JP2006286026A - Opinion collection/analysis device, opinion collection/analysis method used therefor and its program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To automatically extract opinions of persons from a set of documents present in the Internet or the like and to compare and analyze the opinions from various viewpoints. <P>SOLUTION: An object name set input means 11 of an input/output device 1 is used for inputting a plurality of object names by a user. An object name selection means 12 selects an object name of an analysis object out of a set of the object names inputted by the user by using the object name input means 11. A document set storage part 21 of a data processor 2 stores the document set, and an opinion extraction means 22 extracts opinions included in the document set stored in the document set storage part 21 and related to the object name inputted from the object name input means 11. An object comparison/analysis means 23 analyzes features of the opinions (partial opinions) related to the object name selected by the object name selection means 12 within the opinions (the whole opinions) extracted by the opinion extraction means 22 and related to the object names inputted by the object name set input means 11. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

The present invention relates to an opinion collection / analysis apparatus method and program for use therein, and more particularly to an opinion collection / analysis apparatus that compares and analyzes opinions extracted from a document set from various viewpoints.

It is important to collect and analyze consumer opinions for market research and customer satisfaction surveys. Conventionally, as a method for collecting and analyzing this opinion, a method has been used in which opinions are collected from consumers using a questionnaire, and the overall characteristics and trends are analyzed from the answer sentences.

For example, in the questionnaire response analysis system disclosed in Japanese Patent Laid-Open No. 2001-266060, knowledge of the rule format is automatically extracted from a free answer sentence from a consumer by a text classification engine, and the extraction result is transmitted to the user. Presented as an analysis result.

On the other hand, focusing on the point of collecting opinions, there is a method of extracting opinions from information described on an Internet Web page. For example, “Reputation information retrieval from the Internet” by Kenji Tateishi, Yoshihide Ishiguro, Shunichi Fukushima [Proceedings of the 62nd National Convention of Information Processing Society (3), p123-124, 2001] Opinions related to products input by the user are automatically extracted from various information and presented as search results to the user.

However, in the above-mentioned questionnaire response analysis system, people's opinions are collected using questionnaire questionnaires, but the questionnaire implementer selects the questionnaire survey cost and high quality response monitors and requests the questionnaire. There is a problem that there are many cases where opinions of a sufficient number of free answer sentences cannot be obtained even if the investigation is conducted with such efforts.

In addition, the above-described reputation information search system can automatically collect opinions, but simply listing the collected opinions in the search results will allow the user to determine what the product is about. There is a problem that it is impossible to know the overall characteristics and trends such as what is good or bad or how good it is compared to other products.

Furthermore, when the above questionnaire response analysis system and reputation information search system are combined, it is not possible to compare and analyze the extracted opinions simply by inputting the reputation information search system output into the questionnaire response analysis system. There's a problem.

Furthermore, when the above questionnaire response analysis system and reputation information search system are combined, the reputation information search system does not have sufficient opinion extraction accuracy, so the output of the reputation information search system is simply input to the questionnaire response analysis system. Therefore, there is a problem that a highly accurate analysis result may not be obtained.

Accordingly, an object of the present invention is to provide an opinion collection / analysis apparatus that can solve the above-described problems and present a result of analyzing the characteristics of an opinion related to an object selected by a user among all extracted opinions to a client It is another object of the present invention to provide a method for collecting and analyzing opinions and a program therefor.

Another object of the present invention is to provide an opinion collection and analysis apparatus capable of realizing accurate opinion extraction and presenting a highly accurate analysis result of the extracted opinion to a client, an opinion collection and analysis method used therefor, and To provide that program.

Another object of the present invention is to provide an opinion collection / analysis apparatus and an opinion collection / analysis method used therefor that can present a client with the result of analyzing the characteristics of the opinion related to the category selected by the user among the extracted opinions. To provide that program.

An opinion collection / analysis apparatus according to the present invention includes an object name set input unit for inputting a plurality of object names, and an object to be analyzed from the object name set input from the object name set input unit. Object name selection means for selecting a name, document set storage means for storing a document set, and an object included in the document set stored in the document set storage means and input from the object name set input means An opinion extracting means for extracting an opinion related to the name, and an object comparison analyzing means for analyzing the characteristics of the opinion related to the object name selected by the object name selecting means among the opinions extracted by the opinion extracting means; It has.

Another opinion collecting / analyzing apparatus according to the present invention includes an object name input means for inputting one or more object names, and an object to be analyzed in the object name set input from the object name input means. Object name selection means for selecting an object name, object name expansion means for expanding an object name related to the object name input from the object name input means, and a document set are stored. A document set storage means, an object name included in the document set stored in the document set storage means and input from the object name set input means, and an object name extended by the object name extension means Opinion extraction means for extracting related opinions, and object comparison analysis means for analyzing the characteristics of the opinion related to the object name selected by the object name selection means among the opinions extracted by the opinion extraction means ing.

Another opinion collection / analysis apparatus according to the present invention includes an object name input means for inputting one or more object names, an analysis condition input means for inputting an analysis condition, and a document set storage means for storing a document set. An opinion extraction unit that extracts an opinion that is included in the document set stored in the document set storage unit and that is related to the object name input from the object name input unit, and that is extracted by the opinion extraction unit Category setting means for setting various categories for opinions, and comparative analysis means for analyzing the characteristics of opinions that satisfy the conditions specified by the analysis condition input means among the opinions extracted by the opinion extraction means Yes.

An opinion collection and analysis method according to the present invention includes a step of inputting a plurality of object names, a step of selecting an object name to be analyzed from the input object name set, and a document set storing a document set Extracting opinions related to the input object name included in the document set stored in the storage means, and analyzing the characteristics of the opinion regarding the selected object name in the extracted opinions And steps.

Another opinion collection and analysis method according to the present invention includes a step of inputting one or more object names, a step of selecting an object name to be analyzed from the input object name set, and the input A step of expanding a target name related to the target name based on the target name and a document set stored in a document set storage means for storing the document set and the input target name and the expanded The method includes a step of extracting an opinion related to the object name, and a step of analyzing a feature of the opinion related to the selected object name among the extracted opinions.

Another opinion collection and analysis method according to the present invention is included in the document set stored in the document set storage means for storing the document set, the step of inputting one or more object names, the step of inputting the analysis conditions, and the document set. A step of extracting an opinion related to the input object name, a step of setting various categories for the extracted opinion, and a feature of the opinion satisfying the analysis condition in the extracted opinion. Analyzing.

A program for an opinion collection and analysis method according to the present invention includes a process for inputting a plurality of object names to a computer, a process for selecting an object name to be analyzed from the input object name set, and a document set. Processing for extracting an opinion related to the input object name and included in the document set stored in the document set storage means for storing the comment, and an opinion related to the selected object name in the extracted opinion And processing for analyzing the characteristics of the.

Another opinion collection and analysis method program according to the present invention includes a process for inputting one or more object names to a computer and a process for selecting an object name to be analyzed from the set of input object names. A process of expanding the object name related to the input object name based on the input object name, and the input object name included in the document set stored in the document set storage means for storing the document set And processing for extracting an opinion related to the expanded object name, and processing for analyzing the characteristics of the opinion regarding the selected object name in the extracted opinion.

Another opinion collection and analysis method program according to the present invention is stored in a document set storage means for storing one or more object names, a process for inputting analysis conditions, and a document set in a computer. A process of extracting opinions related to the input object name included in the document set, a process of setting various categories for the extracted opinions, and the analysis condition in the extracted opinions. And a process of analyzing the characteristics of the opinions to be satisfied.

That is, the opinion collection / analysis apparatus of the present invention analyzes an object name set input means for a user to input a plurality of object names, and an object name set input by a user using the object name set input means. The object name selection means for selecting the target object name, the document set storage section for storing the document set, and the object name set input means included in the document set stored in the document set storage section. An opinion extracting means for extracting opinions related to the object name, and an opinion (overall opinion) on the object name input by the object name set input means extracted by the opinion extracting means. And object comparison analysis means for analyzing the characteristics of opinions (partial opinions) related to the object name selected by the object name selection means.

By adopting such a configuration, the opinion collection / analysis apparatus of the present invention extracts the opinions of a plurality of object names from the document set, and thereby relates to the object selected by the user among the extracted opinions as a whole. The result of analyzing the characteristics of the opinion can be presented to the client.

In addition, the opinion collection and analysis apparatus of the present invention includes an evaluation expression storage unit in which a set of evaluation expressions, which are expressions indicating evaluation of things, and an evaluation expression stored in the input object and the evaluation expression storage unit. Opinion candidate extraction means for extracting a character string or sentence including an object and an evaluation expression as an opinion candidate when there is a certain relationship, and an opinion determination means for determining the likelihood of the opinion candidate extracted by the opinion candidate extraction means And have.

By adopting such a configuration, the opinion collection and analysis apparatus of the present invention realizes accurate opinion extraction by determining the likelihood of opinion extracted from the Internet by a method such as syntactic suitability or objective suitability. As a result, a highly accurate analysis result of the extracted opinion can be presented to the client.

Furthermore, the opinion collection and analysis apparatus of the present invention stores an object name input means for a user to input one or more object names, an analysis condition input means for a user to input analysis conditions, and a document set. A document set storage unit, an opinion extraction unit that extracts an opinion related to the object name input from the object name input unit included in the document set stored in the document set storage unit, and the opinion extraction unit Among the category setting means for setting various categories for the extracted opinions, and opinions regarding the object name input by the object name input means extracted by the opinion extraction means (overall opinions), And comparative analysis means for analyzing the characteristics of opinions (partial opinions) that satisfy the conditions specified by the analysis condition specification selection means.

By adopting such a configuration, the opinion collection / analysis apparatus of the present invention automatically sets various categories to the opinions extracted from the document set, and the user selects among the extracted opinions as a whole. The result of analyzing the characteristics of opinions regarding the category can be presented to the client.

As described above, the opinion collection / analysis apparatus of the present invention inputs a plurality of object names, selects an object name to be analyzed from the input object name set, and stores a document set By extracting an opinion related to the input object name included in the document set stored in the set storage means, and analyzing characteristics of the opinion related to the object name selected in the extracted opinion, As a result, it is possible to present the client with the result of analyzing the characteristics of the opinion regarding the object selected by the user in the extracted opinions as a whole.

Another opinion collection / analysis apparatus of the present invention inputs one or more object names, selects an object name to be analyzed from the input object name set, and inputs the input object name. Based on the object name related to the expanded object name, the document name stored in the document set storage means for storing the document set is extracted and opinions related to the input object name and the extended object name are extracted. By analyzing the characteristics of the opinions related to the selected object name in the extracted opinions, it is possible to realize accurate opinion extraction, and to provide the client with the highly accurate analysis results of the extracted opinions. The effect that it can be presented is obtained.

Another opinion collection / analysis apparatus according to the present invention includes one or more object names, an analysis condition, and a document set stored in a document set storage means for storing the document set. By extracting opinions related to the object name, setting various categories for the extracted opinions, and analyzing the characteristics of the opinions that satisfy the analysis conditions in the extracted opinions, the entire opinion is extracted. Thus, an effect that the result of analyzing the feature of the opinion regarding the category selected by the user can be presented to the client is obtained.

Next, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration of an opinion collection / analysis apparatus according to a first embodiment of the present invention. In FIG. 1, the opinion collection / analysis apparatus according to the first embodiment of the present invention comprises an input / output device 1 and a data processing device 2.

The input / output device 1 includes an object name set input unit 11 for a user to input a plurality of object names, and an object name to be analyzed among the object name sets input by the user using the object name set input unit 11. The object name selecting means 12 for selecting and an output means 13 for outputting the analysis result of the object comparison / analysis means 23.

The data processing apparatus 2 is mainly composed of a computer, and is input from a document set storage unit 21 for storing a document set and an object name set input unit 11 included in the document set stored in the document set storage unit 21. The opinion extraction means 22 for extracting an opinion related to the object name and the opinion on the object name (the overall opinion) input by the object name set input means 11 extracted by the opinion extraction means 22 Object comparison / analysis means 23 for analyzing the characteristics of opinions (partial opinions) related to the object name selected by the object name selection means 12 and a recording medium for storing a program for realizing each means of the data processing device 2 24. Each means of the data processing device 2 can be configured by hardware.

The object names in the object set input means 11 are the same field such as “product“ M ””, “product“ K ””, “product“ S ””, “product“ R ”” [for example, computer or It is effective to input a plurality of object names in a field such as PDA (Personal Digital Assistance). The object name is not limited to a product name, and may be a company / organization name such as “N Company” or a personal name such as “Taro Yamada”.

The document set in the document storage unit 21 may be in any form as long as it is a digitized document set. However, documents on the Internet from which various people freely transmit information are effective for opinion collection analysis. Is.

A feature of the object comparison / analysis means 23 is that a plurality of pairs of an object and an opinion are input, and an analysis result output is classified into, for example, “1) if-then-else type rule to classify an opinion as an object. "Expression as a rule", "2) Expression as a correlation rule that associates an opinion with an object by an if-then-or type rule", "3) Expression of an object and its characteristic word in a positioning map", " 4) “Expression in a list in which opinions are arranged in a typical order with respect to the object”.

In particular, the 1) if-then-else rule can be obtained by the method disclosed in “Questionnaire Answer Analysis System” of Japanese Patent Laid-Open No. 2001-266060.

FIG. 2 is a block diagram illustrating a configuration example of the opinion extraction unit 22 of FIG. In FIG. 2, the opinion extraction unit 22 includes an evaluation expression storage unit 221 in which a set of evaluation expressions, which are expressions indicating evaluation of things, and an evaluation expression stored in the input object name and evaluation expression storage unit 221. Opinion candidate extraction means 222 that extracts a character string or a sentence including the object name and the evaluation expression as an opinion candidate from the Web document storage unit 25 (document set storage unit 21), and an opinion candidate An opinion determination unit 223 that determines whether the opinion candidate extracted by the extraction unit 222 is recognized as an opinion is included.

The evaluation expression stored in the evaluation expression storage unit 221 is an expression indicating an evaluation of things. An example of the evaluation expression relating to the product is shown in FIG. In FIG. 3, the evaluation expressions corresponding to the product field “common” include “good”, “like”, “best”, “recommended”, and “satisfied”, and the evaluation expression corresponding to the product field “book” Has "interesting", "masterpiece", "difficult to read", "easy to understand", and evaluation expressions corresponding to the product field "computer" are "fast", "heavy", "fragile", "noisy" There are “unstable”, and the evaluation expressions corresponding to the product field “food” include “delicious”, “taste”, and “persistent”.

On the other hand, expressions that do not indicate evaluation for things such as “play”, “move”, and “research” are not evaluation expressions. In addition, the types of evaluation expressions include expressions indicating human feelings and feelings such as “good”, “like” and “delicious”, and expressions indicating characteristics and characteristics of human objects such as “fast” and “heavy”. There is. As a storage format of evaluation expressions, a method of storing in a unit that absorbs a difference in notation due to utilization of each evaluation expression is effective in terms of saving storage capacity, but is not limited to this method.

The case where the object and the evaluation expression are in a certain relationship in the opinion candidate extraction means 222 is, for example, a case where the object and the evaluation expression are in a certain number of characters or bytes, or syntactic dependence on the object and the evaluation expression. This is the case.

There are two methods for determining whether an opinion is recognized as an opinion in the opinion determination unit 223: a method for determining the syntactic opinion likelihood of an opinion candidate by the syntactic aptitude determination unit 223a, and an opinion candidate by the objective aptitude determination unit 223b are objective. There is a method for determining whether or not.

The syntactic opinion likelihood of the opinion candidate in the syntactic aptitude determination means 223a is determined using the syntactic characteristics of the opinion as shown in FIG. In FIG. 4, the syntactic characteristics of the opinion include “(1) the object name and the evaluation expression exist in the same sentence”, “(2) the object name exists in the main part of the evaluation expression”, There are “(3) an evaluation expression exists in a statement predicate” and “(4) an unnecessary sentence ending expression does not exist in the statement predicate”.

The unnecessary sentence ending expression (4) is, for example, “Is the product“ M ”good? Or a hypothetical expression such as “I wish the product“ M ”should be a little lighter”.

The syntactic appropriateness determining means 223a takes the syntactical characteristics of the opinion as described above as a numerical value indicating how much the opinion candidate satisfies, and if the numerical value exceeds a predetermined threshold value, the syntactic appropriateness determining means 223a determines that the opinion is syntactically likely . As a method of making a numerical value, for example, a method of considering the ratio of the number of syntactic characteristics that the opinion candidate satisfies to the total number of syntactic characteristics is conceivable. A method is also conceivable in which a weight is assigned to each syntactic characteristic and the sum of the syntactic characteristic weights satisfied by the opinion candidates is obtained.

In the objective aptitude determination means 223b, the opinion candidate is objective means that the opinion is described by a third party. For example, considering that “product“ M ”is recommended” is an opinion candidate, this opinion candidate satisfies all of the four syntactic characteristics in the syntactic appropriateness determination means 223a, so that it is determined to be syntactically an opinion. The However, if this opinion candidate is described by the developer of “product“ M ”” or the seller of “product“ M ””, it cannot be trusted as an opinion.

The objective suitability determination means 223b prepares a list of writers who are not recognized as such third parties, and the opinion candidates are not objective when the writers of opinion candidates include matches or partial matches in this list Otherwise, it is determined that the opinion candidate is objective. If the document is a document on the Internet, this list can be represented by a URL (Uniform Resource Locator), and FIG. 5 shows an example of a URL list in the field of computers.

5, the above list includes “so.co.jp” as the URL of “So Company”, “ibb.co.jp” as the URL of “Ib Company”, and “tt.co” as the URL of “T Company”. .Jp "," Nnn.co.jp "as the URL of" N Company "," aaa.co.jp "as the URL of" Company A "," sh.co.jp "," URL of "Sh Company" “Ccc.co.jp” as the URL of “Company C”, “fff.co.jp” as the URL of “Company F”, “io.co.jp” as the URL of “Io Company”, URL of “J Company” As “jjj.co.jp”.

6 is a flowchart showing the processing operation of the opinion collection / analysis apparatus according to the first embodiment of the present invention, and FIG. 7 is a flowchart showing the processing operation of opinion extraction in FIG. The processing operation of the opinion collection / analysis apparatus according to the first embodiment of the present invention will be described with reference to FIGS. 6 and 7 is realized by the data processing apparatus 2 executing a program stored in the recording medium 24.

First, the user selects a set of object names from the object name set input means 11 and an object name to be analyzed from the object name selection means 12 to the opinion collection / analysis apparatus (step S1 in FIG. 6). . Here, “product“ M ””, “product“ K ””, “product“ S ”” and “product“ R ”” are input by the object name set input unit 11, and “object“ R ”is input by the object name selection unit 12. A case where the product “M” is selected will be described.

Next, an opinion related to the object name input by the object name set input unit 11 included in the document set stored in the document set storage unit 21 is extracted by the opinion extraction unit 22 (step S2 in FIG. 6).

This opinion extraction process is performed by the opinion candidate extraction unit 222 in which an evaluation expression storage unit 221 in which a set of evaluation expressions that are expressions indicating evaluation of things is registered, and the input object and the evaluation expression stored in the evaluation expression storage unit 221. When the expression has a certain relationship, a process of extracting a character string or a sentence including the object and the evaluation expression as an opinion candidate (step S11 in FIG. 7), and the opinion determination means 223 causes the opinion candidate extraction means 222 to And processing for determining the likelihood of opinion of the extracted opinion candidate (step S12 in FIG. 7).

An example of the opinion extracted by the above opinion extraction process is shown in FIG. In this example, an opinion candidate is first extracted based on a certain relationship between the object name and the evaluation expression in the opinion candidate extraction unit 222 as a condition that both are within 80 bytes, and then the syntactic aptitude of the opinion by the opinion determination unit 223, It is the result of determining objective suitability.

The syntactic characteristics are conditional on satisfying at least three of the above four characteristics. For example, the first line of FIG. 8 is an opinion extracted with the object name product “M” and the evaluation expression name “good”, and all the syntactic characteristics of the four opinions are also satisfied.

Next, the opinion collecting / analyzing apparatus applies the object regarding the opinion (the entire opinion) about the object name input by the object name set input means 11 extracted by the opinion extracting means 22 by the object comparison analyzing means 23. The feature of the opinion (partial opinion) regarding the object name selected by the name selection means 12 is analyzed, and the result is output via the output means 13 (step S3 in FIG. 6).

That is, by inputting a plurality of pairs of an object and an opinion related thereto, for example, 1) a classification rule, 2) a correlation rule, 3) a positioning map, 4) one of four comparative analysis results of a typical sentence list Output one.

The classification rule 1) expresses the relationship between opinions and objects in an if-then-else format. For example, as shown in FIG. 9, OS “D” → product “M” (70%), monochrome → product “M” (70%), e-mail → product “M” (100%), word processor → product “ “M” (100%), transmission → product “M” (90%), otherwise → product “M” (78.9%) is “if the word“ OS ”D” appears in the opinion Is related to the product “M”, otherwise, if the word “monochrome” appears in the opinion, the opinion is related to the product “M” ”.

Here, in the above case, “OS“ D ””, “monochrome”, and the like are characteristic words related to the object “M”. These feature words are selected in descending order of the information amount scale from words appearing in a set of opinions obtained by removing all opinions in which the upper words appear from the whole.

The information amount scale is calculated as follows. For example, if there are N opinions, and there are N1 and N0 objects for the object A and those other than A, respectively, the word or word combination “w” appears. Suppose that the number is T, among which there are T1 and T0 objects A and other than A, respectively, and the number of opinions in which the word “w” does not appear is S, among which the object is If there are S1 and S0 for A and other than A, there is a relationship of T + S = N, T1 + S1 = N1, T0 + S0 = N0, and the information scale of “w” is, for example, “ As the value obtained by subtracting the amount of information I (w) held in the opinion set after classification through the appearance of “w” from the information amount I in the opinion set before classification through the appearance of “w”, that is, ΔI (W) = I−I (w)... (1)

Here, I uses, for example, a stochastic complexity amount proposed by the document “J. Rissanen: FisherInformation and StochasticComplexity. IEEE Etrans. OnInformationTheory, 42 (1), pp: 40-47, 1996”. , I = N · H (N1 / N) + (1/2) log [N (π / 2)] (2). Further, H (x) = [(− x) log (x)] − [(1−x) log (1−x)], π is the circumference, and the base of the logarithm is 2.

Similarly, I (w) is I (w) = T · H (T1 / T) + (1/2) log [T (π / 2)] + S · H (S1 / S) + (1/2 ) Log [S (π / 2)] (3). “W” is selected such that ΔI (w) has the largest value.

Also, instead of stochastic complexity, for example, the document “K. I = min {N1, N−N1} + B√NlogN (4) I (w) = min {T1, T−T1} + B√TlogT + min {S1, S−S1} + B√SlogS .. (5), and then the information scale can be calculated by the equation: ΔI (w) = I−I (w) (6).

As shown in FIG. 10, the object comparison / analysis unit 23 that outputs such a classification rule includes a morpheme analysis unit 231, an attribute extraction unit 232, and an information amount scale calculation unit 233.

The morpheme analyzing means 231 breaks down the opinion sentence into words. For example, “the morpheme analyzing means decomposes the sentence of opinion into words” is decomposed into “morpheme / analysis / means / has / opinion / of / sentence / decomposed into / word / decomposed”.

The attribute extraction means 232 calculates the value of the expression (1) for each word w obtained by morphological analysis, and selects only those whose values are greater than a given threshold. The information amount scale calculation means 233 calculates the value of equation (1) for the entire opinion set for a word or a plurality of combinations of words, and selects the largest one.

The opinion including the selected word is removed from the whole, and the value of the formula (1) is calculated for the word or a plurality of combinations of words for the remaining opinions, and the maximum value is selected from the values. . This is repeated below.

The association rule 2) represents the relationship between an opinion and an object in an if-then-or format. For example, the same figure as the above classification rule can be read as a correlation rule. In that case, “If the word OS“ D ”appears in the opinion, the opinion relates to the product“ M ”, or if the word“ monochrome ”appears in the opinion, the opinion is the product“ M ”. And read.

Here, in the above case, “OS“ D ””, “monochrome”, and the like are characteristic words related to the object “M”. These feature words are selected from the words appearing in the entire opinion set in descending order of the information amount scale Δ (w) in the equation (1).

As shown in FIG. 11, the object comparison / analysis means 23 that outputs such an association rule includes a morpheme analysis means 234, an attribute extraction means 235, and an information amount scale calculation means 236.

The morpheme analyzing means 234 breaks the sentence of opinion into words. The attribute extraction means 235 calculates the value of the expression (1) for each word “w” obtained by morphological analysis, and selects only those values that are greater than a given threshold value.

The information amount scale calculation means 236 calculates the value of the expression (1) for the entire opinion set for a word or a plurality of combinations of words, and arranges the values in order from the largest.

The positioning map 3) is obtained by arranging a plurality of objects and feature words related to them in a two-dimensional map diagram. For example, FIG. 12 shows a positioning made up of a plurality of objects such as a product “M”, a product “K”, a product “S”, and a product “R”, and feature words calculated by the correlation rule of 2). Represents a map. Specifically, the crosses shown in FIG. 12 indicate the positions on the axes of the product “M”, the product “K”, the product “S”, and the product “R”, and are used as feature words for the product “M”. “Pass”, “Find”, and “Send” are the characteristic words for the product “K”. “Debut”, “FIVA”, “MD”, “Unexpected”, and “Sorry” are the characteristic words for the product “S”. "Product" J "", "Product" C "", "Memory card" are the characteristic words for the product "R", "Floppy (R)", "Heat", "CPU" P "" It shows that it exists in the positional relationship.

Once such objects and their characteristic words have been obtained, such a positioning map is applied to a standard correspondence analysis method as described in the document “JP Benzecri: Correspondence Analysis Handbook, MercelDecker, 1992”. Therefore created.

That is, an eigenvector corresponding to the largest eigenvalue and an eigenvector corresponding to the second largest eigenvalue are obtained from a matrix obtained by appropriately symmetrizing a matrix in which the frequency of feature words is arranged over the entire object to be analyzed. And plotting the first component and the second component on the vertical axis and the horizontal axis.

As shown in FIG. 13, the object comparison analysis unit 23 that outputs such a positioning map includes a morpheme analysis unit 237, an attribute extraction unit 238, an information amount scale calculation unit 239, and a correspondence analysis unit 240.

The morpheme analyzing means 237 breaks down the opinion sentence into words. The attribute extraction means 238 calculates the value of the expression (1) for each word “w” obtained by morphological analysis, and selects only those values that are greater than a given threshold value.

The information amount scale calculation means 239 calculates the value of the expression (1) for the entire opinion set for a word or a plurality of combinations of words, and the value is the number given in order from the largest. Or, a word having an information amount scale equal to or greater than a given threshold is extracted.

The correspondence analysis unit 240 creates a correspondence matrix between the extracted words and the analysis target, and the element is the appearance frequency of each word, and the first component and the second component of the first eigenvector and the second eigenvector of this matrix are used. Create a positioning map by finding and plotting.

The typical sentence list of 4) calculates the representative level of each opinion (degree that is typical as the opinion of the target object) for each object, and sorts the values in descending order of the representative level.

In the case where the product “M” is an object, for example: Product "M" is difficult to use and slow 2. 2. Product “M” has become larger, smaller is better This product "M" is my favorite Marvelous. The weak points of this product “M” are arranged in order of darkness.

The representative degree of the sentence B in the object c is, for example, the sentence B = w1,. . . . , WM, Bayes' posterior probability is p (c | B) = p (c) (Πjp (wj | c)) / (Σcp (c) (Πjp (wj | c) ))) ...... It is given by the equation (7).

Here, p (c) is the appearance probability of the object c, and is calculated from the data by the following equation: p (c) = (Nc + 1/2) / (N + J / 2) (8). Here, J is the total number of objects, N is the total number of opinions, and Nc is the number of appearances of the object c therein.

Further, p (wj | c) is a probability that the word wj appears, and this is a Laplace-type estimation amount, p (wj | c) = (Kj + 1/2) / (K + J / 2) (9) Calculated by the formula. Here, J is the total number of words, K is the total number of all words appearing in the opinion data, and is calculated including duplication. Kj is the number of occurrences of the word wj in it.

As shown in FIG. 14, the object comparison analysis unit 23 that outputs such a typical sentence list includes a morpheme analysis unit 241, an attribute extraction unit 242, a representative degree calculation unit 243, and a sort unit 244.

The morpheme analyzing means 241 breaks down the opinion sentence into words. The attribute extraction unit 242 calculates the value of the expression (1) for each word w obtained by morphological analysis, and selects only those values that are greater than a given threshold value.

The representative degree calculating unit 243 calculates the value of the expression (7) for each opinion sentence B with respect to the whole word selected by the morpheme analyzing unit 241. The sorting means 244 outputs the result of sorting the opinion sentences in descending order or ascending order according to the value of equation (7).

FIG. 15 is a block diagram showing the configuration of the opinion collection / analysis apparatus according to the second embodiment of the present invention. In FIG. 15, the opinion collection / analysis apparatus according to the second embodiment of the present invention is an object name in which a user inputs one or more object names to the input / output device 3 instead of the object name set input means 11. Except for providing the input means 31 and providing the data processing device 4 with the object name expansion means 41 for expanding the object name related to the object name input by the object name input means 31. The configuration is the same as that of the opinion collection and analysis apparatus according to the first embodiment, and the same components are denoted by the same reference numerals. The operation of the same component is the same as that of the first embodiment of the present invention.

The object name expanding unit 41 automatically expands the object name input by the object name input unit 12 using a dictionary in which similar object names are registered in advance as a group. For example, “product“ M ””, “product“ K ””, “product“ S ””, “product“ R ”” are registered as a group, and “product“ M ”” is input by the object name input means 12. “Product“ K ””, “Product“ S ””, and “Product“ R ”” are expanded.

In the second embodiment of the present invention, the data processing device 4 is mainly composed of a computer, and each means in the device is realized by the data processing device 4 executing the program stored in the recording medium 42. The Each means of the data processing device 4 can be configured by hardware.

FIG. 16 is a block diagram showing the configuration of the opinion collection / analysis apparatus according to the third embodiment of the present invention. In FIG. 16, the opinion collection / analysis apparatus according to the third embodiment of the present invention comprises an input / output device 5 and a data processing device 6.

The input / output device 5 outputs the analysis result of the object name input means 31 for the user to input one or more object names, the analysis condition input means 51 for the user to input the analysis conditions, and the comparison analysis means 62. And output means 13 for performing the above operation.

The data processing device 6 is mainly composed of a computer, and is included in a document set storage unit 21 for storing the document set, and included in the document set stored in the document set storage unit 21 and input from the object name input unit 31. Opinion extraction means 22 for extracting opinions related to the object name, category setting means 61 for setting various categories for the opinions extracted by the opinion extraction means 22, and object names extracted by the opinion extraction means 22 Comparative analysis means 62 for analyzing the characteristics of opinions (partial opinions) that satisfy the conditions specified by the analysis condition input means 51 with respect to opinions (overall opinions) regarding the object name input by the input means 31; And a recording medium 63 for storing a program for realizing each means of the data processing device 6. Each means of the data processing device 6 can be configured by hardware.

The categories in the category setting means 61 include (a) the function category of the object, (b) the model of the object model, (c) the opinion evaluation category, (d) the opinion evaluation expression category, (e There is a category of opinion description time, and (f) a category of opinion description location. Setting a category means giving a label to each category.

Analysis conditions in the analysis condition input means 51 are described by designating a category and a label of the opinion. As the category, in addition to the categories (a) to (f) set by the category setting means 61, the object name input by the object name input means 31 is also handled as the category of the object name (g). For example, the analysis condition is specified as “object name = product“ M ”& opinion evaluation = affirmation”. This means that an opinion in which the “object name category” label is “product“ M ”” and the “opinion evaluation category” label is “affirmed” is to be analyzed.

Features in the comparative analysis means 62 are expressed as knowledge of the rule format. For example, when the text classification engine in the “questionnaire answer analysis system” described in Japanese Patent Application Laid-Open No. 2001-266060 is used, it is possible to learn the characteristics of a part of the opinion with respect to the entire opinion by an if-then-else type rule. it can.

FIG. 17 is a flowchart showing the processing operation of the opinion collection / analysis apparatus according to the third embodiment of the present invention, and FIG. 18 is a flowchart showing the processing operation of category generation in FIG. The processing operation of the opinion collection / analysis apparatus according to the third embodiment of the present invention will be described with reference to FIGS. Note that the processing shown in FIGS. 17 and 18 is realized by the data processing device 6 executing a program stored in the recording medium 63.

First, when the user inputs one or more object names using the object name input unit 31 (step S21 in FIG. 17), the opinion extraction unit 22 is included in the document set stored in the document set storage unit 21 and is the target. Opinions related to the object name input by the object name input means 31 are extracted (step S22 in FIG. 17).

Next, the category setting means 61 sets various categories for the opinions extracted by the opinion extraction means 22 (step S23 in FIG. 17). Further, when the user inputs the analysis condition by the analysis condition specifying means 51 (step S24 in FIG. 17), the comparative analysis means 61 relates to the object name extracted by the opinion extraction means 22 and input by the object name input means 31. The characteristics of opinions (partial opinions) that satisfy the conditions specified by the analysis condition specifying means 51 are analyzed for the opinions (total opinions) (step S25 in FIG. 17).

The specific operation of the above category setting process is an operation when “object” is replaced with “category” in the operation of step S3 of the opinion collection and analysis apparatus according to the first embodiment of the present invention. It is the same.

The operation of the category setting means 61 first selects one of one or more prepared categories (step S31 in FIG. 18). The prepared categories include (a) the function category of the object, (b) the category of the object model, (c) the opinion evaluation category, (d) the opinion evaluation expression category, and (e) the opinion. (F) The category of the place where the opinion is written.

(A) The category of the function of the object classifies opinions from the viewpoint of the function of the object. For example, if the object is a mobile phone, the functions of the object may be “phone function”, “charging function”, “mail function”, “calendar function”, “phone book function”, “schedule management function”, etc. It is done.

(B) The category of the model of the object classifies opinions from the viewpoint of the model of the object. For example, the product name “product“ M ”” is “product“ MII-550 ”” “product“ M-K1 ””, the company name “N” is “N software”, and “N system construction”. Thus, it is a lower-order object name including the input object name.

(C) The opinion evaluation category classifies opinions from the viewpoint of evaluation. Opinion evaluation indicates whether an opinion writer views the object positively or negatively. Therefore, it can be expressed in the form of two values “positive” or “negative” or their degree.

(D) The category of opinion evaluation expression classifies opinions from the viewpoint of the evaluation expression. The evaluation expression of the opinion is an evaluation expression stored in the evaluation expression storage unit 221 used by the opinion extraction unit 22.

(E) The opinion description time category classifies opinions from the viewpoint of the description time. Opinion writing time indicates when an opinion writer described the opinion.

(F) The category of the description location of the opinion is classified from the viewpoint of the description location. An opinion description location indicates where the opinion writer described the opinion.

Next, the category setting means 61 gives an opinion label for the selected category (step S32 in FIG. 18). The method of assigning the opinion label depends on the selected category.

(A) In the case of the function category of an object, as a method of assigning an opinion label, for example, a list of expressions indicating functions is prepared, and the expression is included in the opinions extracted by the opinion extraction unit 22 May have a method of giving the expression as a label of opinion. For example, if the object is a mobile phone, “phone”, “charge”, “mail”, “calendar”, “phone book”, “schedule management”, and the like can be considered as expressions indicating functions.

In addition, as a method of assigning opinion labels, for example, there may be a method of automatically collecting expressions indicating the above functions using co-occurrence words. The co-occurrence word is, for example, “function”. In other words, the expression indicating the function is automatically collected by using a regular expression as a series of nouns and sagittal nouns adjacent to the front of the co-occurrence word, such as “XX function”.

(B) In the case of the category of the model of the object, as a method of assigning the opinion label, there is a method of preparing a table in which the object name and the model name are associated, for example. In this case, when the model name is expanded from this table using the target name input by the target name input means 31 as a key, and then the model name is included in the opinion extracted by the opinion extraction unit 21 The model name is assigned as a classification label.

For example, “N company” → “N software”, “N system construction”, “product“ M ”” → “product“ MII-R550 ””, “product“ M ” −K1 ””.

(C) In the case of an opinion evaluation category, as a method of assigning an opinion label, for example, a table (basic evaluation table) in which an evaluation expression and its basic evaluation are associated with each other, and a negative expression list (negative expression list) There is a method of using.

In this case, first, the basic evaluation of the opinion is obtained from the evaluation expression included in the opinion extracted by the opinion extraction unit 22 using the basic evaluation table. Next, when a negative expression included in the negative expression table is included in the vicinity of the evaluation expression included in the opinion, an evaluation obtained by inverting the basic evaluation is given as an opinion label. Conversely, if a negative expression is not included, the basic evaluation is given as an opinion label.

Here, “good” → “affirmation”, “bad” → “denial”, “fast” → “affirmation”, “heavy” → “denial” correspond to the evaluation expression in the basic evaluation table and the basic evaluation. To do. Also, as negative expressions, grammatical negative expressions that deny predicates by grammar, such as “not”, “not”, “not”, and “not”, “question”, “can't say” There are lexical negative expressions that negate predicates by vocabulary, and register them.

(D) In the case of opinion evaluation expression category, as a method of assigning an opinion label, for example, there is a method of assigning an evaluation expression included in an opinion extracted by the opinion extracting means 22 as a classification label.

(E) In the case of a category based on opinion description time, as a method of assigning an opinion label, for example, if the document is a document on the Internet, the update time of the document including the opinion extracted by the opinion extracting means 22 There is a method of giving as a label.

(F) In the case of a category based on an opinion description location, as a method of assigning an opinion label, for example, if the document is a document on the Internet, the URL of the document containing the opinion extracted by the opinion extracting means 22 There is a method of giving as a label.

In addition, as a method of assigning an opinion label, for example, if the document is a document on the Internet, there is a method of assigning the URL domain of the document including the opinion extracted by the opinion extracting means 22 as an opinion label.

Finally, if there is a category that has not yet been selected, the process returns to step S31, and if all categories have been selected, the process ends (step S33 in FIG. 18).

In this way, by extracting opinions of multiple object names from the document set, the characteristics of opinions regarding the object selected by the user in the extracted opinions are analyzed, and the results are sent to the client. Can be presented.

In addition, it is possible to extract accurate opinions by judging the likelihood of opinions extracted from a set of documents by a method such as syntactic aptitude or objective aptitude, and the requester can obtain a highly accurate analysis result of the extracted opinions. Can be presented.

Furthermore, by automatically setting various categories for the opinions extracted from the document set, the characteristics of the opinions related to the category selected by the user in the extracted opinions are analyzed, and the results are sent to the client. Can be presented.

It is a block diagram which shows the structure of the opinion collection analyzer by 1st Example of this invention. It is a block diagram which shows the structural example of the opinion extraction means of FIG. It is a figure which shows the example of the content of the evaluation expression dictionary memory | storage part of FIG. It is a figure which shows the example of the syntactic characteristic in the syntactic aptitude determination means of FIG. It is a figure which shows the example of the URL list | wrist in the objective aptitude determination means of FIG. It is a flowchart which shows the processing operation of the opinion collection analyzer by 1st Example of this invention. It is a flowchart which shows the processing operation of opinion extraction of FIG. It is a figure which shows the example of an output of the opinion extraction means of FIG. It is a figure which shows the example output by the classification rule of the target object comparison analysis means of FIG. It is a block diagram which shows the structural example by the classification rule of the target object comparison analysis means of FIG. It is a block diagram which shows the structural example by the correlation rule of the target object comparison analysis means of FIG. It is a figure which shows the example of an output by the positioning map of the target object comparison analysis means of FIG. It is a block diagram which shows the structural example by the positioning map of the target object comparison analysis means of FIG. It is a block diagram which shows the structural example by the typical sentence list of the target object comparison analysis means of FIG. It is a block diagram which shows the structure of the opinion collection analyzer by 2nd Example of this invention. It is a block diagram which shows the structure of the opinion collection analyzer by the 3rd Example of this invention. It is a flowchart which shows the processing operation of the opinion collection analyzer by 3rd Example of this invention. It is a flowchart which shows the processing operation | movement of the category production | generation of FIG.

Explanation of symbols

1, 3, 5 I / O devices 2, 4, 6 Data processing device 11 Object name set input means 12 Object name selection means 13 Output means 21 Document set storage unit 22 Opinion extraction means 23 Object comparison analysis means 24, 42 , 63 Recording medium 31 Object name input means 41 Object name extension means 51 Analysis condition input means 61 Category setting means 62 Comparison analysis means 221 Evaluation expression storage section 222 Opinion candidate extraction means 223 Opinion determination means 223a Syntactic aptitude determination means 223b Objective aptitude determination means 231, 234, 237, 241 Morphological analysis means 232, 235, 238, 242 Attribute selection means 233, 236, 239 Information amount scale calculation means 240 Corresponding analysis means 243 Representativeness analysis means 244 Sort means

Claims (15)

Object name set input means for inputting a plurality of object names;
Object name selection means for selecting an object name to be analyzed from the object name set input from the object name set input means;
Document set storage means for storing the document set;
An evaluation expression storage means in which a set of evaluation expressions, which are expressions indicating evaluations of things, are registered;
From a document set stored in the document set storage means, a character string or a sentence including each object name input from the object name set input means and an evaluation expression stored in the evaluation expression storage means , Opinion extraction means for extracting as opinions about each object name,
For each word or combination of words extracted from the opinion, an index indicating the probability that the opinion of the portion that is the opinion on the object name to be analyzed occupies in the overall opinion that is the opinion extracted by the opinion extracting means And an object comparison / analysis unit that selects and outputs a word or a plurality of combinations of words based on the index.
An object name input means for inputting one or more object names;
An object name selection means for selecting an object name to be analyzed from the object name set input from the object name input means;
An object name expanding means for expanding an object name related to the object name based on the object name input from the object name input means;
Document set storage means for storing the document set;
An evaluation expression storage means in which a set of evaluation expressions, which are expressions indicating evaluations of things, are registered;
From the document set stored in the document set storage means, each object name input from the object name set input means, the object name extended by the object name extension means, and the evaluation expression storage means An opinion extraction means for extracting a character string or a sentence including a stored evaluation expression as an opinion on each object name;
Opinions of portions that are opinions about the object name selected by the object name selecting means and the object name expanded by the object name expanding means in the entire opinion that is the opinion extracted by the opinion extracting means An object comparison analysis unit that obtains an index indicating the probability occupied by each word or combination of words extracted from the opinion, and selects and outputs a plurality of words or combinations of words based on the index An opinion collection and analysis device characterized by
An object name input means for inputting one or more object names;
An analysis condition input means for inputting an analysis condition;
Document set storage means for storing the document set;
An evaluation expression storage means in which a set of evaluation expressions, which are expressions indicating evaluations of things, are registered;
From a document set stored in the document set storage means, a character string or a sentence including each object name input from the object name set input means and an evaluation expression stored in the evaluation expression storage means , Opinion extraction means for extracting as opinions about each object name,
Category setting means for setting a label based on the content of the opinion, a description time and / or a description location for each predetermined category in the opinion extracted by the opinion extraction means;
An index that accounts for the probability that the opinion of the portion that is the opinion that satisfies the condition specified by the analysis condition input means in the entire opinion that is the opinion extracted by the opinion extraction means is the word or word extracted from the opinion And an object comparison / analysis unit which obtains each combination and selects and outputs a word or a plurality of combinations of words based on the index.
4. The opinion collection and analysis apparatus according to claim 1, wherein the object comparison and analysis unit obtains the index using probabilistic complexity. 5.
4. The opinion collection and analysis apparatus according to claim 1, wherein the object comparison and analysis unit obtains the index using extended probabilistic complexity. 5.
4. The opinion collection and analysis apparatus according to claim 1, wherein the object comparison and analysis unit obtains the index using a Bayesian posterior probability and a Laplace-type estimation amount. 5.
Entering multiple feature names;
Selecting an analysis target name from the input feature name set;
A set of evaluation expressions, which are expressions indicating evaluations for each object name and things input in the step of inputting the object name, are registered from the document set stored in the document set storage means for storing the document set. Extracting a character string or a sentence including an evaluation expression stored in the evaluation expression storage means as an opinion on each object name;
An index indicating the probability that the opinion of the portion that is the opinion on the object name to be analyzed in the entire opinion that is the extracted opinion is occupied is obtained for each word or combination of words extracted from the opinion, and the index Selecting and outputting a word or multiple combinations of words based on:
An opinion collection and analysis method characterized by comprising:
Entering one or more feature names;
Selecting an analysis target name from the input feature name set;
Extending a related object name based on the input object name;
Each object name input in the step of inputting the object name and the object name expanded in the step of expanding the object name from the document set stored in the document set storage means for storing the document set Extracting a character string or a sentence including an evaluation expression stored in an evaluation expression storage means in which a set of evaluation expressions, which is an expression indicating an evaluation of things, is registered as an opinion on each object name;
In the overall opinion that is the extracted opinion, an index that indicates the probability that the opinion of the portion that is the opinion related to the target name of the analysis target and the extended target name occupies the word or word extracted from the opinion Obtaining for each combination, selecting and outputting a word or a plurality of combinations of words based on the index;
An opinion collection and analysis method characterized by comprising:
Entering one or more feature names;
Entering analysis conditions;
A set of evaluation expressions, which are expressions indicating evaluations for each object name and things input in the step of inputting the object name, are registered from the document set stored in the document set storage means for storing the document set. Extracting a character string or a sentence including an evaluation expression stored in the evaluation expression storage means as an opinion on each object name;
Setting a label based on the content of the opinion, a description time and / or a description location for each of the extracted opinions for each predetermined category;
An index indicating the probability that the opinion of the portion that is the opinion that satisfies the analysis condition in the entire opinion that is the extracted opinion is obtained for each word or combination of words extracted from the opinion, and based on the index Selecting and outputting a word or multiple combinations of words;
An opinion collection and analysis method characterized by comprising:
The opinion collection and analysis method according to any one of claims 7 to 9, wherein the index indicating the probability is obtained using stochastic complexity.
The opinion collection / analysis method according to claim 7, wherein the index indicating the probability is obtained by using an extended probabilistic complexity.
10. The opinion collection and analysis method according to claim 7, wherein the index indicating the probability is obtained using a Bayesian posterior probability and a Laplace-type estimator. 11.
The process of entering multiple feature names,
A process of selecting a target name of an analysis target from the input target name set;
From the document set stored in the document set storage means for storing the document set, the set of evaluation expressions, which are expressions indicating the evaluation of the respective object names and things inputted, are stored in the registered evaluation expression storage means. A process of extracting a character string or sentence including an evaluation expression as an opinion on each object name,
An index indicating the probability that the opinion of the portion that is the opinion on the object name to be analyzed in the entire opinion that is the extracted opinion is occupied is obtained for each word or combination of words extracted from the opinion, and the index Processing to select and output a word or multiple combinations of words based on
A program that causes a computer to execute.
The process of entering one or more feature names,
A process of selecting a target name of an analysis target from the input target name set;
A process of extending the input object name to a related object name based on the input object name;
From the document set stored in the document set storage means for storing the document set, the respective object names input from the object name set input means and the object names and things expanded by the object name extension means A process of extracting a character string or a sentence including an evaluation expression stored in an evaluation expression storage means in which a set of evaluation expressions, which is an expression indicating an evaluation, is extracted as an opinion regarding each object name;
In the overall opinion that is the extracted opinion, an index that indicates the probability that the opinion of the portion that is the opinion related to the target name of the analysis target and the extended target name occupies the word or word extracted from the opinion A process for obtaining each combination and selecting and outputting the word or a plurality of combinations of words based on the index;
A program that causes a computer to execute.
Processing to input one or more target names Processing to input analysis conditions,
From the document set stored in the document set storage means for storing the document set, a set of evaluation expressions, which are expressions indicating the evaluation of the respective object names and things inputted from the object name set input means, are registered. Processing to extract character strings or sentences including evaluation expressions stored in the evaluation expression storage means as opinions relating to respective object names The contents and descriptions of the extracted opinions for each predetermined category Processing to set labels based on time and / or location,
The extracted opinions (total opinions) and the opinions satisfying the analysis conditions (partial opinions) are classified according to the presence / absence of each word or word combination extracted from the opinions. And comparing and analyzing the difference in the index representing the probability that the opinion of the part before and after the classification occupies the whole opinion as an information measure, and based on the information measure of the word or a plurality of combinations of words The process of outputting the selection as a feature of the opinion of the part,
A program that causes a computer to execute.

Images