of morphemes included in each sentence is the Kansei dictionary Means for associating attribute tags indicating any one of the above-mentioned positive / negative attributes combined with the expression, whether or not it corresponds to a sensitivity term stored in the storage means; Negative expression storage means for storing negative expressions used when composing sentences, expressions composed of morphemes or combinations of morphemes included in each sentence are Hands that determine whether or not the negative expression stored in the current storage means corresponds, and if so, a hand that associates a negative expression tag indicating that it is a negative expression to the expression, an attribute associated with each sentence Based on the combination pattern of tag type and presence / absence of negative expression tag, it is determined whether the sentence is a negative evaluation sentence or a positive evaluation sentence. Means to attach positive judgment tags to sentences, search for particles that constitute the main character in each sentence, and subject recognition means that recognizes independent words attached with the particles as the subject; An intention expression storage means for storing a combination of an intention expression indicating the type of the intention and a type of the intention, an intention type tag that means “dissatisfied” is associated with the sentence with the negative determination tag, and the positive determination tag With Whether or not an expression composed of a morpheme or a combination of morphemes included in each sentence corresponds to the intention expression stored in the intention expression dictionary storage means. A means for associating an intention type tag indicating a corresponding type with the expression, if applicable, a character string indicating a plurality of types of evaluation axes, and a correspondence between similar expressions included in each evaluation axis The evaluation axis similar expression storage means for storing the relationship, whether the subject recognized in each sentence corresponds to the similar expression stored in the evaluation axis similar expression storage means, and if so, A means for associating a corresponding evaluation axis with a sentence, and for a set of sentences within a predetermined range, based on time information associated with each sentence, the number of sentences is totaled for each predetermined time interval, Have a corresponding relationship Means for generating time series information, means for generating a time series change graph based on the time series information, means for identifying a change point on the time series change graph based on the time series information, each intention type Consideration rule storage means for storing a plurality of consideration rules composed of combinations of application conditions defined by the number and ratio of sentences to which tags are added and a consideration sentence template, a predetermined period of interest including the change point Each of the above consideration rules is applied to the set of sentences in the list to determine whether or not the application condition is met, and each of the above considerations is set to the set of sentences within the entire period including the above change point. Means for applying a rule and determining whether or not it matches the application condition, and if there is a consideration rule that only matches a set of sentences within the period of interest, describe it in the template of the consideration rule Discussion sentence generation means for extracting the contents as discussed statements about the change point, it is characterized in that to function as means for outputting the time-series variation graph and discussion statement.

  In the sensibility analysis system according to claim 1 and the sensibility analysis program according to claim 8, a predetermined time interval (for example, “day” or “week”, Generate time-series information that has a corresponding relationship between the number of "months" and the number of cases, generate a time-series change graph based on this time-series information, and change the number of cases dramatically on this time-series graph Analyst in charge because it has a mechanism to identify points and automatically output a consideration sentence according to the number and ratio of intention types set in each sentence belonging to a predetermined period of interest including this change point. Makes it possible to accurately grasp the tendency of the end user's writing before and after the change point.

  The sensitivity analysis system according to claim 2 has a mechanism for outputting a consideration sentence in which the most representative opinion in each sentence belonging to the period of interest is output, so that the analyst is more persuasive. It is possible to present a strong thought sentence.

  The Kansei analysis system according to claim 3 is provided with a mechanism for outputting a consideration sentence in which a ratio of sentences to which a specific intention type tag is given in the period of interest is output. On the other hand, it is possible to present a more persuasive consideration sentence.

  According to the sensibility analysis system described in claim 4, it is possible to output different time series change graphs and consideration sentences for each designated evaluation axis unit or intention type unit, and detailed consideration sentences for the analyst. Can be presented.

  According to the sensibility analysis system described in claim 5, it is possible to output different time-series change graphs and consideration sentences for each designated period, and to present detailed consideration sentences to the person in charge of analysis. It becomes possible.

  According to the sensibility analysis system described in claim 6, since the change point is specified based on the specified time series analysis model, the person in charge of analysis performs the optimal analysis according to the characteristics of the analysis target data. The result can be obtained.

  According to the sensibility analysis system described in claim 7, since a consideration sentence is generated based on a consideration rule template having the highest priority when a plurality of consideration rules match, an analyst Makes it possible to obtain a consideration sentence most suitable for the period of interest.

  FIG. 1 is a block diagram showing the overall configuration of a sensibility analysis system 10 according to the present invention. A document registration unit 11, a document storage unit 12, a document analysis unit 13, a negative / positive determination unit 14, and a sensibility dictionary. Storage unit 15, negative expression storage unit 16, subject recognition unit 17, intention extraction unit 18, intention expression storage unit 19, analysis result storage unit 20, analysis result evaluation unit 21, evaluation axis similar expression storage Unit 22, time series analysis unit 23, consideration rule storage unit 24, and screen generation unit 25.

The document registration unit 11, the document analysis unit 13, the negative / positive determination unit 14, the subject recognition unit 17, the intention extraction unit 18, the analysis result evaluation unit 21, the time series analysis unit 23, and the screen generation unit 25 are The CPU is realized by executing necessary processes according to the OS and application programs.
In addition, the document storage unit 12, the sensitivity dictionary storage unit 15, the negative expression storage unit 16, the intention expression storage unit 19, the analysis result storage unit 20, the evaluation axis similar expression storage unit 22, and the consideration rule storage unit 24 are server computers. In the external storage device.

This sensibility analysis system 10 is connected to a Web server 26 via a network, and the Web server 26 is connected to a plurality of client terminals 28 via the Internet 27.
Each client terminal 28 includes a PC or the like equipped with an OS and a web browser.

Next, a method of using the sensitivity analysis system 10 will be described.
First, the user accesses the Web server 26 from the client terminal 28, inputs an ID and password, logs in, and then selects “Kansei analysis service” from the service menu.
As a result, an analysis condition designation screen is transmitted from the Web server 26 to the client terminal 28.

  FIG. 2 shows an example of the analysis condition designation screen 40, which includes an analysis target file designation column 41 and an evaluation axis designation column 42.

First, the analysis target file designation column 41 is a column for designating a document file prepared in advance by the user. That is, when the user clicks the browse button 43, a file selection window showing the drive structure of the client terminal 28 is displayed on the Web browser. On the other hand, the user clicks and expands a specific drive name or a folder name under it to specify a specific document file.
In the figure, a document file “03-11 questionnaire result.csv” is designated.

In this document file, a plurality of digitized documents collected from customers of the user “ABC Tourist Hotel” are stored.
Each document corresponds to a questionnaire response document described by an individual customer, and each document includes a plurality of sentences. In addition, each document is added with time information such as reply date and time, update date and time, and attribute information of the respondent (gender, age, occupation, prefecture, postal code, etc.).

The evaluation axis designation column 42 is a column for designating analysis items for each sentence included in the analysis target file, and lists a plurality of standard evaluation axes set for each industry to which the user belongs.
In the figure, since the industry code associated with the user ID is “Hotel / Ryokan industry”, “Price”, “Cooking”, “Room”, “Bath”, “Location”, “Banquet”, “Equipment” Etc. are displayed.
On the other hand, the user can exclude the evaluation axis from the analysis items by unchecking the check box of the unnecessary evaluation axis.

When the user who has completed the above settings clicks the analysis start button 44, a sensitivity analysis request is transmitted from the client terminal 28 to the Web server 26.
At this time, the document file data designated by the user and the evaluation axis identification code set by the user are transmitted to the Web server 26.
Receiving this, the Web server 26 transmits data such as the document file and the business type code uploaded from the client terminal 28 to the sensibility analysis system 10 and requests analysis.

Hereinafter, the processing procedure in the sensitivity analysis system 10 will be described with reference to the flowchart of FIG.
First, the document registration unit 11 of the sensitivity analysis system 10 stores the document file transmitted from the web server 26 in the document storage unit 12 (S10). This document file is associated with the evaluation axis identification code set by the user and the user's business type code.

  Next, the document analysis unit 13 retrieves the document file stored in the document storage unit 12, assigns a document identification code to each document in the document file, and simultaneously assigns a sentence identification code to each sentence included in each document. (S11).

Next, the document analysis unit 13 performs a morphological analysis process for each sentence in the document file (S12).
Here, “morpheme analysis” refers to a process of decomposing a sentence written in a natural language into morphemes, which are the smallest meaningful language units, and specifying each part of speech. This morpheme analysis process is a known technique. For example, the following free software can be used as the morpheme analysis engine.
■ MeCab (http://mecab.sourceforge.net/)
■ ChaSen (http://chasen.naist.jp/hiki/ChaSen/)

Next, the document analysis unit 13 performs syntax analysis processing for each sentence decomposed into morphemes (S14).
Here, “syntactic analysis” refers to a process of specifying a dependency structure between clauses included in each sentence. This syntax analysis itself is also a known technique. For example, the following free software can be used as the syntax analysis engine.
■ KNP (http://nlp.kuee.kyoto-u.ac.jp/nl-resource/knp.html)
■ CaboCha (http://chasen.org/~taku/software/cabocha/)

When the processing by the document analysis unit 13 is completed, negative / positive determination processing by the negative / positive determination unit 14 is executed (S16).
Hereinafter, a specific procedure of the negative / positive determination process will be described with reference to the flowchart of FIG.

  First, the negative / positive determination unit 14 refers to various sensitivity dictionaries stored in the sensitivity dictionary storage unit 15, and corresponds to the sensitivity term registered in the sensitivity dictionary in the morpheme or combination of morphemes constituting each sentence. Search for what to do (S16-01).

  As shown in FIG. 5, the sensitivity dictionary storage unit 15 stores a general-purpose dictionary group 50 and a plurality of domain-specific dictionary groups 55. The general-purpose dictionary group 50 includes a word dictionary 51, a binary relation dictionary 52, an idiomatic phrase dictionary 53, and a mimetic word dictionary 54. Further, each domain dictionary group 55 includes a word dictionary 56, a binary relation dictionary 57, and a mimetic word dictionary 58, respectively.

FIG. 6 exemplifies registered data of the word dictionary 51 belonging to the general-purpose dictionary group 50, and stores a large number of records having data items of “Kansei term”, “Reading” and “Polarity”. .
Here, the “Kansei term” data item stores nouns, adjectives, etc. that represent human subjective value judgments regarding positive / negative (good / bad) things such as “delicious”, “delicious”, “bad”, etc. Has been. Further, in the polarity data item, a sign “+” is set when the meaning content of the sensitivity term is positive, and a sign “−” is set when it is negative.

In the binary relation dictionary 52 belonging to the general-purpose dictionary group 50, a number of records having data items of “sensibility term”, “reading”, and “polarity” are registered, although illustration is omitted.
Here, the “binary relationship” means a word that causes a specific value judgment for the first time when two words are combined through a particle, for example, “high price”.
In other words, the adjective “high” generally has a positive meaning, but it depends on the subject, such as “high price”, “high risk”, and “high corrosiveness”. May have negative implications. Or, conversely, the adjective “low” is generally conceived of a negative meaning, but like “low risk”, “low toxicity”, “low illegality” The subject may have positive implications.
For this reason, the binary relation dictionary 52 is given a polarity of “+” or “−” for a word that expresses a specific value judgment by a combination of two words.

In the common phrase dictionary 53 belonging to the general-purpose dictionary group 50, a number of records having data items of “sensibility term”, “reading”, and “polarity” are registered, although illustration is omitted.
Specifically, “Nenbutsu in the ears of the horse (sensitivity term) / − (polarity)”, “It rains and solidifies (sensitivity term) / + (polarity)”, “Hit stomachache (sensitivity term) / − (polarity) ”,“ Do not return to the basin (sensitivity term) / − (polarity) ”,“ Honor of the indigo (sensitivity term) / + (polarity) ”, and idioms with value judgments such as proverbs and maxims Is filled in the data item of “Kansei term”, and “+” is used when the meaning of each idiom is positive, and “−” is “polarity” when it is negative. The data items are filled.

Although not shown in the mimetic word dictionary 54 belonging to the general-purpose dictionary group 50, a large number of records having data items of “sensitive term”, “reading”, and “polarity” are registered.
Specifically, "Waku Waku (Sensitive term) / + (Polarity)", "Hot (Sensitive term) / + (Polarity)", "Zawazawa (Sensitive term) /-(Polarity)", "Gitto (Sensitive term) Like “/-(polarity)”, a mimetic word with a value judgment is filled in the data item of “Kansei term”, and “+” is negative when the meaning of each mimetic word is positive. If it is a negative one, “-” is filled in the “polarity” data item.

Although not shown, the word dictionary 56, the binary relation dictionary 57, and the mimetic word dictionary 58 that belong to each domain dictionary group 55 are provided with data items of “sense term”, “reading”, and “polarity”. A lot of records are registered.
However, each domain dictionary group is prepared for each specific category (industry, product field, service field), and sensitivity terms and polarities specific to the category are registered.

  When a sensitivity term matching a specific morpheme or combination of morphemes exists in the sensitivity dictionary, the negative / positive determination unit 14 adds a tag (<+> or <->) indicating the polarity of the sensitivity term. It is given to the expression (S16-02).

In this case, for example, if there is an expression “price is a little high” in the sentence, and there is a registered example of “high price (sensitivity term) / − (polarity)” in the binary relational dictionary 52, the negative / positive By referring to the processing result of the syntax analysis, the determination unit 14 can add a <-> tag to “a little expensive”.
That is, since the analysis data indicating that the “price” clause in the sentence relates to the “high” clause is generated by the parsing process by the document analysis unit 13, the negative / positive determination unit 14 performs this analysis. Based on the data, the polarity of “high price” in the binary relational dictionary 52 can be applied to the expression “price is a little high”.

  Next, the negative / positive determination unit 14 refers to the negative expression storage unit 16 to search for an expression that matches the negative expression pattern registered in the negative expression storage unit 16 in the morpheme or combination of morphemes constituting each sentence. (S16-03).

Although not shown in the figure, the negative expression storage unit 16 stores a large number of negative expression patterns used when creating negative sentences. For example, "I don't think", "I don't think", "I can't say", "I can't say", "I can't say", "I don't", "I don't feel", "I don't feel", "I don't feel" No ”,“ I don't think ”, etc.
If a negative expression pattern that matches the morpheme or combination of morphemes in the sentence is registered in the negative expression storage unit 16, the negative / positive determination unit 14 performs a negative expression for the expression (morpheme or combination of morphemes). A tag (<Negation>) is assigned (S16-04).

Next, the negative / positive determination unit 14 determines whether the entire sentence is negative or positive by applying a unique grammar rule to the combination of the polarity tag and the negative expression tag given to each sentence ( S16-05).
When a negative determination result is obtained, the negative / positive determination unit 14 adds a negative determination tag of “<N> (Negative)” to the negative evaluation sentence, and determines “positive”. If a result is obtained, a positive determination tag of “<P> (Positive)” is assigned to the positive evaluation sentence (S16-06).

  For example, as shown in FIG. 7A, in the case of the sentence “I thought the dish was delicious,” the <+> polarity tag was attached to the “delicious” morpheme, and the <deny> tag Is not assigned, the negative / positive determination unit 14 recognizes the entire sentence as “positive” and associates the positive determination tag of <P>.

  On the other hand, as shown in FIG. 7B, in the case of the sentence “I didn't think the dish was delicious,” the “+” polarity tag was attached to the “delicious” morpheme. Since the <No> tag is attached to the “I don't think” part that is located behind it, the negative / positive determination unit 14 recognizes the entire sentence as “Negative” and the <N> negative Associate a judgment tag.

  In addition, as shown in FIG. 7 (c), in the case of the sentence “Cooking was bad”, the <-> polarity is assigned to the “morphic” morpheme, and the <negative> tag is assigned. Therefore, the negative / positive determination unit 14 recognizes the entire sentence as “negative” and associates the negative determination tag of <N>.

  On the other hand, as shown in FIG. 7 (d), in the case of the sentence “Cooking was not bad”, the <-> polarity was given to the “poor” morpheme. Since the <Negated> tag is assigned to the “no” part located at the rear, the negative / positive determining unit 14 recognizes the entire sentence as “positive” and associates the <P> positive determining tag.

When the negative / positive determination processing by the negative / positive determination unit 14 is completed, the subject recognition processing by the subject recognition unit 17 is executed (S18 in FIG. 3).
Here, the “subject recognition process” means a process for specifying a negative / positive evaluation target.

  In order to identify this subject, the subject recognition unit 17 first searches for the case elements "~ ha", "~ ga", and "~ mo" in the sentence, and uses the independent words (major case) to which these particles are attached as the subject. And attach a <subject> tag to that effect. In the case of each example sentence shown in FIG. 7, “cooking” corresponds to the subject.

  In some sentences, the main character may be omitted. For example, as shown in FIG. 8, the main character is omitted in “(2) It was musty” following the sentence “(1) The room was too hot and hard to sleep”. In such a case, the subject recognition unit 17 passes the “room” that is the subject of the previous sentence to the subsequent sentence, and recognizes “room” as the subject of the sentence (2).

Inheritance of this subject is not limited between successive sentences, but is inherited one after another when sentences without a main character continue.
In FIG. 8, the “room” in the sentence (1) is inherited as the main character in the sentence “(3) It was a salvation at least for quiet.” Then, when the sentence (4) having a new main character “Cooking” appears, the inheritance as the theme of “room” is stopped.

If there is no main character in the sentence and there is no preceding sentence with a main character that should be inherited, the subject recognition unit 17 searches for whether or not a combination modifier is included in the sentence. If it is found, the modified morpheme is recognized as the subject.
For example, in the case of the sentence “good scent.”, The main character is omitted, but there is a combined modifier “good (<+>)” that modifies the noun (body name) “scent”. The subject recognition unit 17 recognizes “scent” as the subject of the sentence.

When the subject certification process by the subject certification unit 17 is completed, the intention extraction process by the intention extraction unit 18 is executed (S20 in FIG. 3).
Here, “intent” refers to the dualistic value judgment of “good / bad” by a speaker (writer) on an individual subject, as well as the purpose or speculation behind the value judgment, or the good or bad This is a concept that includes an assertion or message to the reader of a speaker who is far from the judgment of the value of, and is "popular", "dissatisfied", "intention", "request", "question", "unexpected", "interesting" , “No interest”, etc.

  First, the intention extraction unit 18 adds a <Popular> intention type tag to a sentence with a <P> positive determination tag, and <N> for a sentence with a <N> negative determination tag. A dissatisfaction> intention type tag is assigned.

Next, the intention extraction unit 18 refers to the intention expression storage unit 19 and executes matching with expressions (morpheme or combination of morphemes) in each sentence to extract intention types other than <popular> and <dissatisfaction>. .
That is, as shown in FIG. 9, each record stored in the intention expression storage unit 19 includes data items of “intention expression” and “intention type”. Therefore, the intention extraction unit 18 searches each sentence for an expression that matches the intention expression, and if a corresponding expression is found, associates the tag of the corresponding intention type with the expression.

For example, as shown in FIG. 10 (a), when the sentence “Since dinner is poor, please increase the number of items a little more.” Is given, the intention extraction unit 18 displays the character string “please” as the intention expression storage unit. After detecting that it exists in 19, an intention type tag of <request> corresponding to “request” which is an intention type of “please” is associated with the expression.
Incidentally, since the negative determination tag of <N> indicating that the whole sentence is “negative” is assigned to this sentence, the <unsatisfied> intention type tag is redundantly assigned.

Also, as shown in FIG. 10B, when a sentence “I was surprised that the price is conscientious” is given, the intention extraction unit 18 sends a character string “surprise” to the intention expression storage unit 19. After detecting the presence, an <unexpected> intention type tag corresponding to “surprise” is assigned to the expression.
For this sentence, the <P> positive determination tag indicating that the entire sentence is “positive” is assigned, and therefore, the <popular> intention type tag is also assigned.

Further, as shown in FIG. 10 (c), when the sentence “I want to drive this time because the access is bad” is given, the intention extraction unit 18 stores the character string “I want” in the intention expression storage unit 19. After detecting that it exists, an intention type tag of <intention> corresponding to “I want to do” is given to the sentence.
For this sentence, since the negative determination tag of <N> indicating that the whole sentence is “negative” is assigned, the intention type tag of <dissatisfaction> is also assigned. .

Furthermore, as shown in FIG. 10 (d), when a sentence “Is it possible to ask us?” Is given, the intention extraction unit 18 stores the intention expression memory string. After detecting that it exists in the part 19, an intention type tag of <question> corresponding to “Is it?” Is given to the sentence.
In the case of this sentence, no tag indicating the negative / positive determination result is added, and therefore no other intention type tag is assigned.

When the intention extraction process is completed for all sentences included in the analysis target file, the intention extraction unit 18 stores the analysis results thus far in the analysis result storage unit 20 (S22 in FIG. 3).
Here, “analysis result” corresponds to analysis result data in which a subject tag, a polarity tag, a negative expression tag, a negative / positive determination tag, and an intention type tag are assigned to each sentence.

Next, the analysis result evaluation unit 21 generates an analysis result screen (S24 in FIG. 3).
This analysis result screen is transmitted to the client terminal 28 via the Web server 26 (S26 in FIG. 3).

FIG. 11 shows an example of the analysis result screen 60 displayed on the Web browser of the client terminal 28, and a negative / positive distribution graph 61 showing a negative / positive analysis result is displayed in the main area.
Further, the sidebar is provided with an intention classification chart 62 in which the number of intention extraction results is displayed in a tree shape.

First, in the negative / positive distribution graph 61, positive, negative, and neutral percentages are shown by band graphs for each evaluation axis of total, price, food, room, bath, and location.
By referring to this graph 61, for example, it can be read that the customer is generally satisfied with the price because the ratio of the document with the positive determination is larger than the ratio of the negative determination document.
On the other hand, regarding the “room”, since the proportion of documents with negative judgments is overwhelmingly large, it can be understood that many customers are buying a bad experience.
In addition, with respect to “total” derived from the total value of each evaluation axis, the ratio of positive and negative is antagonizing, so the overall result can be recognized as “no good or bad”.

  The way of showing the negative / positive analysis result is not limited to the above-mentioned band graph, and the negative / positive ratio and number can also be displayed by a radar chart or a pie chart.

  In the intention classification chart 62, the total number of corresponding sentences is displayed in parentheses for each evaluation axis of price, cooking, room, bath, and location. For example, 282 sentences are associated with “Price” and 152 sentences are associated with “Cooking”.

When the user clicks each evaluation axis, popular intentions, dissatisfactions, intentions, requests, questions, unexpected, interested, uninteresting, and unknown intention types are developed.
In addition, the total number of sentences with the corresponding intention type tag is displayed in parentheses for each intention type. For example, it is shown that there are 74 sentences for “popular” under “price” and 121 sentences for “dissatisfied”.
“Unknown” indicates the number of sentences that are related to the relevant evaluation axis but are not assigned any intention type tag.

Here, when the user clicks on any intention type, a representative opinion list in which a plurality of representative opinions included in the intention type are listed is developed.
For example, when the user clicks an intention type “popular” under the evaluation axis “price”, as shown in FIG. 12, “low price is attractive. (17)”, “price is reasonable” in the sidebar. The representative opinion list 63 listing representative opinions such as “12)” and “Good for cheap” (15) is displayed.
The numbers in parentheses attached to the end of each representative opinion indicate the total number of approximate opinions included in the representative opinion.

On the other hand, when the user selects any representative opinion with the mouse pointer, a representative opinion breakdown list 64 listing the representative opinions and their approximate opinions is displayed in the main area.
The list 64 includes a document number and a display item for an analysis target sentence. In the display item of the analysis target sentence, a sentence that is the same as or approximate to the representative opinion is displayed in each sentence. The relationship between the representative opinion and the approximate opinion will be described in detail later.
When the user clicks the “select” button 65 of an arbitrary sentence, although not shown, another window is activated on the display, and the entire document including the representative opinion or approximate opinion is displayed. Various information (polarity tag, negative expression tag, negative / positive determination tag, theme tag, evaluation axis, intention type tag, attribute information, time information, etc.) is given to these documents.

Next, a processing procedure related to generation of the analysis result screen 60 will be described with reference to the flowchart of FIG.
First, the analysis result evaluation unit 21 acquires “price”, “dish”, “room”, “bath”, and “location”, which are evaluation axes associated with the document file (S24-01).
Next, the analysis result evaluation unit 21 refers to the industry-specific evaluation axis similar expression dictionary stored in the evaluation axis similar expression storage unit 22, and specifies the similar expression (expanded word) of each evaluation axis (S24-02). .

FIG. 14 shows a specific example of an evaluation axis similar expression dictionary for the hotel / ryokan industry stored in the evaluation axis similar expression storage unit 22.
For example, for the evaluation axis of “price”, synonyms and related words such as “fee”, “price”, “accommodation fee”, “accommodation fee”, and the like are stored as similar expressions in addition to “price”. These similar expressions are compiled for each industry to which the user belongs while referring to a general synonym dictionary or the like and considering the characteristics of each industry.
For this reason, on the evaluation axis of “price”, in addition to general names such as “price”, “price”, and “price”, “hotel charges” and “price” are unique to the hotel and ryokan industry. Names are listed.
In addition, when providing services to users belonging to the food and beverage industry, similar expressions such as “food and beverage charges” and “beverage charges” are listed instead of “hotel charges” and “hotel charges”. The evaluation axis similarity expression dictionary is applied.

Next, the analysis result evaluation unit 21 matches the character string recognized as the “subject” of each sentence included in the document file with the above similar expression, and associates the corresponding evaluation axis with the corresponding sentence (S24-). 03).
For example, in the case of the sentence “Dinner is poor, please increase the number of items a little more” in FIG. 10 (a), the character string “dinner” which is the subject is registered as a similar expression of the evaluation axis “dish”. Therefore, the analysis result evaluation unit 21 associates the evaluation axis identification code of “cooking” with the sentence.

  Also, in the case of the sentence “I was surprised that the price is conscientious” in FIG. 10 (b), since the character string of “the price” that is the subject is registered as a similar expression of the evaluation axis “price”, The analysis result evaluation unit 21 associates the “price” evaluation axis identification code with the sentence.

  Also, in the case of the sentence “Access is bad and I want to drive this time” in FIG. 10 (c), since the character string “access” that is the subject is registered as a similar expression of the evaluation axis “location”, The analysis result evaluation unit 21 associates the evaluation axis identification code of “location” with the sentence.

  On the other hand, in the case of the sentence “Wouldn't it be my request?” In Fig. 10 (d), the character string “How to request” that is the subject is registered as a similar expression on any evaluation axis. Therefore, the analysis result evaluation unit 21 associates an evaluation axis identification code of “unknown”.

  Next, the analysis result evaluation unit 21 extracts a sentence to which identification codes of a plurality of evaluation axes selected by the user are added from the document file, and then sets negative / positive determination tags <P> and <N set in each sentence. The number of> is counted (S24-04).

For example, there are a total of 80 sentences with “price” evaluation axis identification codes, of which 45 are tagged with the <P> tag, and those with the <N> tag. When there are five sentences to which none of <P> and <N> is assigned, the analysis result evaluation unit 21 generates the following total result.
[Evaluation axis: Price]
<P> (Positive): 45 <N> (Negative): 30
(Neutral): 5

In addition, there are 60 sentences with “Cooking” identification code in total, 18 of which are tagged with <P> and 25 of which are tagged with <N>. When there are 17 sentences to which neither <P> nor <N> is assigned, the analysis result evaluation unit 21 generates the following tabulation results.
[Evaluation axis: Cooking]
<P> (Positive): 18 <N> (Negative): 25
(Neutral): 17

Next, the analysis result evaluation unit 21 generates a band graph (negative / positive distribution graph 61) reflecting the above-described total result (S24-05).
At this time, the analysis result evaluation unit 21 calculates the “total number of positives” that is the total number of positives of each evaluation axis, and also calculates the “total number of negatives” that is the total number of negatives of each evaluation axis, and the neutral number of each evaluation axis. Calculate the “total neutral number”, and reflect the percentages in the “total” of the band graph.

  Next, the analysis result evaluation unit 21 aggregates the number of sentences for each evaluation axis and intention type assigned to each sentence, and generates an intention classification chart 76 (S24-06).

Next, the analysis result evaluation unit 21 extracts representative opinions from the document file in units of “evaluation axis × intention type” (S24-07).
For this purpose, the analysis result evaluation unit 21 first extracts sentences having the same emotional expression for each group of sentences having the same evaluation axis and intention type.
For example, as shown in FIG. 15 (a), when the sentences (1) to (6) that share the evaluation axis “price” and the intention type “dissatisfaction” are given, the analysis result evaluation unit 21 performs FIG. As shown in b), three sentences (1), (2), and (5) that share the emotional expression “high” are extracted.
Next, as shown in FIG. 15 (c), the analysis result evaluation unit 21 reads the sentence “(2) with the smallest number of characters among (1), (2), and (5)“ obviously too expensive ”. It is recognized as the representative opinion in the three sentences.
Next, the analysis result evaluation unit 21 recognizes the sentences (1) and (5) that share the representative opinion with the evaluation axis, intention type, and sensitivity expression as approximate opinions included in the representative opinion of (2). Associate with the representative opinion. In addition, “3”, which is the total number of representative opinions and approximate opinions, is counted as the number of representative opinions.

Although illustration is omitted, following the same procedure, the analysis result evaluation unit 21 extracts the sentences (3) and (6) that share the sensitivity expression `` bottling '', and then the sentence (3) with a smaller number of characters. I think it's only a rip-off. "
The analysis result evaluation unit 21 recognizes the sentence (6) as an approximate opinion included in the representative opinion of (3) and associates it with the representative opinion.

The screen generation unit 25 generates the analysis result screen 60 by filling a predetermined template with the negative / positive distribution graph and the intention classification chart delivered from the analysis result evaluation unit 21 (S24-08).
The analysis result screen 60 is transmitted to the client terminal 28 via the Web server 26 as described above.

In addition, when selection information of a specific intention type related to a specific evaluation axis is transmitted from the client terminal 28 via the web server 26, the screen generation unit 25 displays the representative opinion and each representative opinion passed from the analysis result evaluation unit 21. A screen listing the number of cases (representative opinion + number of approximate opinions) is generated and transmitted to the client terminal 28 via the Web server 26.
As a result, as shown in FIG. 12, on the Web browser of the client terminal 28, a screen 60 in which representative opinions related to a specific evaluation axis and intention type are listed in the sidebar is displayed.

Further, when selection information of a specific representative opinion is transmitted from the client terminal 28 via the Web server 26, the screen generation unit 25, the specific contents of the corresponding representative opinion and approximate opinion passed from the analysis result evaluation unit 21 Is generated and transmitted to the client terminal 28 via the Web server 26.
As a result, as shown in FIG. 12, a screen 60 on which a list of representative opinions and approximate opinions is arranged in the main area is displayed on the Web browser of the client terminal 28.

  When the user clicks the “time series analysis result” button 66 in the screen 60, the time series analysis processing by the analysis result evaluation unit 21 and the time series analysis unit 23 is executed, and a screen reflecting the analysis result is displayed on the screen. Generated by the generation unit 25 and transmitted to the client terminal 28 via the Web server 26.

FIG. 16 shows the time-series analysis result screen 70. In the main area of the screen, a time-series change graph 71 in which the number of cases is set on the vertical axis and the time is set on the horizontal axis is displayed. Yes.
As for the change point P in the time series change graph 71, “There are a lot of negative opinions, and 53% of the opinions are cautions. The consideration sentence 72 with the content “It has appeared” is clearly indicated. The significance of the “change point” will be described later.

Hereinafter, a processing procedure related to the time series analysis will be described with reference to the flowchart of FIG.
First, the analysis result evaluation unit 21 that has received a request for time series analysis from the client terminal 28 via the Web server 26 (S30), based on the time information (writing date, etc.) of each document included in the analysis result data, The number of documents is totaled (S32), and a time series change graph is generated based on the time series information (S34).
Next, the analysis result evaluation unit 21 passes the daily number information (time series information) to the time series analysis unit 23 and requests detection of a change point.

Receiving this, the time series analysis unit 23 specifies the change point by putting the above time series information into a predetermined time series analysis model (algorithm) (S36).
Here, the “change point” means the point in time when the basic properties of the data change in a set of multiple random time-series data, and multiple optimal analysis models are prepared according to the characteristics of the data. . FIG. 18 shows the characteristics of the main time series analysis model.
The time series analysis unit 23 specifies a change point by applying an “ARIMA model” by default, and returns it to the analysis result evaluation unit 21.

  On the other hand, the analysis result evaluation unit 21 generates a consideration sentence corresponding to the change point and appends it to the time series change graph as described above (S38). A method for generating the consideration sentence will be described in detail later.

  The screen generation unit 25 generates a time series analysis result screen 70 based on the time series change graph passed from the analysis result evaluation unit 21 (S40).

As described above, the time series analysis based on the “ARIMA model” is executed for “all analysis result data” by default, but the user limits the analysis target period and intention type, or a different time series analysis model. Can also be requested.
Specifically, the user narrows down the analysis target period by selecting a start period and a period in the period designation field 73 prepared in the sidebar.
The user narrows down the intention type of the target sentence by checking a check box of a necessary intention type in the target data designation column 74. For example, if “unsatisfied” under “general” is checked, a sentence to which a tag of <dissatisfied> related to all evaluation axes is assigned is designated as a time-series analysis target. On the other hand, when “dissatisfied” under “price” is checked, only sentences with the <dissatisfaction> tag related to the evaluation axis of “price” are designated as time series analysis targets.
Furthermore, the user can obtain a change of the analysis model to be applied by designating an arbitrary analysis model in the analysis model designation field 75.

When the user who has completed the setting in the sidebar clicks the “redisplay” button 76, the specified condition data is transmitted from the client terminal 28 to the Web server 26.
Upon receiving this specified condition data via the Web server 26, the analysis result evaluation unit 21 creates a time-series change graph based on the analysis result data that matches the specified condition, and specifies the change point based on the data. The time series analysis unit 23 is requested.
Receiving this, the time series analysis unit 23 inputs time series data into the designated time series analysis model, identifies a new change point, and returns it to the analysis result evaluation unit 21.
The analysis result evaluation unit 21 generates a consideration sentence related to the change point.
The screen generation unit 25 generates a time series analysis result screen based on the time series change graph and the consideration sentence passed from the analysis result evaluation unit 21.
This time series analysis result screen is transmitted to the client terminal 28 via the Web server 26.
As a result, a new time series analysis result screen is displayed on the Web browser of the client terminal 28 (not shown).

  When the user clicks the “automatic consideration result” button 77 in the screen 70, the analysis result evaluation unit 21 executes a consideration sentence generation process, and a screen describing the consideration sentence is generated by the screen generation part 25. It is transmitted to the client terminal 28 via the server 26.

  FIG. 19 shows this time-series analysis result screen 80, which includes “Comprehensive consideration”, “Consideration by respondent attribute (1)”, “Consideration by respondent attribute (2)”, “Price”. Consideration sentence display fields 81 such as "Considerations on", "Considerations on cooking", "Considerations on rooms", and "Considerations on baths" are provided, and each consideration sentence display field 81 has a tendency of description contents in the analysis target document. Is described in simple words (comments).

  For example, “Comprehensive consideration” is displayed as “Generally, there are a lot of positive opinions, accounting for 75% of the opinions. By referring to this, the user can roughly grasp the tendency of the entire document file.

  In addition, as “Consideration by respondent attribute (1),” “Attribute '30's male” has a very large number of positive opinions, accounting for 62% of the opinions. Is displayed, and by referring to this, the user can grasp the evaluation tendency specific to the attribute of the respondent describing the document.

  In addition, as a “consideration on price”, “83% of positive opinions are related to price. As a popular opinion, there are many opinions that“ the price is cheap ”” is displayed. By referencing, the user can roughly recognize the evaluation tendency regarding “evaluation axis: price”.

Hereinafter, a processing procedure related to generation of a consideration sentence related to “price” will be described.
First, the analysis result evaluation unit 21 determines the number of sentences associated with the “price” evaluation axis, and “dissatisfaction”, “popularity”, “request”, “unexpected”, “question”, “intention” Is applied to the consideration rule stored in the consideration rule storage unit 24 to generate a consideration sentence.

FIG. 20 shows an example of the consideration rule, and each rule includes data items of “priority”, “consideration name”, “judgment condition”, and “template”.
On the other hand, as shown in FIG. 21, the analysis result evaluation unit 21 checks the match / mismatch between the determination condition and the analysis target data in descending order of priority. Insert a specific numerical value (percentage of positive) in the column, insert the most representative opinion in the Opinion Insertion Column, and insert “About Price” at the beginning of the sentence, which is a pillow related to the price evaluation axis. To complete the discussion.
Here, the “most representative opinion” means a representative opinion having the largest number of representative opinions + approximate opinions among sentences given “price” as an evaluation axis.

  Even in the case of generating a consideration sentence relating to another evaluation axis, the analysis result evaluation unit 21 follows the same procedure as described above to complete a consideration sentence unique to the evaluation axis.

Even when generating a comprehensive consideration sentence, the analysis result evaluation unit 21 basically follows the same procedure as described above.
In other words, the rules for considering the total number of documents from price to location and the number of intent types of “dissatisfied”, “popular”, “request”, “unexpected”, “question”, and “intention” included in the document. After identifying the corresponding template, insert a specific numerical value into the “ratio insertion field” and insert the “most frequent opinion” with the highest frequency through all the evaluation axes into the “opinion insertion field”. Completion of the consideration sentence by inserting the phrase “generally” into the beginning of the sentence.

Next, discussion text generation processing for each attribute of the respondent will be described.
First, the analysis result evaluation unit 21 classifies all sentences to be analyzed in attribute group units of “age × gender” associated with each document, and totals the number of cases.
Next, as shown in FIG. 22, the analysis result evaluation unit 21 applies the determination condition of the consideration rule to the analysis target sentence in units of “age × sex” to derive the determination result (true / false).
Then, a consideration rule template corresponding to only one attribute group is extracted as a template of the consideration sentence, and the consideration sentence is completed through insertion of necessary numerical values and words.

  For example, in the case of the consideration name “very popular”, only the attribute group of “male in their 30s” has “judgment result = true”, so the corresponding template is “very positive opinion, [Ratio]%, and it is worth paying attention. ”Is taken out, the necessary numerical value and representative opinion are inserted into the“ ratio insertion column ”and the“ most opinion insertion column ”, and“ attribute “in men in their 30s” ”is inserted at the beginning of the sentence. Is generated.

  In addition, as for the consideration name “question”, only the attribute group of “female 30s” has “judgment result = true”, so a lot of questions “[Opinion]”, which is the corresponding template, appear. Is extracted, then “most opinion” is inserted, and “attribute“ for women in their 30s ”” is inserted at the beginning of the sentence to generate a consideration sentence.

  On the other hand, in the case of the consideration name “request”, since “judgment result = true” is obtained for the attribute group of “male in the 20s” as well as “male in the 40s”, the attribute is based on the corresponding template. Another consideration sentence is not generated.

  In FIG. 22, only the attribute group of women in their 20s to 40s is shown for convenience of illustration, but in reality, the above determination processing is executed for all attribute groups, and only one attribute group is matched. A consideration sentence is generated based on the template of the consideration rule.

Next, a method for generating the consideration sentence 72 displayed in the time series analysis shown in FIG. 16 will be described.
First, as shown in FIG. 23A, the analysis result evaluation unit 21 recognizes a predetermined period (for example, one week) centered on the change point P as a focus period.
Next, the analysis result evaluation unit 21 applies the consideration rule to the analysis result data within the target period.
At the same time, the analysis result evaluation unit 21 applies the consideration rule to the analysis result data within the entire period including the period of interest.
Then, a consideration sentence is generated based on a consideration rule template corresponding only to the period of interest.

  For example, as shown in FIG. 23 (b), regarding the consideration rule with the consideration name “substantially dissatisfied”, only the period of interest is “true” and the whole period is “false”. As a dissatisfied opinion, “[Opinion]” is cited as a template, and a specific numerical value or character string (maximum) A consideration sentence is completed by inserting a representative opinion).

In this way, the determination result based on the analysis result data within the entire period is compared with the determination result based on the analysis result data within the period of interest, and a consideration sentence is generated based on the consideration rule template suitable only for the latter. This makes it possible to derive a consideration sentence peculiar to the period of interest.
Therefore, the user can easily recognize the unique evaluation tendency in the vicinity of the change point.

  In addition, regarding the consideration name “dissatisfied”, only the period of interest is “true” and the determination result is “false” for all periods. However, because the priority is inferior, the template related to this consideration rule is applied. There is no.

  The user can return to the analysis result screen 60 by clicking the “analysis result” button 85 on the screen 80 (see FIG. 11).

  In the above, the sensitivity analysis system 10 according to the present invention has been described as an example in which the various screens (HTML files) are transmitted via the Web server 26 to the client terminal 28 equipped with the Web browser. Of course, it is also possible to implement a so-called stand-alone system in which a computer program for realizing the system 10 is set up in a computer such as a PC.

It is a block diagram which shows the function structure of the Kansei analysis system which concerns on this invention. It is a figure which shows an example of an analysis condition designation | designated screen. It is a flowchart which shows the process sequence in a sensitivity analysis system. It is a flowchart which shows the procedure of a negative / positive determination process. It is a schematic diagram which shows the structural example of a sensitivity dictionary. It is a table which shows the registration data of the word dictionary which belongs to a general purpose dictionary group. It is a figure which shows the execution example of the negative / positive determination process with respect to a specific sentence. It is a figure which shows the example of execution of the subject inheritance process with respect to a specific sentence. It is a table which shows the registration data of an intention expression memory | storage part. It is a figure which shows the execution example of the intention expression extraction process with respect to a specific sentence. It is a figure which shows an example of an analysis result screen. It is a figure which shows an example of an analysis result screen. It is a flowchart which shows the process sequence which concerns on the production | generation of an analysis result screen. It is a table which shows the registration data of an evaluation axis similarity expression dictionary. It is a figure which shows the procedure of representative opinion extraction. It is a figure which shows a time series analysis result screen. It is a flowchart which shows the process sequence which concerns on the production | generation of a time series analysis result screen. It is a figure which shows the characteristic of various time series analysis models. It is a figure which shows an automatic consideration result screen. It is a figure which shows an example of a consideration rule. It is a figure which shows the production | generation procedure of a consideration sentence. It is a figure which shows the production | generation procedure of the attribute-specific consideration sentence. It is a figure which shows the production | generation procedure of the consideration sentence which concerns on the change point at the time series analysis.

10 Kansei analysis system
11 Document Registration Department
12 Document storage
13 Document Analysis Department
14 Negative / Positive judgment part
15 Sensitivity dictionary storage
16 Negative expression storage
17 Subject Certification Department
18 Intention extraction unit
19 Intention expression storage
20 Analysis result storage
21 Analysis result evaluation department
22 Evaluation axis similarity expression storage
23 Time Series Analysis Department
24 Consideration rule memory
25 Screen generator
26 Web server
27 Internet
28 Client terminal
40 Analysis condition specification screen
41 Analysis target file specification column
42 Evaluation axis specification field
43 Browse button
44 Start analysis button
50 general dictionaries
51 word dictionary
52 Binary relational dictionary
53 Phrasebook Dictionary
54 Mimicry Dictionary
55 Domain-specific dictionaries
56 word dictionary
57 Binary relational dictionary
58 Mimicry Dictionary
60 Analysis result screen
61 Negative / Positive Distribution Graph
62 Intention classification chart
63 Representative Opinion List
64 Breakdown list of representative opinions
65 Select button
66 Time Series Analysis Results button
70 Time Series Analysis Result Screen
71 Time series change graph
72 Thoughts
73 Period designation field
74 Target data specification column
75 Analysis model specification field
76 Refresh button
77 “Automatic Result” button
80 Time series analysis result screen
81 Discussion text display column
85 “Result” button

Claims (8)

A sensitivity dictionary storage means for storing a combination of a sensitivity term representing an affirmative / negative value judgment for an object and an attribute indicating whether it is affirmative / negative;
A morpheme analyzing unit that performs morpheme processing on a document file and decomposes a plurality of text sentences included in the document file into morpheme units;
It is determined whether or not an expression composed of a morpheme or a combination of morphemes included in each sentence corresponds to a sensitivity term stored in the sensitivity dictionary storage unit, and if applicable, is combined with the expression. Means for associating an attribute tag indicating any of the positive / negative attributes;
Negative expression storage means for storing negative expressions used when constructing negative sentences;
It is determined whether or not an expression composed of a morpheme or a combination of morphemes included in each sentence corresponds to a negative expression stored in the negative expression storage means, and if so, is a negative expression for the expression. Means for associating a negative expression tag indicating
Based on the combination pattern of the attribute tag type associated with each sentence and the presence / absence of a negative expression tag, it is determined whether the sentence is a negative evaluation sentence or a positive evaluation sentence, and for negative evaluation sentences, a negative determination tag And a means for adding a positive determination tag to a positive evaluation sentence,
Searching for a particle for constructing the main character in each sentence, and subject recognition means that recognizes an independent word attached to the particle as a subject;
An intention expression storage means for storing a combination of an intention expression representing a speaker's intention and a type of the intention;
The sentence with the negative determination tag is associated with an intention type tag that means “dissatisfied”, and the sentence with the positive determination tag is associated with an intention type tag that means “popular”. It is determined whether or not an expression composed of a morpheme or a combination of morphemes corresponds to an intention expression stored in the intention expression dictionary storage means, and if so, an intention indicating a corresponding type for the expression Means for associating a type tag;
An evaluation axis similar expression storage means for storing a correspondence relationship between a character string indicating a plurality of types of evaluation axes and a similar expression included in each evaluation axis;
A means for determining whether the subject recognized in each sentence corresponds to the similar expression stored in the evaluation axis similar expression storage means, and, if applicable, associating a corresponding evaluation axis with the sentence;
Means for counting a number of sentences for each predetermined time interval based on time information associated with each sentence for a set of sentences in a predetermined range, and generating time-series information having a correspondence relationship between the time interval and the number of cases When,
Means for generating a time series change graph based on the time series information;
Means for identifying a change point on the time series change graph based on the time series information;
A consideration rule storage means for storing a plurality of consideration rules composed of combinations of application conditions defined by the number and ratio of sentences to which each intention type tag is assigned and a consideration sentence template;
Apply each of the above consideration rules to a set of sentences within a given period of interest that includes the above change point, determine whether or not the application condition is met, and determine a sentence within the entire period including the above change point Means for applying each of the above consideration rules to the set of and determining whether or not the application condition is matched;
When there is a consideration rule that matches only a set of sentences within the period of interest, a consideration sentence generation unit that extracts a template of the consideration rule as a consideration sentence regarding the change point;
Means for outputting the time series change graph and the consideration sentence;
Kansei analysis system characterized by having. The above template has a field for inserting opinions in advance.
Means for classifying each sentence into a similar sentence group that is a set of similar sentences having a common evaluation axis, intention type, and sensibility term, and certifying a similar sentence with the least number of characters in each similar sentence group as a representative opinion With
The said consideration sentence production | generation means inserts the representative opinion with the largest number of the similar sentences contained in the corresponding similar sentence group in the said representative opinion in the said opinion insertion column. Kansei analysis system. The template has a ratio insertion field for inserting a ratio of a sentence to which a specific intention type tag has been assigned in advance.
Means for calculating a ratio of sentences to which the intention type tag is attached;
The sensitivity analysis system according to claim 1 or 2, wherein the consideration sentence generation means inserts the ratio into a ratio insertion field of the template.   The above-described time-series change graph generation process, change point identification process, and consideration sentence generation process are executed on a set of sentences associated with a specified evaluation axis or intention type tag. Item 4. The sensitivity analysis system according to any one of Items 1 to 3.   5. The time series change graph generation process, the change point identification process, and the consideration sentence generation process are executed for a set of sentences belonging to a specified period. Kansei sensitivity analysis system.   6. The change point specifying process is executed by applying one time series analysis model designated among a plurality of time series analysis models prepared in advance. Kansei sensitivity analysis system. Each of the above consideration rules has a priority,
The consideration sentence generation means generates a consideration sentence based on a template related to a consideration rule having a high priority when there are a plurality of consideration rules that match only a set of sentences within the period of interest. The sensitivity analysis system according to any one of claims 1 to 6. Computer
A sensitivity dictionary storage means for storing a combination of a sensitivity term representing an affirmative / negative value judgment for an object and an attribute indicating whether it is positive / negative;
Morpheme analysis means for performing morpheme processing on a document file and decomposing a plurality of text sentences included in the document file into morpheme units;
It is determined whether or not an expression composed of a morpheme or a combination of morphemes included in each sentence corresponds to a sensitivity term stored in the sensitivity dictionary storage unit, and if applicable, is combined with the expression. Means for associating an attribute tag indicating any of the positive / negative attributes;
Negative expression storage means for storing negative expressions used when constructing negative sentences,
It is determined whether or not an expression composed of a morpheme or a combination of morphemes included in each sentence corresponds to a negative expression stored in the negative expression storage means, and if so, is a negative expression for the expression. Hand associating a negative expression tag indicating that
Based on the combination pattern of the attribute tag type associated with each sentence and the presence / absence of a negative expression tag, it is determined whether the sentence is a negative evaluation sentence or a positive evaluation sentence, and for negative evaluation sentences, a negative determination tag And a means for adding a positive determination tag to a positive evaluation sentence,
Search for a particle for composing the main character in each sentence, and determine the subject recognition means to recognize the independent word attached with the particle as the subject,
An intention expression storage means for storing a combination of an intention expression representing a speaker's intention and a type of the intention;
The sentence with the negative determination tag is associated with an intention type tag that means “dissatisfied”, and the sentence with the positive determination tag is associated with an intention type tag that means “favored”. It is determined whether or not an expression composed of a morpheme or a combination of morphemes corresponds to an intention expression stored in the intention expression dictionary storage means, and if so, an intention indicating a corresponding type for the expression Means for associating type tags,
Evaluation axis similar expression storage means for storing a correspondence relationship between a character string indicating a plurality of types of evaluation axes and a similar expression included in each evaluation axis;
Means for determining whether the subject recognized in each sentence corresponds to the similar expression stored in the evaluation axis similar expression storage means, and in this case, associating a corresponding evaluation axis with the sentence;
Means for counting a number of sentences for each predetermined time interval based on time information associated with each sentence for a set of sentences in a predetermined range, and generating time-series information having a correspondence relationship between the time interval and the number of cases ,
Means for generating a time series change graph based on the time series information;
Means for identifying a change point on the time series change graph based on the time series information;
Consideration rule storage means for storing a plurality of consideration rules composed of combinations of application conditions defined by the number and ratio of sentences to which each intention type tag is assigned and a consideration sentence template,
Apply each of the above consideration rules to a set of sentences within a given period of interest that includes the above change point, determine whether or not the application condition is met, and determine a sentence within the entire period including the above change point Means for applying each of the above consideration rules to the set of and determining whether or not the application condition is met,
A consideration sentence generating means for extracting a consideration rule template as a consideration sentence related to the change point when there is a consideration rule that only matches a set of sentences within the period of interest;
Means for outputting the time-series change graph and the consideration sentence;
Kansei analysis program characterized by functioning as

Images