JP3402598B1 - Information processing apparatus, information processing method, and information processing program - Google Patents

Abstract

Abstract: [PROBLEMS] To provide an information processing device or the like that can automatically extract an expression representing emotionalness from a sentence. SOLUTION: A word representing an emotion and a word for emphasizing or negating the word are prepared in advance as a dictionary. The input sentence is analyzed morphologically and syntactically, and the resulting syntactic structure and the meaning of the words are used to create the words that are the main emotional words and expressions that emphasize or deny them. Extract. In addition, the words representing emotionality registered in the dictionary are given numerical values of the emotional strength. On the other hand, words that emphasize or deny the word expressing emotionality have a certain emphasis. Operators and the like indicating the degree of are given. Then, the emotional strength and the like of the expression extracted using these numerical values and operators can be expressed numerically.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an information processing device, an information processing method, and an information processing program, and for example, relates to one that analyzes and processes a sentence.

[0002]

2. Description of the Related Art In recent years, techniques for analyzing and processing sentences created by humans, such as morphological analysis and syntactic analysis, have been developed. These text analysis processing techniques are basic techniques applicable to various fields. For example, in the field of information retrieval, a search formula is generated from natural sentences entered by the user,
It can support the execution of a search, and in the field of home appliances, it can analyze the commands uttered by the user and control the home appliances.

There are various types of techniques for analyzing and processing sentences, and for example, there is a technique of extracting a specific word such as a noun from a text and classifying it. This is an analysis based on the part-of-speech of a word appearing in the text, and is a morphological analysis of the text, and as a result, a noun or verb in the text that meets a specific part-of-speech condition is extracted. This method can also handle compound nouns. In addition, as other methods, for example, the frequency of appearance of words is used, or words that easily co-occur are extracted.
Some use statistical methods based on parts of speech and face.

[0004]

However, all of these conventional techniques focus on the morpheme of the text,
The analysis was done using the types of parts of speech and statistical methods. Then, expressions such as “smooth skin” and “beautiful Asahi” that imply qualitative and subjective evaluations are not expressed as morphemes but as words or meanings. There is no system that recognizes and extracts from the traditional structure.

Therefore, an object of the present invention is to provide an information processing apparatus or the like which can automatically extract an expression expressing emotionality from a sentence and quantify the intensity of emotionality.

[0006]

In order to achieve the above object, in the invention according to claim 1, a sentence acquisition for obtaining a sentence including an emotion expression word which is a word expressing an emotion for a predetermined object. means, emotional expression word specifying means for specifying the emotional expression word used in the sentence in which the acquired, the
Increased the emotional intensity of emotional expressions in sentences
Affect emotional expression words such as
Function word specifying means for specifying a function word that is a word to be changed
And a form in which the sentence is morphologically analyzed and a morpheme string is output.
Elementary parsing means and a parser for parsing the morpheme string
And using the results of parsing by the parsing means
Is specified as an emotional expression word by the emotional expression word specifying means.
The specified word as a function word, or the function word
A word specified as a functional word by the specifying means is an emotional expression word
Re-identifying means for re-identifying, and the emotion expression word identifying means,
Or using the emotional expression word specified by the re-identification means ,
An information processing apparatus comprising: an extracting unit that extracts an emotional expression portion that expresses an emotional expression from the sentence, and an output unit that outputs the extracted emotional expression portion. In the invention described in claim 2, wherein the extraction means, of the sentence, the emotional expression word identifying means also
Is the emotion expression word identified by the re-identification means and the machine
The information processing apparatus according to claim 1, wherein a portion including a function word that acts on the emotion expression word specified by the noun word specifying unit or the re-specifying unit is extracted. The invention according to claim 3 further comprises a numerical value associating unit for associating a numerical value with the emotional expression word included in the extracted emotional expression portion, and using the numerical value associated by the numerical value associating unit, 3. The method according to claim 1, wherein the degree of emotional strength of the emotional expression portion is quantified.
And an information processing device described in 1. In the invention described in claim 4, the information processing apparatus according to claim 3, characterized in that in the numerical correlating means equipped the further numerical changing means for changing the numerical value associated with the emotion expression word provide. In the invention according to claim 5 , further comprising an operation associating unit for associating an operation method with a function word acting on the emotion expression word included in the extracted emotion expression portion, and a numerical value associated with the emotion expression word. and by performing a calculation by the calculation method associated with the operation correlating means, claim 3, characterized in that varying degrees of emotional intensity of the emotion representation part, or to claim 4 An information processing device as described is provided. In the invention of claim 6, wherein the emotional expression word specific means, the morphemes constituting the morpheme string output by morphological analysis means, and search for expression word database storing emotion expression word, the morpheme the expression claim from claim 1, characterized in that identifying the morpheme as emotional expression word when retrieved in the word database 5
An information processing apparatus according to any one of the above claims is provided. In the invention according to claim 7 , the function word identifying means searches the function word database storing the function words for the morphemes forming the morpheme string, and when the morpheme is searched for in the function word database. The information processing apparatus according to claim 6 , wherein the morpheme is specified as a function word. According to the invention described in claim 8 ,
Said extracting means, using the obtained dependency structure in the syntax analysis means, any one of the claims of claims 1 to 7, characterized in that extracting the emotional representation part
An information processing apparatus according to the above item 1 is provided. In the invention according to claim 9 , a sentence acquisition means, an emotion expression specifying means, and a machine
Noun specifying means, morphological analysis means, syntactic analysis means,
In a computer including re-identification means, extraction means, and output means, a sentence acquisition step of acquiring a sentence including an emotion expression word, which is a word expressing emotion for a predetermined object, in the sentence acquisition means. When, in the emotional expression word specifying means, and emotion expression word specifying step of specifying the emotional expression word used in the acquired text, the sentence
Changes emotional strength by acting on emotional expressions
A function word that specifies a function word that is a word by the function word specifying means
The specific step and the morphological analysis means shape the sentence.
Morphological analysis step of performing morphological analysis and outputting a morphological sequence
And the syntactic analysis means parses the morpheme string.
Syntax analysis step and analysis result by the syntax analysis means
By using the re-identification means,
A word specified as an emotional expression word by the word specifying means is a functional word.
Re-specify or as a function word by the function word specifying means
Re-identification step of re-identifying the identified word as an emotional expression word
If the emotion expression word specific means, or using said emotional expression word that is especially <br/> constant re specifying means, the emotional representation part representing the emotional expression from the sentence by the extraction means An information processing method, comprising: an extraction step of extracting; and an output step of outputting the extracted emotion expression part by the output means. In the invention according to claim 10 , a sentence acquisition function of acquiring a sentence including an emotion expression word that is a word expressing emotion for a predetermined object,
An emotional expression word specifying function for specifying an emotional expression word used in the acquired sentence;
Machines that act on emotional expressions to change emotional strength
A function to specify a noun word and a morphological solution to the sentence
A morphological analysis function for analyzing and outputting a morphological sequence,
A parsing function for parsing a prime sequence, and the parsing means
By using the analysis result by
The words identified as emotional expressions by the fixed method are re-specified as functional words.
Specified or specified as a function word by the function word specifying means
A re-identification function for re-identifying a specified word as an emotional expression word, and
Using the emotion expression word specifying function, or the emotion expression word specified by the re-specifying function, an extraction function of extracting an emotion expression portion expressing an emotional expression from the sentence, and the extracted emotion expression portion An information processing program for realizing the output function of outputting and a computer is provided.

[0007]

BEST MODE FOR CARRYING OUT THE INVENTION (First Embodiment) Hereinafter, a preferred first embodiment of the present invention will be described with reference to FIGS.
Will be described in detail with reference to. In the present embodiment, a sentence including a subjective evaluation such as "this lipstick has a beautiful color development" is processed. Therefore, first, terms used in processing these subjective sentences will be explained.

(1) Emotionality Emotionality refers to a property that includes qualitative and subjective expressions such as emotional movements or likes and dislikes. For example, expressions such as "this book is interesting" and "the food at the restaurant there was delicious" are expressions that include subjective expressions and express emotionality. However, what represents emotionality or what is subjective can be freely set depending on the operation method of the system, such as how to create the effect term dictionary 22 described later. For example, when the expression “beautiful jewel” is used, “beautiful” can be considered as the objective attribute of the jewel, and also as the subjectivity of the person who expressed it. Even in such a case, it can be widely dealt with by a method of creating a dictionary.

In general, when a subjective expression is used, a subjective evaluation is often expressed, for example, "This lipstick has a beautiful coloring. Therefore, in the present embodiment, a description will be given focusing on the case of processing subjective expressions (including subjective evaluation) among emotional expressions. However, this does not limit the system of the present embodiment to one that processes expressions including subjective expressions. For example, due to the design of a dictionary, etc. In addition, emotional expressions that are not necessarily subjective can be processed. As mentioned above,
In the present embodiment, expressions that include subjective expressions centered on words that express subjective evaluation are considered to be expressions that express emotions widely, without strictly defining emotionality. To do. Thus, emotional expression is a broader concept than subjective expression. Furthermore, expressions that include subjective evaluation are
It is a concept included in subjective expressions. (2) Affect expression Affect expression refers to an expression that includes qualitative and subjective expressions, that is, an expression that includes emotional expression. For example, the expression "this flower is very beautiful" is an expression that includes the subjective expression of the person who saw the flower, and is an affective expression. Also,
In other words, the affective expression can be said to be an emotional expression part expressing emotion in the sentence. Further, when the emotional part represents a subjective expression, it can be referred to as a subjective expression part, and when the emotional part represents a subjective evaluation, it can be referred to as a subjective evaluation expression part.

(3) Affect term An affect term is a morpheme that represents emotionality that forms the center of an affect expression. In the above example, the adjective stem "Beauty", which is the "Beautiful" without its ending, is the Affective Term. In other words, affect terms are emotional expressions that express emotional expressions. Further, when the emotion expression word represents a subjective expression, it can be called a subjective expression word, and when the emotion expression word represents a subjective evaluation, it can be called a subjective evaluation word. (4) Function words Function words are words that give meaning to the affect term such as emphasis, degree, or negation. Then, the function word is combined with the affect term to form one affect expression. For example, with the expression "not very happy",
"Excessive" indicates the degree, and "absent" is a functional word indicating negation. And these function words are combined with the affect term "Ureshi" to form one affect expression. In addition, examples of the function word indicating emphasis include “very” and “very”.

(5) Affectiveness Affectability means that each morpheme in the sequence of morphemes to be processed is classified into either an affect term or a function word, but neither. The property of whether or not it is classified as is called an effect. As will be described later, the effect expression processing apparatus 1 of the present embodiment includes an affect term dictionary that collects affect terms and function words as a database, and by searching each morpheme in a morpheme with this affect term dictionary, each morpheme It is possible to determine whether is an affect term, a functional word, or neither.

(6) Affect information Affect information is information obtained by searching an affect term dictionary for a certain morpheme. As will be described later in detail, as the affect information, for example, information such as whether a certain morpheme is an affect term or a functional word, and whether the emotion represented by the affect term is positive or negative. It includes information such as whether it is emotional, and information that quantifies the emotional strength. (7) Article An article is a group of sentences for extracting an affective expression. Usually, it is a sentence of several lines to several tens of lines, such as information in which an impression sentence for a certain product is written at once. For example, when a large number of impression sentences for lipsticks are posted on the electronic bulletin board, each impression sentence is an article. Therefore, the article in this embodiment has a broader concept than a newspaper article or an article evaluated by a specialist.

(8) Subject of Evaluation The subject of evaluation refers to the subject evaluated by the article being analyzed. For example, if there is an article that "smell lasts long" for a certain perfume, this perfume is the evaluation subject. In the present embodiment, products such as cosmetics are handled as the evaluation subject, but the evaluation subject is not limited to the product, and for example, an abstract object such as service or organization reputation may be used as the evaluation subject. In this embodiment, the articles for analysis are collected in advance for each evaluation subject, but the present invention is not limited to this, and the articles may be automatically classified for each evaluation subject.

(9) Evaluation axis The evaluation axis is a classification of aspects (attributes) for evaluating an evaluation subject. This is set for each evaluation subject. For example, if the evaluation subject is lipstick, the side (attribute) includes color gloss, smell, comfort, package, price, etc.
These can be evaluated. When the subject of evaluation is a car, attributes such as fuel consumption, riding comfort, operability, appearance, price, etc. can be considered. (10) Axis term (evaluated word) The axis term means a direct word (usually a noun) evaluated by a certain expression. For example, in the sentence "coloring is beautiful", "coloring" is the axis term. In this way, the axis term is a word that expresses a concrete expression for a particular aspect of evaluation (evaluation axis). It corresponds to the evaluation axis “color gloss”. In addition, axis terms such as “comfort”, “feel”, and “moisture” relate to the evaluation axis “comfort”. Further, for example, in the above example sentence, “coloring” is also an emotional word that is the target of emotion of a word expressing the emotion of “beautiful”. If the expression is not limited to the subjective evaluation and widely expresses an emotional expression, the axis term can be said to be an emotional word.

The terms described above can be used, for example, as follows. Lipstick A (specific lipstick product)
In the article whose evaluation subject is, the affective expression "very glossy" is an evaluation for the axis term "coloring", and "coloring" is a word belonging to the evaluation axis "coloring gloss" in the evaluation of lipstick.

By the way, it is not always easy to determine whether a certain article (evaluation subject) is treated positively or negatively, so it is necessary to strictly define emotionality. It can be difficult. For example, just because an article affirms an evaluation subject does not necessarily mean that the creator of the article favorably evaluates the evaluation subject. Consider the following example sentence. "That actress is very beautiful, but I don't like it." Being beautiful as an actress is very important, and while this statement affirms her existence as an actress, she has a negative personal preference.

Therefore, in the present embodiment, the meaning of the word "emotionality" is understood as broadly as possible, and the system is interpreted so that it can be flexibly customized. Then, in the present embodiment, a non-quantitative part is evaluated, whereas a numerical value or the like that can be quantitatively evaluated, such as “specification / performance of personal computer”. For example, the evaluation subject is qualitatively evaluated by collecting comments from many people, and
Quantify emotional strength. In addition, we identify the axis terms that are the subject of the assessment, and analyze what aspects of the assessment subject the axis terms are assessing. In addition, among qualitative expressions, those with relatively high subjectivity in which emotional expression is expressed are treated.

FIG. 1 is a diagram showing a logical configuration of an effect expression processing apparatus 1 of this embodiment. Although the hardware configuration of this device will be described later, it can be realized by using, for example, a personal computer (hereinafter referred to as a PC). In this case, the effect expression processing program is installed in the PC, and this program is executed by the CPU.
Each module is configured by software by executing. Also, each dictionary is stored as a database in a storage device such as a hard disk.

The effect expression processing apparatus 1 includes an input unit 10 for reading articles from an article database 20 which is a database of articles, a morphological analysis unit 11 for morphological analysis of the input articles, and a morphological analysis dictionary 2 used for morphological analysis.
1. an effect information assigning unit 12 that assigns affect information to a morpheme; an affect term dictionary 22 that stores information about affect terms and function words; a parsing unit 1 that parses a morpheme string to which affect information is attached.
3, the parsing dictionary 23 used at the time of parsing, the effect expression extraction unit 14 that extracts the effect expression from the parsed sentence, the axis term determination unit 16 that determines the axis term of the extracted effect expression, and the axis term It is composed of an axis term dictionary 24 in which information is registered, an output unit 15 for outputting the extracted affect expression, and the like. The details of each of these components will be described below.

FIG. 2 is a diagram showing an example of a logical configuration of article data stored in the article database 20. In the article database 20, articles are stored for each evaluation subject. As an example, the figure is a collection of articles with lipstick as the evaluation subject. For example, with regard to lipstick A, there is no smell or strange taste. For each lipstick product (title). Articles have been collected in.

The effect expression processing apparatus 1 does not automatically specify the evaluation subject from the article, and therefore collects articles in advance for each evaluation subject. These articles can be collected, for example, from comments on the electronic bulletin board on the Web regarding product evaluation or questionnaire surveys conducted to users. In the present embodiment, lipstick is used as an evaluation subject as an example, but in addition to this, collection of articles can be performed for various fields such as lipstick, perfume, hot springs, restaurants, watches, and cars.

The input unit 10 (FIG. 1) is a module that retrieves and reads an article from the article database 20 and outputs it to the morphological analysis 11, and constitutes a sentence acquisition means. Although not shown, the input unit 10 includes an article database 2
You can set the field to read from 0. For example, when lipstick is designated as the field, articles related to lipstick are read from the article database 20. It is possible to identify which evaluation subject (specific lipstick product) the read article relates to. Input section 10
Reads an article from the article database 20 one sentence at a time and outputs it to the morpheme analysis unit 11. The components below the morphological analysis unit 11 sequentially process the articles acquired from the input unit 10.

The morpheme analysis unit 11 is a morpheme analysis means for performing morpheme analysis on the input article and outputting a morpheme sequence. A morpheme is a phrase that is more detailed than a bunsetsu and is divided into levels of independent words and adjuncts. The morpheme analysis is to decompose a sentence into morphemes and generate a morpheme string in which information such as a part of speech is added. For example,
The sentence "I got well with those words" is morphologically analyzed as follows. "Ano (adjunct) / word (noun) / de (particle) / energetic (noun) / ni (particle) / became (verb:
Become)". The morphological analysis dictionary 21 stores data used for morphological analysis. As the morphological analysis method and the morphological analysis dictionary 21, known publicly used ones are used. Note that the verb includes an inflectional ending for simplicity, but there is a grammar that processes the verb stem and the verb inflection ending.

The effect information adding unit 12 adds the effect information to the morpheme string while searching the affect term dictionary 22. That is, the emotion expression word specifying means for specifying the emotion expression word is configured. First, the effect term dictionary 22 will be described, and then the effect information adding unit 12 will be described. FIG. 3 is a diagram showing an example of a logical data structure of the affect term dictionary 22. One row of the table in the figure represents information for one word. In the affect term dictionary 22, function words are registered in addition to the affect terms. In addition, there are words that become affect terms or function words depending on the context, and both such words are registered and the default is specified.

In the headline 31, the registered affect terms and function words are recorded. Heading 31
Since affect terms and function words are recorded as morphemes, the words to be used are recorded without the inflection ending. In the figure, the inflection ending is put in (). Note that the heading “Yasu” is a supplementary word that adds meaning to an independent word (verb) such as “easy to understand” and does not mean that the price is low. In addition, the heading 31 “na” is a supplementary word that adds a negative meaning to a idiom, such as “not strange”, and does not mean that there is / does not exist.

The part-of-speech 32 is a part-of-speech recorded for each word of the headline 31. The vocabulary is uniquely determined by the headline 31 and the part of speech 32. Therefore, by searching the headline 31 and the part of speech 32 using the word and the part of speech obtained from the morphological analysis as search keys, it is possible to obtain the effect term information 33, the function word information 34, the default affectiveness 35, etc., which will be described later.

The effect term information 33 is composed of two attributes, polarity information and strength. The polarity is an attribute indicating the emotional direction of the word of the headline 31, and the headword 3
When the affectiveness represented by the word 1 is positive, it is set to +, and when it is negative, it is set to −. The strength is a parameter that numerically represents the strength of emotionality, and is set in advance by the system designer. The larger the value, the greater the emotional strength. The direction and size of emotional
It can later be used to obtain quantitative information from qualitative and subjective sentences.

The affect term having a positive polarity may be called a positive affect term, and the affect term having a negative polarity may be called a negative affect term. By the way, which affect term becomes + or-depends on how to make a dictionary, so the positive or negative of an affect term is not a general one and is decided by the designer of the dictionary.

The function word information 34 is composed of two attributes: calculation (type) and strength. This is used to change the strength of the affect term when the function word acts on the affect term included in the affect expression. That is, using the strength set in the function word information 34 for the strength of the effect term, the function word information 3
The calculation set in 4 is performed. For example, when the type of operation of the function word “very” is multiplication and the strength is 2, the operation “2 ×” is applied to the strength of the effect term. So, if there is a word that is simply "interesting", the score for this word will be +2 points,
If the function word is acting like "very interesting", the score is + 2 × 2 = 4 points.

The default effect property 35 is heading 3
Set the default value for the affectivity of the word 1. As described above, one word may be either an affect term or a function word depending on the syntactic structure. For example, if there is an affective expression such as "This book is amazing",
"Sugo" is an affect term that evaluates the axial term "book". On the other hand, if there is an affective expression such as "This book is very interesting,"
It is a function word that emphasizes the affect term "interesting". Therefore, the affect term dictionary 22 holds both the information as an affect term and the information as a function word for such a word, and sets which meaning to use as a default in the default affectivity 35. ing. At the stage of giving effect information described later, the information set as default is added, but if necessary, the effect property is changed in the later parsing.
As described above, the affect term dictionary 22 serves both as an expression word database storing subjective evaluation words (affect terms) and a function word database storing function words. The subjective evaluation word is associated with a numerical value indicating the degree of emotional strength, that is, the degree of subjective strength, and the functional word is associated with a calculation method.

Returning to FIG. 1, the effect information adding section 12
Searches the effect term dictionary 22 for words corresponding to each morpheme that constitutes the morpheme string using the "heading" and "part of speech" as search keys, and if that word is registered in the effect dictionary 22, Affect information is added.
The affectity given at this stage is default information. Then, the morpheme sequence to which the effect information is added is output to the syntax analysis unit 13. The value set as the default value may be changed by the processing in the subsequent process. Alternatively, the affectiveness may be canceled and re-judged as being neither an affect term nor a functional word, or the affect term may become an affective expression by itself. Furthermore, there are cases where what was originally not said to be an affective expression becomes an affective expression by being combined with another word. In this way, the effect expression processing device 1
Does not extract the affect expression only by adding the affect information. In this way, the effect information adding section 12 constitutes a numerical value associating means for associating a numerical value with the subjective evaluation word and an operation associating means for associating the operation method with the function word.

The syntactic analysis unit 13 is syntactic analysis means for creating a clause list from the morpheme list. Creating a clause
Basically, an independent word and an adjunct word in the morpheme list are combined into a clause. The syntactic analysis unit 13 performs the same processing as that of a well-known syntactic analysis engine that is generally used, and also has the following two unique functions. (1) Perform affect action processing. Normally, in the syntax analysis, only the syntax tree is assembled, but the syntax analysis unit 13 changes the affectivity given by the effect information addition unit 12 according to the situation. In addition, in the present embodiment, the process of the affectability changing process is performed at the same time as the syntactic analysis, but this process and the syntactic analyzing process can be separated. In this case, it is possible to use a parsing system used in another system.

(2) Even if the structure of one sentence remains divided into a plurality of structures, each structure is accepted as a syntax tree. An ordinary syntactic analysis system aims to create one structure such as one syntactic tree and one semantic structure as a result of analyzing one sentence. If this fails, the parsing itself is considered a failure. However, the affective expression processing device 1 may process a sentence that is not necessarily linguistically sophisticated or a sentence that includes misuse of a word, such as a sentence posted on an electronic bulletin board on the Internet. Therefore, when the analysis reaches the end of the sentence, each sentence is accepted as a syntax tree even if the whole sentence does not have one structure, that is, even if it is divided into a plurality of structures.
However, the morpheme position information for indicating that the plurality of syntax trees are included in one sentence is stored and associated. Therefore, even if the syntax tree is not one, the order relation of morphemes can be recognized.

The function of the syntax analysis unit 13 will be specifically described below. The syntactic analysis unit 13 searches the syntactic analysis dictionary 23 that stores the grammar for performing syntactic analysis,
The morpheme sequence is analyzed to form bunsetsu. Although various methods of parsing have been proposed, here, as an example, a method using a binary relation will be described. This is to analyze a sentence structure using two adjacent morphemes of interest and the structure of a dependent.

When focusing on adjacent morphemes, the grammar is searched using the part of speech obtained as a result of the morphological analysis as a search key. When focusing on the structure of the dependency, the grammar is searched using the structure name (particle phrase, verb phrase, etc.) as a search key. In addition to this, the affectivity (information such as an affect term, a functional word, or neither) that is possessed thereafter can be used as a search condition.

As an example, referring to FIG. 4, the example sentence "I buy a book" is parsed. This example sentence does not need to be an affect action process. The following three grammars for parsing are stored in the parsing dictionary 23. Grammar 1) If there is a noun + particle, put the noun on the particle and make a particle phrase. Grammar 2) If there is a particle phrase + verb, apply the particle phrase to the verb to make a verb phrase. Grammar 3) If there is a particle phrase + verb phrase, apply the particle phrase to the verb phrase to make a verb phrase.

The result of morphological analysis of the case sentence is as follows. "I (noun) / ha (particle) / book (noun) / is (particle) / buy (verb)". Note that the verb includes an inflectional ending for simplicity, but there is a grammar that processes the verb stem and the verb inflection ending.

Step 1: Since the first morpheme "I" and the next morpheme "ha" satisfy the grammar 1, the noun "I" is applied to the particle "ha", and "Iwa" is made a particle phrase. Step 2: The syntactic analysis dictionary 23 is searched under the conditions of "I" and "Book", but since there is no grammar "particle phrase + noun", it is not possible to set a dependency relationship for both.
Therefore, grammar 1 is found by ignoring the first dependency structure and searching the syntactic analysis dictionary 23 for the next two morphological nouns "book" and particle "wo". Therefore, "book" is multiplied by "wo" and "book is" is used as a particle phrase.

Step 3: Next, the first two structures, that is, the particle phrase "Iwa" and the particle phrase "book", are parsed into the syntactic analysis dictionary 23.
Search with, but the corresponding grammar does not exist. Therefore, the first composition is ignored, and the process moves to the verb "buy" adjacent to the particle phrase "book" which is the second structure. When the syntactic analysis dictionary 23 is searched for the relation of these two terms, the grammar 2 matches. Therefore, "book" is put into "buy" and made into a verb phrase. Step 4: Adjacent structures are the particle phrase "I" and the verb phrase "buy a book". When this condition is searched in the syntactic analysis dictionary 23, the grammar 3 matches. Therefore, "I am" is used as a verb phrase by "buying a book". In this way, the sentence structure is analyzed. In the above example, the part-of-speech of the result of the morphological analysis is used, but the accuracy of the syntactic analysis can be improved by using the case pattern of the semantic element or the verb.

By the way, as described above, the effect expression processing apparatus 1 accepts the analysis result as the analysis result when the analysis result reaches the end of the sentence even if the sentence is not integrated into one structure. For example, if there is no grammar 3, the structure of step 3 described above cannot be transferred to the structure of step 4.
In this case, the structure of step 3 is accepted as the final form.

The syntax analysis method using the above binary relation is
It is easy for the system developer to understand intuitively, and since the binary relation is processed, it is easy to add the function unique to the effect analyzer (for example, the effect / action function). Further, when the structure of one sentence is not integrated into one, it is easy to expand the function such as accepting this as a plurality of structures. On the other hand, as another parsing method, for example, there is one using a phrase structure rule.
In order to speed up the processing, the algorithm is complicated like the LR parser, and its operation cannot be intuitively grasped. Therefore, it is difficult to add or expand the function. The L of the LR parser is (Left to
Right, where R is (Rightmost D
Emission) is represented.

Next, the case where an effect action process is performed at the time of parsing will be described. This process is a process of changing the affectivity of the morpheme or dependency currently focused on at the time of parsing. If the morpheme being processed by the syntactic analysis has the affect information, this morpheme may be either an affect term or a function word, such as "Sugo (i)". The default affectivity is temporarily designated by the affect information adding unit 12 for this morpheme, but it may be necessary to change this in some cases. Then, this change is described in the syntax analysis dictionary 23 as a grammar of the change method, and the change is performed according to this.

As an example, description will be made by using an example sentence "This book is very interesting". In this example, the inflectional ending of "sugoi" is a continuous form, which modifies the idiom of "interesting" to act as an adverb and acts as a functional word. In addition, adjectives representing other degrees have the same general effect. By the way, the default of the "wow" effect is set to the effect term. However, in the above example sentence, since it acts as a function word, it is necessary to switch the affectiveness of "Sugoi" from the affect term to the function word. Therefore, the following grammar 4 is described in the syntactic analysis dictionary 23. Grammar 4) Degree Adjective stem + Consecutive adjective Inflection, at the same time as making an adjective phrase, change the affectiveness from the default affect term to a functional word. The syntactic analysis unit 13 searches the syntactic analysis dictionary 23 for the grammar 4, and switches the affectiveness of the “sugoi” from the affect term to the function word. In this way, the syntactic analysis unit 13 has a function as a re-identifying unit that re-identifies the affect term and the function word.

Next, another example of the effect action processing will be described. The auxiliary phrase has the word "easy". This word is, for example, "This book is easy to read."
When a non-affect term adjective is concatenated before "yasu", it may mean something rather good as a whole. Therefore, the default is "easy".
The effect term information of polarity “+” is set to.

By the way, consider an example sentence "This package is fragile". It is assumed that the "break" is an affect term with a polar "-". in this case,
The effect of the auxiliary idiom “Yasu” is unnecessary.
Further, if left as it is, “break” and “easy” will be cut off in the extraction of the effect expression described later. Therefore, it is necessary to delete the affectiveness of “easy” or change it to a functional word. For that purpose, the following grammar 5 may be described in the syntactic analysis dictionary 23. Grammar 5) In the binary relation, when the word at the beginning of the sentence is an affect term and the word at the end of the sentence is an auxiliary adjective or an affect term, the latter's affectiveness is made a functional word.

In more detail, in order to apply the grammar 5, it is necessary that the inflectional ending of the word at the beginning of the sentence inherits the effect. For example, suppose that there is an example sentence that says "I'm irritated." And let's say "chi" is an affect term. As shown in FIG. 5, when this is related to the inflectional ending "chi",
Inherits the affectiveness of being an affect term. That is, if the "frustration" part does not inherit the affectiveness, the grammar 5 when checking "frustration" and "easiness"
Does not work, and both parties proceed to parse with an affect term.

The following grammars for other effect actions are considered. (1) Affectiveness / affect information is newly added to those that have no affectiveness (those that are not registered in the affect term dictionary 22). (2) Affectiveness-Delete them from the words to which the affect information is added. (3) Delete the affectiveness / affect information from all the analyzed results.

In the syntactic analysis using the binary relation described above, the front morpheme (or the dependent structure) is applied to the rear morpheme (or the dependent structure), and the structure name is given to the newly created structure. Although the grammar was constructed, in the case of using the binary relation, in addition to this, the grammar should be constructed by, for example, the one in which the back is put forward, the one in which the morpheme (or the dependent structure) in the front or the rear is discarded. You can also

The effect expression extraction unit 14 (FIG. 1) extracts the effect expression centering on one affect term by analyzing the dependency structure and the part-of-speech arrangement resulting from the syntactic analysis. Further, when the function word is combined with the affect term, the affect expression is extracted in the form of combining the affect term and the function word. For example, the example sentence "not very beautiful" is an affective expression centered around the affect term "beauty", which combines "beautiful" with the emphasized functional word "very" and the negative functional word "not". It is one effect expression. In this way, the affective expression extracting unit 14 constitutes an extracting means and a function word specifying means for specifying a function word.

FIG. 6 is a block diagram showing a more detailed structure of the effect expression extraction unit 14. The function of the effect expression extraction unit 14 will be described below with reference to the example sentence "Coloring is beautiful, but it is not so good compared to other products because it is sticky when applied." FIG. 7 shows an example of the result of parsing this example sentence. In addition, in order to avoid complication of explanation, the inflectional inflection endings are summarized as the inflectional stem.

The cutoff unit 41 cuts off one branch having a dependency relationship by using the result of the syntactic analysis, and the connection particle separation unit 42.
Output to. As a result, the following branch is obtained. "The color is beautiful but not good", "It's bad because I'm trying to put it on", "It's not good because it's sticky", "Not good compared to other products", "Not so good". Here, depending on the morpheme, it may be used in multiple branches, but in the end, the affect terms cannot be used in duplicate. At the stage of the cutout section 41, the branches are cut off temporarily by allowing the overlap.

The connecting particle separating unit 42 creates a branch separated at a portion where the connecting particle exists, from the result of branch cutting. As a result, the following branches of ~ shown in Fig. 7 are obtained. The color is beautiful, is intended to be attached, is sticky, is not good compared to other products, is not very good.For example, the branch "ga" of "color is beautiful but not good" is a connection particle, so it is separated. And "the color is beautiful"
It is what

The pattern extraction unit 43 (1) deletes a portion not containing an affect term, (2) decomposes it into a structure centering on one affect term, (3) finds and cuts a cutting pattern, (4) It has functions such as effect expression extraction. (1) Deletion of a portion that does not include an affect term In the example sentence of FIG. 7, a portion with a wavy line is an affect term, and a portion with an underline is a function word. As is clear from the figure, the branch that is supposed to be attached is a branch that does not contain an affect term, so this is deleted.

(2) Decomposition into a structure centering on one affect term Although the example sentence of FIG. 7 does not correspond to this, at this stage, for example, a plurality of affect terms and function words are arranged in a complicated manner as follows. There is a possibility that the branches of the pattern are left. K1 + K2 + A1 + A2 + K3 + A3 + K4, where K1 to K4 are function words and A1 to A3 are affect terms. In such a case, this function determines to which affect term each function word belongs by the part of speech, and decomposes it into a structure for each affect term. By this, for example, K1 + K2 + A1, A2 + K3, A3 + K
It can be divided into three structures centered on A1, A2, and A3 such as 4. That is, for example, K3 is a function word for A2, K1 is a function word for A1, and the like.

Next, an example of the judgment using the part of speech will be shown. Although the auxiliary adjective "no" is a functional word, it is attached to the immediately preceding morpheme because it adds an auxiliary meaning to the immediately preceding independent word, such as denying the previous adjective. For example, in the sentence "This beautiful book is not interesting", "being beautiful" and "not interesting" because "not" belongs to "interesting"
It is decomposed into two expressions. Also, in the sentence "This beautiful book is interesting", since "not" belongs to "beautiful", it is decomposed into two expressions, "not beautiful" and "interesting".

(3) Finding and cutting a cutting pattern This function is to cut a portion corresponding to a preset cutting pattern. An example of a cutting pattern is shown. If there is an affect term that is not related to the preceding function word between the function words, disconnect it immediately before the following function word. For example, suppose there is a sentence that says "It's neither beautiful nor interesting." If this sentence is simply decomposed, two affective expressions of "not beautiful" and "not interesting" are extracted, and the former becomes double negative. Here, if this cutting pattern is used, an affect term "interesting", which has no relationship in terms of word order, is sandwiched between "null" in the front and "null" in the rear, so just before the functional word in the rear, That is, it is possible to disconnect just before "not" of "not beautiful" and to extract two affective expressions "not beautiful" and "not interesting". In addition, from the sentence "neither beautiful nor interesting", "not beautiful"
Two affective expressions such as “not interesting” are extracted, and the analysis results of these two example sentences are the same.

In the case of the example sentence "Neither beautiful nor interesting", it is possible to recognize it as a parallel expression by performing a strict syntactic analysis. In this way, double negation can be eliminated in advance. However, the parsing engine used in the effect expression processing apparatus 1 sets a weak parsing condition in order to deal with unsophisticated and misused sentences such as sentences posted on electronic bulletin boards on the Web. And I aimed to make the parsing as successful as possible.

(4) Extraction of Affect Expression This function deletes a part which is not related to the effect expression in addition to the part deleted by the deletion function of the part not including the affect term of (1). Is. As a result, the following effect expressions are extracted. The coloring is beautiful → beautiful, is already deleted by the function of (1), it is sticky → sticky, not good compared to other products → not good, not so good → not good For example, in the case of " Chromophore has been removed because it does not contain the effect term, the effect term "beautiful"
However, the affect term "Yasu"
When the affect term is a supplementary word, as in the case above, it may be traced back to the preceding independent word and incorporated into a part of the affect expression.

The inclusion portion deletion unit 44 (FIG. 6) judges the inclusion relation between the effect terms and deletes the included affect term. For example, "not good" in
"Not very good". Therefore, delete. As a result, the following affect expression is finally extracted. Beautiful, (remove), greasy, (remove), not very good

The score calculator 45 gives a score for each affect expression. If the affect expression does not have a function word, the polarity and the strength assigned to the affect term in the affect term dictionary 22 are simply added.
When the affect term has a function word, the effect term is calculated by using the operation and strength given to the function word in the affect term dictionary 22 and the polarity and strength of the affect term.
If there are a plurality of function words and the operation is only addition / subtraction or multiplication / division, there is no problem. For example, the following calculation procedure can be considered. (1) Calculate from the functional words that are close to the effect term. (2) Calculate from the front. (3) Calculate from the back. Note that the calculation is not limited to addition, subtraction, multiplication and division, and for example, a predetermined function may be set and the calculation may be performed using this.

In this way, the score calculation section 45
A score can be calculated for each affect expression.
By the way, first, the effect term having a polarity of + in the affect term dictionary 22 is defined as a positive effect term, and the effect term having a polarity of − is defined as a negative effect term. Therefore, correspondingly, an effect expression having a score of + will be referred to as a positive effect expression, and an effect expression having a score of-will be referred to as a negative effect expression.

Next, an example of calculating the score of the affect expression will be shown. Suppose there was an expression "not very beautiful". Then, it is assumed that the following registration is made in the affect term dictionary 22. "Very" ... functional words of emphasis (operation = x, strength =
2), "Beautiful" ... Affect term (polarity = +, strength =
1), “none” ... Negative function word (operation = ×, strength = −)
1), in this case, "very / beautifully / not" score = 2 x 1
X (-1) =-2, and the score of the affective expression is-
It becomes 2. Therefore, this expression is a negative affect expression.

In addition, an effect expression in which the score is 0 and is neither positive nor negative is also conceivable. For example, suppose there was an expression "Is it beautiful?"
In addition, it is assumed that “kana” is an ambiguous expression and is registered as a function word having a meaning of weakening (operation = ×, strength = 0) in the effect term dictionary 22. in this case,
The score of “Is it beautiful?” Is 1 × 0 = 0.

The term positive / negative is used as follows, for example. The total score of one sentence's affective expression is used as "this sentence is positive" or "this sentence is negative". Total the scores of one or more articles with the same evaluation subject,
"The comfort of this lipstick A has a positive score."
Or "lipstick C color has a negative score"
And so on. For different articles with the same evaluation subject,
It can also be used for the score of an article, such as "Article 1 is positive for product A, but article 2 is negative".

The axis term dictionary 24 is a database in which words that may become axis terms and evaluation axes for classifying these words are registered in advance. FIG. 8 is a diagram showing an example of a logical configuration of the axial term dictionary 24 customized for lipstick. The axis term dictionary 24 is composed of three items: headline, part of speech, and evaluation axis. The headline is
The assumed axis terms for the evaluation subject are registered. For example, in the case of a lipstick axis term, it is a word that the user of the lipstick would use to evaluate the lipstick. For example, when an article such as “This lipstick is glossy”, “Good color tone”, and “Color development is not good enough” is assumed, gloss, color tone, color development, etc. can be considered as axial terms. The part of speech of the axial term dictionary 24 represents the part of speech of the axial term, and is usually a noun.

The evaluation axis is a classification of preset aspects (attributes) to be evaluated by the evaluation subject. Assume that the lipstick has an evaluation subject. As aspects of the user's evaluation of the lipstick, for example, color, odor, comfort, package, mochi, price, ... For example, comparing the sentence "This lipstick has a beautiful color development" with the sentence "This lipstick has a beautiful design", the aspects evaluated by these sentences are clearly different. It is assumed that the former sentence evaluated the aspects related to color gloss, and the latter sentence evaluated the aspects related to packaging. Therefore,
The axis term “coloring” can be classified into the evaluation axis “color gloss”, and the axis term “design” can be classified into the evaluation axis “package”. In the axis term dictionary 24, the aspect of the evaluation of the evaluation subject, that is, the evaluation axis is set in advance in this way, and a database is stored to which evaluation axis each axis term belongs. That is, an evaluation attribute database in which the axis term is associated with the evaluated attribute of the evaluation subject is configured. The axis term determining unit 16 described later can search the axis term dictionary 24 using the headline and the part of speech as keys, and acquire the evaluation axis of the axis term.

The axis term determining unit 16 determines the axis term in the branch and further determines the evaluation axis of this axis term. That is,
The affective word identifying means for identifying the axis term and the evaluation attribute associating means for associating an attribute with the identified affective word are configured. Further, by associating the emotion word with the attribute by the associating means, it is possible to specify the attribute (evaluation axis) in which the emotion expression word expresses the emotion, and thus also has a function as the attribute specifying means. In the following description, the expression "one branch" will be used in the following meaning. To the point where you can follow the direct parent (only one), and also to the child (there may be multiple) to one where you can follow one of the branches, this is called a branch. To do. In the present embodiment, the case where the axis term is determined from the noun in the same branch as the effect expression will be described. However, the present invention is not limited to this, and the axis term is determined from a branch different from the branch including the effect expression. Sometimes the term is decided. For example, if there is a sentence "This scent. I've always loved it.", The axis term of the affective "love" in the second sentence is "fragrance" in the first sentence. In this way, the axis term may be extracted from a sentence different from a sentence having an affect expression.

For example, from the example sentence shown in FIG. 7 "The coloring is beautiful, but it is not so good compared to other products because it is sticky when attached." can get. Color is beautiful but not good

When determining the axis term, the axis term determining unit 16 determines, as the axis term, a noun satisfying the following conditions among the effect expressions. (1) A noun on the same branch as the affect term, "Noun +
No. + Affect term "and the noun is registered in the axis term dictionary 24. For example, the sentence "deliciousness of cooking" satisfies the above pattern, so "cooking" is in the axial term dictionary 2.
If registered in 4, "cooking" is the axis term. (2) A noun that is in the same branch as the affect term, forms a pattern of "affect term + no + noun", and the noun is registered in the axis term dictionary 24. For example, the sentence “impressive service” satisfies the above pattern, so if “service” is registered in the axial term dictionary 24, “service” is the axial term. The patterns (1) and (2) in which the axial term determining unit 16 determines the axial terms have been described above, but these are examples, and the axial terms may be determined using other patterns.

Further, after determining the axis term, the axis term determining unit 16 determines the evaluation axis of the axis term in the affect expression. In the present embodiment, the evaluation axis is mainly determined by the following two methods. (1) If a morpheme recognized as an axis term (here, only nouns, but not limited to nouns) is registered in the axis term dictionary 24, the corresponding evaluation axis is recorded. Therefore, the evaluation axis is determined using this. (2) If the affect term itself has a default evaluation axis, use that information. Here, the default evaluation axis is an evaluation axis specified by this information when the affect term itself has information about what to evaluate, and the affect term dictionary 2
Register in 2 etc. For example, with lipstick "shiny"
Speaking of which, it usually means the evaluation axis "color gloss", and it is almost unlikely that the appearance of the package is "shiny". In this way, the affect term itself may be able to provide the evaluation axis.

As described in (1) above, when the axis term is registered in the axis term dictionary 24, the evaluation axis can be obtained by searching with the axis term as a key, but the axis term is the axis term dictionary. There is a possibility that it is not registered in 24. In that case, the evaluation axis can be obtained by the method (2). The case (2) will be described in more detail below with an example.

(Example 1) Consider the expression "red is beautiful". Now, it is assumed that the axis term "red" is not registered in the axis term dictionary 24. Usually, "beautiful" is evaluated as "red" because various aspects are evaluated.
Unless is registered, it is not possible to specify that the beautiful thing is "color gloss". Since neither "red" nor "beautiful" has information about the "evaluation axis", it cannot be finally specified.

(Example 2) Next, consider the expression "red is vivid". Similar to the example 1, it is assumed that the axis term “red” is not registered in the axis term dictionary 24. Here, if it is clear that the adjective "colorful" is an expression that evaluates the "color gloss" aspect of lipstick, "red" also represents a color classification, etc.
It seems that this expression evaluates color gloss. That is, although the evaluation axis is not known from the axis term “red”, an attribute designated by default as the axis term “colorful” can be used.

(Example 3) Consider the expression "it is very vivid". Here, even if you look at the sentences before and after, "color"
It is assumed that there is no axial term such as "coloring".
In this case as well, it is almost the same as in Example 2, and it is finally determined that there is no axial term. However, the side (evaluation axis) being evaluated can be specified as “color gloss”. That is, "color vividness" holds information that the effect expression itself must evaluate "color gloss", and when the evaluation axis (attribute) of the axis term is not clear,
Use "Color Shine" by default. "Beautiful" is so versatile that it cannot be determined to be "color gloss", so it cannot be processed as in Example 2 without the default aspect.

When there is no axis term, for example, using the default evaluation axis of the affect term,
It is possible to determine which evaluation axis the affect expression belongs to. Also, the axis term dictionary is 24
Even when the evaluation axis information is registered in, it is possible to use the default evaluation axis on the affect term side before using the evaluation axis information.

From the viewpoint of system tuning, the effect term dictionary 22 may be composed of a general-purpose dictionary that does not specify a field and a specialized dictionary that specializes in a specific field such as "lipstick". In this case, the default evaluation axis is registered in the specialized dictionary. This is because the concept of the evaluation axis is quite artificial and therefore changes depending on which field the effect expression processing apparatus 1 is applied.

The output unit 15 is the effect expression extraction unit 14
Is an output means for outputting the extracted effect expression and score, positive / negative distinction, and the like. The output data is stored in the storage device and can be used by various application software and the like. For example, a score for each product (title) can be calculated from the result of the questionnaire survey.

FIG. 9 is a diagram showing an example of the hardware configuration of the effect expression processing apparatus 1. In the present embodiment, the effect expression processing device 1 is configured using a personal computer. The effect expression processing device 1 is
To the control unit 51, an input device 55 via a bus line 58,
An output device 56, a communication control device 57, a storage device 63, a storage medium drive device 60, an input / output interface 59, etc. are connected and configured.

The control section 51 includes a CPU 53, a ROM (Re
ad Only Memory) 52, RAM (Ran)
Dom Access Memory) 54 and the like. The control unit 51 operates in accordance with various programs such as an affect expression processing program, extracts various effect expressions from various articles, and calculates various scores such as calculating these scores. Also, the control unit 5
1 also controls the entire effect expression processing apparatus 1.

The CPU 53 is a central device of the control unit 51, loads a program from the ROM 52, the storage device 63, or a storage medium driven by the storage medium drive unit 60, and controls the control unit 51 according to the program. To do.

The ROM 52 is a read-only memory that stores various programs, data and parameters for the CPU 53 to perform various calculations and controls. ROM5
Reference numeral 2 denotes a non-volatile memory, which retains stored contents even when the power supplied to the ROM 52 is zero. Further, since the ROM 52 is a read-only memory,
Normally, no data is written.

The RAM 54 is a random access memory used as a working memory by the CPU 53.
The CPU 53 can write and erase programs and data in the RAM 54. In the present embodiment, the RAM 54 can secure an area for extracting the effect expression and calculating the score.

The input device 55 is composed of an input device such as a keyboard and a mouse. The keyboard is
This is a device for inputting information such as characters and numbers to the effect expression processing device 1. The keyboard is composed of keys for inputting kana and English characters, ten keys for inputting numbers, various function keys, cursor keys and other keys. The keyboard is used, for example, when the user inputs a login ID or password for logging in to the affective expression processing apparatus 1 or operates a program.

The mouse is a pointing device. GUI (Graphical User Inte)
When operating the affective expression processing device 1 by using a rface) or the like, it is possible to input predetermined information by clicking a button or icon displayed on the display device with a mouse.

The output device 56 is composed of, for example, a display device and a printing device. The display device is a device for presenting information on a screen, which is configured by, for example, a CRT (Cathode Ray Tube) display, a liquid crystal display, a plasma display, or the like. The display device can display a keyboard or mouse input result, a screen displayed by the program, and the like.

The printing device is a device for printing the execution result of the program on a printing medium such as paper. For example, when the program is an editor, the edited sentence can be printed, and when the program is a database, data can be printed. The printing device is composed of various printer devices such as an inkjet printer, a laser printer, a thermal transfer printer, and a dot printer.

The communication control device 57 is a device for connecting the effect expression processing device 1 to a network such as the Internet via a communication line (not shown), and is composed of a modem, a terminal adapter and other devices. The communication control device 57 is connected to, for example, the Internet, a LAN (Local Area Network), or the like, and can transmit / receive data to / from a server device or the like connected to these networks.
The communication control device 57 is controlled by the CPU 53 and transmits and receives according to a predetermined protocol. The affection expression processing device 1 can access the electronic bulletin board via the communication control device 57 to collect articles.

The storage device 63 is composed of a readable / writable storage medium and a drive device for reading / writing programs and data from / to the storage medium. A hard disk is mainly used as the storage medium, but other storage media such as a magneto-optical disk, a magnetic disk, and a semiconductor memory can be used instead.

In the storage device 63, storage areas such as the program storage section 61 and the data storage section 62 are secured. The program storage unit 61 stores an affect expression processing program, a browser, and an OS (Operating System) which is basic software for operating the effect expression processing device 1.
tem), a communication program for controlling the communication control device 57, and enabling communication via the Internet. In the data storage unit, the article database 20,
The morphological analysis dictionary 21, the effect term dictionary 22, the syntax analysis dictionary 23, and other data are stored.

The storage medium drive unit 60 is a drive unit for driving a removable storage medium to read / write data. Examples of removable storage media include magneto-optical disks, magnetic disks, magnetic tapes, semiconductor memories,
There is a paper tape punched with data, a CD-ROM, etc. Note that CD-ROMs and paper tapes can only be read. Data for backup can be written to the storage medium via the storage medium driving device 60, and application software can be installed from the storage medium.

The input / output interface 59 is, for example,
It is composed of a serial interface and other standard interfaces. By connecting an external device corresponding to the interface to the input / output interface 59, the function of the effect expression processing device 1 can be expanded. Examples of such external devices include a storage device such as a hard disk, a speaker, and a microphone.

FIG. 10 is a flowchart showing a procedure in which the affective expression processing apparatus 1 extracts an affective expression from an article. First, the CPU 53 reads articles relating to a predetermined field from the article database 20 and loads the RAM 5
4 (step 10). At this time, for example, fields such as lipstick, inn, and hot spring are assumed to be designated by the user in advance. Further, the fields can be configured in a hierarchical structure, for example, lipstick, perfume, ...
It may be configured to provide child fields such as.

Next, the CPU 53 reads the article from the RAM 54, analyzes the morpheme, and generates a morpheme string (step 20). Next, the CPU 53 refers to the affect term dictionary 22 and, if a word corresponding to each morpheme is registered in the affect term dictionary 22, adds the affective information to each morpheme (step 25). next,
The CPU 53 syntactically analyzes the morpheme string to which the effect information is added (step 30). At this time, the CPU 53
Performs affect action processing and converts the affect information if necessary. Even if a plurality of sentence structures are obtained from one sentence, the plurality of sentence structures are accepted as the analysis result.

Next, the CPU 53 extracts an affect expression using the syntax analysis result (step 35). CP
U53 deletes a part not containing an affect term, decomposes into a structure centering on one affect term, finds and cuts a cleavage pattern, when extracting an affect expression.
Affected expression extraction, etc. are performed. Next, the CPU 53
Calculates a score for the extracted affect expression. As a result, information such as the score of each affect expression and whether it is a positive affect expression or a negative affect expression can be obtained.

Next, the CPU 53 determines the evaluation axis of the axis term with reference to the axis term dictionary 24 (step 4).
2). If the axis term is not in the axis term dictionary 24, the default axis is adopted.

Next, the CPU 53 stores the analysis result in the RAM 5
4 or to the storage device 63 or the like. The output data is composed of the extracted affect expression, the score of the affect expression, whether the affect expression is positive or negative, and data that specifies the article from which the extraction is performed. Next, the CPU 53 determines whether or not the effect expressions have been extracted for all articles belonging to the relevant field, and if not, (step 5)
0; N), the process returns to step 10, the next article is read from the article database 20 and processed, and when all articles have been processed (step 50; Y), the processing ends.

Next, one application example of the effect information extraction device 1 will be described. 2. Description of the Related Art Conventionally, there are many sites that perform product evaluation (including specific products such as electric appliances and cosmetics, as well as evaluation of tourist spots such as movies and hotels) on the Web. However, these evaluation methods are largely
Classified into one. (1) Those evaluated by experts in advance. (2) The ranking is decided by the vote of the user. (3) A list of articles describing comments from users and experts, and a calculation of these comments on an electronic bulletin board.

These methods have the following disadvantages. In the case of (1), evaluation is made by the site creator's discretion, and the user's judgment is not reflected at all. In the case of (2), the majority of votes are about the overall judgment of the product, and there are not many detailed votes for various aspects of the product. In the case of (3), unless the article is read in detail, it is not possible to know the evaluation points of the product or whether the evaluation is high or low.

Therefore, an article or the like posted on an electronic bulletin board or the like is processed by the affect expression processing device 1 and the score of each affect term is statistically processed to obtain a large number of evaluations of various aspects of the product. You can For example, when the field is lipstick, it becomes possible for lipstick A to obtain data such that “comfort” is +20 and “coloring” is −1.

As described above, in the first embodiment, the following effects can be obtained. Expressions including emotionality such as subjective evaluation can be automatically extracted from articles. The emotional strength can be quantified. In the known normal parsing, when one sentence finally becomes one structure (syntax tree), the succeeding process is continued, but the parsing proposed in this patent uses one sentence. Even if the final result does not form one structure and remains in a plurality of structures, it can be regarded as successful and the next process can be performed. The relationship between the affect term and the function word can be determined, and the affect term and the function word can be set as an affect expression. The effect expression score can be calculated using the polarity and strength, which are the attributes of the affect term, and the operation and the strength, which are the attributes of the function word.

Further, the axis term which is the direct object of the evaluation is specified, and in the case of a specific field,
The evaluation axis can be used to classify what aspect the axis term represents with respect to the subject of the sentence. In addition, by combining these functions, the emotionality of the article is quantified as a score, and the emotionality is similarly quantified as a score for a particular aspect of the subject of the sentence, or A system that provides emotionality and axis terms in a list and provides them to the user to check specifically in what expression the subject is evaluated and to classify sentences by that expression. Can be provided. Further, the system of this embodiment is a basic library. Further, it is assumed that the user uses some kind of application software, and in that case, the application software is an API (ApplicationPro) provided by this system.
(Gramming Interface), this system can be called to request processing such as effect expression extraction, and the result can be obtained.

In the present embodiment, the effect term dictionary 22 is configured to correspond to all articles, but the present invention is not limited to this. It can also be configured to use a specialized affect term dictionary specialized for a field. Although the first embodiment of the present invention has been described, the present invention is not limited to the described embodiment, and various modifications can be made within the scope of the claims.

(Second Embodiment) In recent years, the number of sites for evaluating products and services on the Web has increased, and has been attracting attention. Consumer evaluations and opinions (so-called word-of-mouth information) posted on these sites are useful information for consumers because they convey the aspects that consumers want to know about products and services in an objective and concrete manner. Yes, again
For companies, it is useful to collect and use information from these sites that have a great social impact.

By the way, these sites had the following drawbacks. For example, in the case of posting the evaluation performed by an expert in advance, the evaluation of the product / service is performed by the site creator, and the user's judgment is not reflected.

[0105] Further, depending on the user's voting order, etc.,
Most of the people who evaluate services are those who vote on the overall judgment of a product, and it is unclear which aspect of the product or service is evaluated. For example, when evaluating a restaurant, as an aspect of evaluation, for example, taste, atmosphere, attitude of employees, price, etc. should be evaluated for each, but if they are integrated, The evaluation for each of these aspects is lost. In addition, voting sites use a simple calculation method such as dividing the score by the number of evaluators to determine the ranking, and the results are displayed only by showing the final score.・ It is difficult to understand the degree of reaction to services.

[0106] On a site that lists articles in which users or experts simply write comments in a bulletin board format, the amount of documents is large, and unless the articles are read in detail, the evaluation points and the high or low of the products / services are high or low. Is hard to understand.

Therefore, in the present embodiment, the comments of general consumers and experts are automatically analyzed, the comments are classified into evaluation axes, and the scores are displayed, so that the evaluation in each aspect of the product / service is performed. So that the user can easily recognize.

FIG. 11 is a diagram showing an example of a network configuration in this embodiment. The client terminal devices 2, 2, 2, ... After that, the client terminal devices 2, 2,
, 2 ... are simply referred to as the client terminal device 2 unless otherwise distinguished.

The client terminal device 2 is equipped with a browser, and from a server device or the like connected to the Internet 3, for example, HTML (Hypertext M
It is possible to receive a file in the arkup Language format and display the screen defined in the HTML file on the display. The client terminal device 2 is also equipped with an input device such as a keyboard and a mouse, and clicks a predetermined button on the screen,
The input information can be transmitted to the server device by inputting a character in a predetermined blank space.

The information processing server 5 is a server device having an effect expression extraction function, and can perform data communication with the client terminal device 2 via the Internet 3. When communicating, for example, TCP / IP (Tra
nsmission Control Protocol
/ Internet Protocol).

FIG. 12 is a diagram showing an example of the logical configuration of the information processing server 5. The same components as those in the first embodiment will be described with the same numbers. The effect analyzer engine 80 is the input unit 1 of the first embodiment.
0, morphological analysis unit 11, effect information addition unit 12, syntax analysis unit 13, effect expression extraction unit 14, axis determination unit 1
6, a module composed of the output unit 15. The effect analyzer engine 80 uses the morphological analysis dictionary 2
1, the article acquired from the article database 20 is analyzed while referring to the effect term dictionary 22, the syntactic analysis dictionary 23, and the axis term dictionary 24, and the affect expression and the score are output to the output engine 81.

The output engine 81 creates screen data to be displayed on the client terminal device 2 in an HTML file or the like while referring to the data of the effect information database 87, and the effect analyzer engine 8
For example, the effect information database is updated using the information acquired from 0. The list screen data 83 is data for displaying a list screen displaying titles for each field such as lipstick, perfume, hot springs, .... In this screen, for example, when a lipstick is selected, a list of brands of lipstick is displayed, and a specific brand can be selected. When the user selects a particular issue, the title screen for this issue is displayed.

FIG. 13 is a diagram showing an example of the list screen. The list screen is the one in which the list screen data 83 transmitted from the information processing server 5 to the client terminal device 2 is displayed on the display by the client terminal device 2 using a browser. The list screen has a field selection area 120. The field selection area 120 is, for example, “lipstick”,
This is an area for selecting a field to be evaluated among cosmetics such as “mascara”, “gloss”, and “perfume”.

The field setting area 120 is a drop-down menu, and the user can select a desired field from each field prepared in the menu.
In the drop-down menu, when the pointer 124 is clicked with the mouse or the like and the triangle at the end of the field setting area 120 is displayed, a list of fields of selection candidates is displayed, and the selected one is input. is there. In the present embodiment, cosmetics are used as an example, but the present invention is not limited to this, and any inn, hot spring, restaurant, or any other object to be evaluated may be used.

Axis selection area 121 is field selection area 1
This is an area for selecting an evaluation axis for evaluating an evaluation subject belonging to the field selected in 20. The axis selection area 121 is also a drop-down menu, and the user can select a desired one from the displayed list of evaluation axes. The list of evaluation axes displayed using the drop-down menu function in the axis selection area 121 changes according to the field selected in the field selection area 120. this is,
This is because the evaluation axis differs depending on the evaluation subject. For example,
When the field is “lipstick”, the evaluation axes include “color gloss”, “mochi”, “smell”, “comfort”, ..., “Other”, “general” and the like. "Comprehensive" is an evaluation axis that is the sum of scores on all evaluation axes.

The normalization selection area 122 is a radio button for selecting whether or not to normalize the data to be displayed. When the data to be displayed is normalized, for example, the scores of the evaluation subject having a large number of articles and the evaluation subject having a small number of articles are adjusted so that they can be compared with each other. Without normalization, data obtained by simply adding the scores of each evaluation subject is displayed.

The execute button 123 is a button for starting a search, and is activated when the pointer 124 is moved to the execute button and clicked. When the execute button 123 is actuated, the information processing server 5 moves from the client terminal device 2.
The information such as the field selected by the user, the evaluation axis, and whether or not to normalize the data is transmitted. And
In the information processing server 5, the output engine 81 searches the effect information database 87 and the article database 20 based on these data, and generates ranking data to be displayed on the client terminal device 2. The generated ranking data is transmitted to the client terminal device 2.

In the ranking area 126, the ranking data generated by the output engine 81 is displayed. In the ranking area 126, products (titles) in the field selected in the field selection area 120 are sorted and displayed in descending order of score on the evaluation axis selected in the axis selection area 121. For example, when the field is “lipstick” and the evaluation axis is “color gloss”, the lipsticks of each company are displayed in the order of high color gloss score.

The ranking area 126 is composed of a rank section, a title section, a points section, a score bar section and the like. In the rank column, the ranking is displayed by numerical values, and in the title column, the name of the product (title) and the maker name of the evaluation subject are displayed. In the points column, the scores obtained for each evaluation axis selected in the axis selection area 121 for each product (title) are displayed. The score is
There is a positive score for positive evaluations (expressed as a positive number) and a negative score for negative evaluations (represented as a negative number). It In the score bar column, the score acquired by each product (title) is displayed so that a positive score and a negative score can be distinguished.

The contents displayed in the point column and the score bar column differ depending on whether or not normalization is made by selection in the normalization area 122. In the case of normalization, the score value is adjusted in consideration of the difference in the amount of articles for each product. On the other hand, when normalization is not performed, the score is displayed without adjusting the score without considering the difference in the amount of articles for each product.

In the ranking content area 125, fields related to ranking, evaluation axes, presence / absence of normalization, etc. are displayed. A scroll bar 127 is provided on the right side of the ranking area 126, and the ranking area 126 can be scrolled by operating this with a mouse or the like. When the title column of the ranking area 126 is selected by clicking with the mouse, the title screen of the product (title) is displayed.

Returning to FIG. 12, title screen data 84
Is data for displaying the title screen of the brand selected on the list screen on the client terminal device 2. The title screen displays the score for each evaluation axis of the product (brand) selected on the list screen.

FIG. 14 shows an example of the title screen when the field is lipstick and the product is "quatroup". The title screen 130 is composed of an evaluation item column, a points column, a score bar column, and the like. Each evaluation axis is displayed in the evaluation item column. The point column displays the product (title) in which the positive score and the negative score obtained for each evaluation axis are combined.

In the score bar column, the product (title) concerned
However, the total value of the positive score and the negative score acquired for each evaluation axis is displayed in a distinguishable manner by a bar. The display of the score column and the score bar column also differs depending on whether or not the data normalization is selected as in the case of the list screen. With regard to this product, there are many positive affective expressions regarding “color gloss” and “price”, and the points are positive, and it is understood that these aspects are highly evaluated. Also, regarding "mochi", there are many negative affective expressions, and the points are negative, and it can be seen that evaluation is low for this evaluation aspect.

Returning to FIG. 12, the detailed screen data 86 is data for displaying the full text of the article of the product. The title screen 99 is provided with a detail button (not shown), and when the user selects it with a mouse operation or the like, the output engine 81 uses the data such as the effect information database 87 and the article database 20 described later for details. The screen data 86 is generated. Then, this data is transmitted from the information server 5 to the client terminal device 2 used by the user. Then, in the client terminal device 2, the CPU can display the detailed screen on the display by using the browser.

FIG. 15 is a diagram showing an example of a detailed screen displayed on the client terminal device 2. Detail screen 13
5, the entire article about the product is posted in the article area 138. Then, in these articles, positive affect terms and negative affect terms appearing in the articles are displayed in a distinguishable manner. In the figure, positive affect terms are shown with wavy lines, and negative affect terms are shown with underlines. That is, the positive terms are "Sugo" and "Cute", and the negative terms are "No" and "Sorry". Alternatively, the positive affect term can be represented in blue and the negative affect term can be represented in red. This is because the output engine 81 outputs articles to the article database 20.
It is created by searching the effect information database 87 for the effect term and its positive / negative property appearing in the article, as well as searching from the effect information database 87.

The details screen 135 is the same as the previous article selection section 136.
And the next article selection unit 137 are provided, and when the user selects these with a mouse operation or the like, the previous article or the next article is displayed. Below the article area 138, an affect term list area 139 is provided. In this area, the analysis result of the effect term that appears in the article is displayed. This area, on the left side, shows the analysis results for the extracted affect terms,
The right side shows the analysis result regarding the axis term evaluated by the affect expression.

The sentence number, the effect expression, and the score item on the left side represent the number of the sentence in which the effect expression is present, the extracted effect expression, and the score of the effect term calculated by the effect analyzer engine 80. . For each item of sentence number, evaluated word, axis type on the right side, the number of the sentence in which the object evaluated by the affect term in the left column appears, the axis term of the affected expression, the axis term or It shows the evaluation axis judged from the effect term. For example, 2
The data on the line is that the affect term "kawaii" is
It has an axis term of "color" and a score of 2.0,
The affect term and the axis term are included in the first sentence.

Returning to FIG. 12, the effect information database 87 is a database of information specifying an article, an affect term that appeared in the article, the appearance position of the affect, the score of the affect expression, the appearance frequency of the affect term, and the like. Is. These data are generated by the output engine 81 from the output of the effect analyzer engine 80. Target articles and products (titles) can be added later.

In the above example, the title screen is displayed for each title, but data relating to a plurality of titles is displayed as one title.
It can also be displayed on one screen so that multiple products can be compared. FIG. 16 shows a screen transition diagram in that case. 1
For example, in the case of displaying title 1 → title 2 for each title, for example, dividing the screen and setting the title to 3
Allows the user to select from the list screen when displaying about one at a time. When displaying multiple titles, select from the title list screen to be displayed.

According to this embodiment, the following effects can be obtained. When there are a large number of articles in which opinions about a certain object (such as a product or service) are described, the affective information processing apparatus 1 according to the first embodiment is used to automatically determine the emotionality expressed in the articles. It can be parsed and scored so that the contributor's rating of the article can be recognized without reading the article.

Further, by setting the evaluation axis, it is possible to evaluate various aspects of the evaluation subject. As a result, the details of the evaluation are further detailed, and it is possible to clearly understand which part of the target is evaluated and how. In addition, this allows us to compare and examine the advantages and disadvantages of multiple target cases from various aspects. Furthermore, by visualizing the evaluation amount for each object, it is possible to understand how much reverberation the object has collected.

Further, it is possible to understand how much opinions are gathered for each object, which enables comparison and comparison of a plurality of objects.
The examination can be done easily. Furthermore, the normalization of the score allows relatively accurate comparison of evaluations for a plurality of evaluation subjects regardless of the number of articles. Also,
It can provide useful information to general consumers, and can provide useful information to companies such as marketing and stopping reputation damage.

(Third Embodiment) First Embodiment,
In the affect term dictionary 20 of the second embodiment, the attributes of each affect term, that is, the polarity and the strength are
It was fixed at a value preset by the dictionary designer. For this reason, it is necessary to use it fixedly or to tune it by hand manually on a regular basis. In that sense
There is a possibility that the evaluation will be limited to within the criteria intended by the site designer. Therefore, in the third embodiment,
The attribute of the affect term and the evaluation result by the user's voting result are organically linked, and the attribute of the affect term is optimized and automatically updated.

The network configuration of this embodiment is the same as that of the second embodiment, and the client terminal devices 2, 2, 2, ... Are arranged so as to be connectable to the information processing server 5 via the Internet 3. It is set up. After that, the second
The components corresponding to those in the embodiment will be described with the same numbers.

FIG. 17 is a diagram showing an example of the logical configuration of the information processing server 5. The same components as those in the first embodiment and the second embodiment will be described with the same reference numerals. The configuration of the effect analyzer engine 80 is the second
The description is omitted because it is the same as the embodiment. The effect analyzer engine 80 uses the morphological analysis dictionary 21,
The article acquired from the article database 20 is analyzed with reference to the affect term dictionary 22, the syntactic analysis dictionary 23, and the axis term dictionary 24, and the affect expression and the score are output to the output engine 81. First embodiment and second
The difference from the embodiment is that the effect term dictionary 22
And the polarity and strength of the effect term is the dictionary adjustment engine 8
It is a point updated by 2.

The output engine 81 creates screen data to be displayed on the client terminal device 2 in an HTML file or the like while referring to the data of the effect information database 87, and the effect analyzer engine 8
For example, the effect information database is updated using the information acquired from 0. The list screen data 83 is the second
This is the same as the embodiment. For example, when “lipstick” is selected in the field and “color gloss” is selected in the evaluation axis, a list of lipsticks is displayed ranked by the color gloss score. Then, the product can be selected by clicking on a specific product name. When the user selects a specific product, a title screen regarding this product is displayed.

As the title screen data 84, the title screen of the brand selected on the list screen is displayed on the client terminal device 2.
This is the data to be displayed in. The title screen shows 1 evaluation subject (1 group) in 1 field (1 group that is the same evaluation subject).
It is a screen that displays information about products and services that are subject to evaluation, such as products, and displays the score of a predetermined product, and also has a function of accepting votes from users.

FIG. 18 shows an example of the title screen when the field is lipstick and the brand is "Shigure". The title screen 99 is one in which the information processing server 5 transmits the title screen data 84 to the client terminal device 2, and the CPU of the client terminal device displays it on the display according to the browser software. The title screen 99 includes an evaluation area 95, a voting status area 96,
The voting column 97 includes a voting button 98 and the like.

The evaluation area 95 is further composed of evaluation items, evaluation points, a score bar and the like. The evaluation items are the items for evaluating the product, listed by item. For example, in the case of lipstick, "color gloss", "mochi", "smell", "comfort", "price", "package" Etc. These items are preset by the system designer according to the product. The point is the sum of the scores of the effect expressions included in the articles of the product stored in the article database 20. For example, when an article has an affective expression such as "colorful gloss", the score of this affective expression is added to the "color gloss" point.

The score bar is a bar representing the total value of the positive score and the negative score acquired by the product, for each evaluation item. A positive score means a positive score, and a negative score means a negative score. For example, the lip gloss “Shigure” may be positively evaluated in some articles to give a positive score, or negatively evaluated in other articles to give a negative score. Then, for example, a positive score is represented by a blue bar and a negative score is represented by a red bar. In this way, by displaying the positive score and the negative score in the distinguishable bar, these aggregated values can be visually expressed. Then, the user can intuitively grasp the evaluation of “Shigure” at a glance.

In the voting area 97, three selection branches of “evaluate”, “not evaluate”, and “don't know” are set corresponding to the evaluation items of the evaluation area 95. The voting area 97 is a field in which the user inputs his / her own evaluation for each evaluation item, and can select either “evaluate”, “not evaluate”, or “don't know” for one item. By default, I don't know
Is selected. In this way, voting is possible not only when the evaluation subject (such as a product) is considered well, but also when it is considered bad. That is, you can vote either positively or negatively.

The selection can be performed using a mouse, for example. When the user operates the mouse provided on the client terminal device 2, the pointer 100 can be moved on the screen. Then, when the pointer is moved to the desired section of the voting area 97 and the left button is clicked, the section can be brought into the selected state. As with the radio button, one of “evaluate”, “not evaluate”, and “don't know” for one evaluation item
When one is selected, the other categories are deselected.

The voting button 98 is used by the user in the voting area 9
It is a button for transmitting the information input in 7 to the information server 5. After the user completes the entry in voting area 97,
When the voting button 98 is clicked, voting data that specifies the content selected by the user for each evaluation item is displayed on the information server 5.
Sent to.

Like the evaluation area 95, the voting status area 96 corresponds to the evaluation items and is “evaluated”,
Three categories are set: "not evaluated" and "don't know". In this area, the voting status of other users up to the present time is displayed and can be used as a reference when the user inputs in the voting area 97. In this case, the output engine 81 retrieves the voting result from the voting result database 89 described later. By the way, in the present embodiment, the title screen 99 is created in the HTML form, but the present invention is not limited to this, and may be anything such as an applet as long as it can transmit the voting result of the user to the information server 5 side.

Returning to FIG. 17, the detailed screen data 86 is data for displaying the entire text of the article of the product. The title screen 99 is provided with a detail button (not shown), and when the user selects it with a mouse operation or the like, the output engine 81 uses the data such as the effect information database 87 and the article database 20 described later for details. The screen data 86 is generated. Then, this data is transmitted from the information server 5 to the client terminal device 2 used by the user. Then, in the client terminal device 2, the CPU can display the detailed screen on the display by using the browser.

FIG. 19 is a diagram showing an example of a detailed screen displayed on the client terminal device 2. Detail screen 10
In 5, the entire article about the product is posted in the article area 106. Then, in these articles, positive affect terms and negative affect terms appearing in the articles are displayed in a distinguishable manner. In the figure, the positive affect terms are shown with wavy lines, and the negative affect terms are shown with underlines. You can also
This is created by the output engine 81 searching the article database 20 for an article, and searching the effect information database 87 for the effect term appearing in the article and its positive / negative property. The detailed screen 105 includes a previous article selection unit 106 and a next article selection unit 107. When the user selects these with a mouse operation or the like, the previous article or the next article is displayed.

Returning to FIG. 17, the effect information database 87 is a database of information specifying an article, an affect term that appeared in the article, an appearance position of the affect, a score of the affect expression, an appearance frequency of the affect term, and the like. Is. These data are generated by the output engine 81 from the output of the effect analyzer engine 80. Target articles and products (titles) can be added later.

The voting result acquisition unit 88 is a module which receives the voting data transmitted from the client terminal device 2 by the user and updates the voting result database 89. The voting result database 89 is a database that stores the voting contents of users. For example, the voting result database 89 stores information that identifies the users who voted, the products that are the voting targets, and the voting contents for each evaluation item. The voting database 89 is
It is accessible from the output engine 81 and the dictionary adjustment engine 82. The output engine 81 is provided with the current voting result. The dictionary adjustment engine 82 is also provided with information for updating the affect term dictionary 22.

The dictionary adjustment engine 82 is a module that acquires the voting result of each user from the voting result database 89 and updates the attribute (polarity and strength) of the affect term registered in the affect term dictionary 22 to the optimum value. . In this way, the dictionary adjustment engine 82 constitutes a numerical value changing means for changing the numerical value indicating the degree of the subjective strength of the subjective evaluation word. The outline is as follows. Statistical processing is performed on the voting result and the appearance frequency of the effect term contained in the article for the product for which the voting is performed. Then, the degree of correlation between the evaluation item and the affect term is determined, and the polarity (positive / negative directionality) or strength (weighting) of the affect term is automatically updated using this.

As described above, the dictionary adjustment engine 82 indirectly adjusts the evaluation of the product by changing the attribute of the affect term without directly changing the evaluation of the product for which voting has been performed. Therefore, even if the votes are not exhaustive and average for all products, even if there are no votes or products with a small number of votes, adjust the evaluation if there is an article. Is possible.

The necessity of adjusting the affect term dictionary 22 will now be described. Effect analyzer 8
In 0, in the initial state (before adjusting the affect term dictionary 22 by voting etc.), the score is calculated using the polarity and strength of the affect term set by the system designer. For example, the word "old" is generally judged to "represent bad feelings" and may be given a negative polarity in advance. In this case, if the word "old" is included in the article, the score of the article is unconditionally lowered. Conversely, if you set the word "new" to a positive polarity, the score will rise unconditionally if the word "new" is included in the article.

However, it is unclear whether the user thinks the target in accordance with the setting of the polarity. For example, suppose an article about a tourist spot mentions "old temple" and "new temple." Here, if the above-mentioned effect information is set for "old" and "new", a bad score is given to the article "I saw an old temple and felt the historical weight and was moved", A good score is given to the article that says, "The temple is new and it made me feel like a tourist spot." In order to improve such a defect, the system designer can perform detailed analysis for each field and customize the affect term dictionary 20, but in the end, the user's opinion is reflected in the attribute of the affect term. A mechanism is needed. Therefore, in this system, it was decided to adjust the attribute of the effect term by using the voting result of the user.

The dictionary adjustment engine 82 has functions such as (1) mapping function, (2) correlation calculation function, and (3) effect attribute changing function. (1) The mapping function is a voting result map that sums up voting results for each affect term for a certain product (title) and an article about the product (title), and the frequency of appearance of the affect term for each affect term ( An appearance frequency map is generated by totaling the number of appearances).

(2) The correlation calculation function uses these two maps to determine the frequency of occurrence of a certain affect term and the actual number of votes and the voting results (evaluate / not evaluate / not know). Calculate whether there is a correlation. For example, no matter how many times the phrase "package" is "rugged" appears, if the number of votes (for rating) on the package of the product is small, the robustness of the package is not considered to be a criterion for evaluation. On the other hand, the more frequently the word "beautiful" appears (the package is), the more votes (evaluate) the package has, and the less votes it receives (not evaluate), the more beautiful the package is. It is presumed that this is directly linked to the evaluation of.

(3) The effect attribute changing function, based on the calculated correlation, the attribute of the affect term (polarity, strength) set in the affect term dictionary 22.
To update. If we draw the example of the above package, we will reduce the strength of the effect term "strong" and increase the strength of "beautiful". Further, in the affect term dictionary 22, attributes for one affect term can be set for each field to which the evaluation subject belongs and these can be changed independently. In this way, the Affect Term Dictionary 22
, The polarity of the above-mentioned affect term “old” can be positive when the field is a temple and negative when the field is fresh food.

In the case described above, the evaluation axis is not taken into consideration in the statistical processing as in the case of comprehensive evaluation of products, but the same word should be evaluated for each different evaluation axis. You can also That is, taking lipstick as an example, the effect term “beautiful” for the evaluation axis “color gloss” and the affect term “beautiful” for the evaluation axis “package” are distinguished. In this case, first, one affect term is fixed, and a map of voting results of evaluation axes corresponding to products (titles) is created. If the affect terms are not processed separately for each evaluation axis, only the total score will be obtained. In addition, a map is created for the frequency of occurrence of affect terms. That is, the product (title) and the appearance frequency of the effect term in the article about the product are mapped. In this case, when the statistical processing is performed in consideration of the evaluation axis, the same affect term is not counted if the evaluation axis is different. If the evaluation axis is not taken into consideration, the total number may be used.

When the update of the attribute of the affect term is adjusted, the output engine 81 recalculates the affect score and updates the affect information database 87. The update of the affect term dictionary 22 is performed at regular intervals or at regular intervals.

As described above, in the present embodiment,
Instead of directly reflecting the voting result on the score of the title (product), once it is reflected on the polarity and strength of the effect term. Therefore, there is a possibility that the score may be adjusted even if there is a title for which there is no vote. For example, consider the case where titles are ranked by score. It is presumed that users tend to look from the highest ranking. In that case, if the result of the voting is directly reflected in the score, it is possible that even if the rank is low in the initial state, it is not even voted, and it is possible that the rank cannot be permanently higher. However, titles that have not been voted may be ranked higher if the score indirectly increases by adjusting the polarity and strength of the affect term.

By the way, when the number of articles is extremely small compared with the number of votes, it is desirable to directly reflect the voting result on the score. Therefore, when ranking titles, it is desirable to give the user the following options. Ranking display is performed only based on the score calculated by the effect analyzer 80. Display the ranking only based on the voting results. Both are given a certain weight, and a weighted average is taken to display the ranking.

FIG. 20 is a flowchart showing the procedure for updating the affect term dictionary 22. First, the CPU of the information server 5 searches the voting result database 89 and creates a voting result map and an appearance frequency map (step 100). Next, the CPU calculates the correlation between the affect term and the voting result by using the voting result map and the appearance frequency map (step 105). Next, CPU
Changes the polarity / intensity of the affect term in the affect term dictionary 22 based on the calculated correlation (step 110). Next, the CPU recalculates the score and the like of each product using the updated affect term dictionary 22, and updates the affect information database 87.

There are various known methods for calculating the correlation, but any method may be used. The correlation is a method of quantitatively expressing the strength of the correlation with a numerical value from -1 to +1. The closer the correlation is to +1, the stronger the positive correlation is, and the closer it is to -1, the stronger the negative correlation is. Further, when there is no correlation, it becomes 0.

The following effects can be obtained in the present embodiment described above. The user's opinion can be drawn up by voting. Then, the strength / polarity of the affect term can be changed according to the user's opinion. This can improve the accuracy of the score of the affect expression. Therefore, it is possible to provide product evaluation information that more appropriately reflects the user's opinion. Also,
By indirectly linking the voting results to product evaluations through articles, even if there are few votes for a product, the score can be calculated automatically if there is an article, and if either one of the votes and articles exists The score can be presented.

Although one embodiment of the present invention has been described above, the present invention is not limited to the described embodiment, and various modifications can be made within the scope of the claims.

[0165]

According to the present invention, an expression expressing emotionality can be automatically extracted from a sentence, and the emotional strength can be quantified.

[Brief description of drawings]

FIG. 1 is an effect expression processing device 1 according to a first embodiment.
It is a figure showing the logical composition of.

FIG. 2 is a diagram showing an example of a logical configuration of article data stored in an article database.

FIG. 3 is a diagram showing an example of a logical data structure of an affect term dictionary.

FIG. 4 is a diagram showing an example of syntax analysis.

FIG. 5 is a diagram showing one application example of grammar.

FIG. 6 is a block diagram showing a more detailed configuration of an effect expression extraction unit.

FIG. 7 is a diagram showing an example of a result of syntax analysis.

FIG. 8 is a diagram showing an example of a logical configuration of an axis term dictionary.

FIG. 9 is a diagram showing an example of a hardware configuration of an effect expression processing device.

FIG. 10 is a flowchart showing a procedure in which the affective expression processing device extracts an affective expression from an article.

FIG. 11 is a diagram showing an example of a network configuration according to the second embodiment.

FIG. 12 is a diagram showing an example of a logical configuration of an information processing server according to the second embodiment.

FIG. 13 is a diagram showing an example of a list screen.

FIG. 14 is a diagram showing an example of a title screen.

FIG. 15 is a diagram showing an example of a detailed screen.

FIG. 16 is a diagram showing an example of a screen transition diagram.

FIG. 17 is a diagram showing an example of a logical configuration of an information processing server according to a third embodiment.

FIG. 18 is a diagram showing an example of a title screen.

FIG. 19 is a diagram showing an example of a detailed screen.

FIG. 20 is a flowchart showing a procedure for updating an affect term dictionary.

[Explanation of symbols]

1 Affect information extraction device 2 Client terminal device 3 Internet 5 Information processing server 10 Input section 11 Morphological analyzer 12 Affect information addition section 13 Parsing part 14 Affected expression extraction unit 15 Output section 20 article database 21 Morphological analysis dictionary 22 Affect Term Dictionary 23 Parsing dictionary 24-axis term dictionary 51 control unit 52 ROM 53 CPU 54 RAM 55 Input device 56 Output device 57 Communication control device 58 bus lines 59 I / O interface 60 storage medium drive 61 Program storage section 62 data storage 63 storage device 80 Afect Analyzer Engine 81 output engine 82 dictionary adjustment engine 83 List screen data 84 title screen data 86 Detailed screen data 87 Affected Information Database 88 Voting Result Acquisition Department 89 Voting Result Database

─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Natsue Natsue No. 108 Hiraishi Wakamatsu, Kawauchi Town, Tokushima City, Tokushima Prefecture 4 Justsystem Co., Ltd. (56) Reference JP-A-6-110920 (JP, A) (58) Fields surveyed (Int.Cl. ⁷ , DB name) G06F 17/21-17/30

Claims (10)

(57) [Claims] 1. A sentence acquiring unit for acquiring a sentence including an emotional expression word that is a word expressing emotion for a predetermined object, and an emotional expression word for specifying an emotional expression word used in the acquired sentence. In the identification means and the sentence, the emotional strength is affected by acting on the emotional expression word.
Function word specifying means for specifying a function word that is a word to be changed
And a morphological solution for morphologically analyzing the sentence and outputting a morphological sequence
And analysis means, and parsing means for parsing the morpheme strings, by using the analysis result by said syntax analysis means,
Specified as an emotional expression word by the emotional expression word specifying means
Re-specify the word as a function word, or
Re-specified the words specified as functional words in the column as emotional expressions
Re-specifying means for determining, an emotion-expressing word specifying means, or an extracting means for extracting an emotion-expressing part expressing an emotional expression from the sentence by using the emotion-expressing word specified by the re-specifying means. An information processing apparatus comprising: an output unit that outputs the extracted emotion expression portion. Wherein said extraction means, of the sentence, the
The emotion expression word specified by the emotion expression word specifying means or the re-specifying means, and the function word specifying means or the re-specifying
The information processing apparatus according to claim 1, wherein a portion including a function word that acts on the emotion expression word specified by the means is extracted. 3. A numerical value associating unit for associating a numerical value with an emotional expression word included in the extracted emotional expression portion, wherein the numerical value associated by the numerical value associating unit is used,
The information processing apparatus according to claim 1 or 2, wherein the degree of emotional strength of the emotional expression portion is quantified. 4. The information processing apparatus according to claim 3 , further comprising a numerical value changing unit that changes the numerical value associated with the emotion expression word by the numerical value associating unit. 5. The apparatus further comprises operation associating means for associating an operation method with a function word acting on an emotion expression word included in the extracted emotion expression part, wherein the numerical value associated with the emotion expression word is calculated by the operation. The information processing according to claim 3 or 4 , wherein the degree of emotional strength of the emotional expression portion is changed by performing an arithmetical operation by an arithmetical method associated by the associating means. apparatus. 6. The emotion expression word identifying means is the morpheme.
The morphemes forming the morpheme string output by the analysis means are searched in the expression word database storing the emotion expression words, and when the morphemes are searched in the expression word database, the morphemes are specified as emotion expression words. The information processing apparatus according to any one of claims 1 to 5 , which is characterized. 7. The function word identifying means searches for a morpheme that constitutes the morpheme string in a function word database that stores function words, and if the morpheme is found in the function word database, functions the morpheme. The information processing apparatus according to claim 6 , wherein the information processing apparatus is specified as a word. 8. The extracting unit extracts the emotional expression portion by using the dependency structure acquired by the syntax analyzing unit .
The information processing apparatus according to claim 1 . 9. A sentence acquisition means, an emotional expression word identification means, a function word identification means, a morpheme analysis means, and a syntactic analysis means.
A computer including a stage, a re-identification unit, an extraction unit, and an output unit, wherein the sentence acquisition unit acquires a sentence including an emotion expression word that is a word expressing an emotion for a predetermined target. A sentence acquisition step, an emotion expression word identifying step of identifying the emotion expression word used in the obtained sentence by the emotion expression word identifying means, and a strength of emotion acting on the emotion expression word in the sentence To
The function word that is the word to be changed is specified by the function word specifying means.
And a morphological analysis of the sentence by the morphological analysis means.
A morpheme analysis step for outputting a sequence of primes; and a syntax for parsing the morpheme sequence by the syntax analysis means.
By using the analysis step and the analysis result by the syntax analysis means,
The emotion expression is specified by the re-identification means and the emotion expression word specification means.
Reidentify the word identified as a word as a functional word, or
Emotion table for words specified as function words by the function word specifying means
And re-specifying step of re-identification and the current word, the emotional expression word specifying means, or using said emotional expression word identified by the re-identification means, the emotional representation part representing the emotional expression from the sentence the An information processing method comprising: an extraction step of extracting by an extraction means; and an output step of outputting the extracted emotion expression part by the output means. 10. A sentence acquisition function for acquiring a sentence including an emotional expression word that is a word expressing emotion for a predetermined object, and an emotional expression word for specifying an emotional expression word used in the acquired sentence. Specific function and in the above sentence, affect the emotional expression word to increase emotional strength.
Function word specifying function that specifies the function word that is the word to be changed
And a morphological solution for morphologically analyzing the sentence and outputting a morphological sequence
And analysis functions, and parsing function parses the morpheme strings, by using the analysis result by said syntax analysis means,
Specified as an emotional expression word by the emotional expression word specifying means
Re-specify the word as a function word, or by the function word specifying means
Re-identify the word specified as a functional word as an emotional expression word
A specific function, an emotion expression word specifying function, or an extraction function of extracting an emotion expression part expressing an emotional expression from the sentence using the emotion expression word specified by the re-specifying function; An information processing program for realizing the output function that outputs the emotional expression part and a computer.