JP2016103156A - Text feature amount extraction device, text feature amount extraction method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a text feature amount extraction device, a text feature amount extraction method, and a program for corresponding to Japanese and estimating an utterance intention of an inputted text in a text feature amount extraction device for determining a feature amount of a sentence described in a text.SOLUTION: The text feature amount extraction device includes a text analysis part for executing a morphological analysis of an inputted text to extract particles and auxiliary verbs as a functional expression, a score calculation part for determining an intention category to which functional expressions extracted from the text belong on the basis of an intention category dictionary representing a correspondence relation between an intention category obtained by integrating intentions expressed by a sentence and a functional expression belonging to the intention category, and outputting scores obtained by counting the number of functional expressions included in the text in each intention category, and an intention category estimation part for calculating a ratio of respective intention categories represented by the text to estimate an intention represented by the sentence on the basis of the scores, and outputting an estimation result obtained by associating the ratio of intention categories with the text.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a text feature quantity extraction device, a text feature quantity extraction method, and a program.

  Conventionally, specific words (texts) in texts such as TF-IDF (Term Frequency-Inverse Document Frequency) (see Non-Patent Document 1) and LDA (Lent Dirichlet Allocation) (see Non-Patent Document 2). ) As a keyword, there is a technique for determining document features (hereinafter referred to as “feature amount extraction”). In text feature extraction using TF-IDF or LDA, the feature of a target sentence is determined by focusing on the words representing the contents of the sentence and the topic (hereinafter referred to as “content expression”), such as nouns, adjectives, and verbs. Thus, sentences having similar (similar) feature quantities can be grouped (classified) into the same keyword group. The classified results can be used for summarizing the contents of sentences and searching for similar sentences.

  In the feature quantity extraction by TF-IDF, nouns and adjectives appearing in sentences are used as they are. On the other hand, in feature amount extraction by LDA, notation that is semantically similar to notation appearing in a sentence is put together to absorb fluctuations in notation treated as another in feature amount extraction by TF-IDF. Can do.

  For example, as in Patent Document 1, a technique of an apparatus using feature amount extraction by TF-IDF or LDA has been proposed. The technique disclosed in Patent Document 1 is a technique of a document topic extraction apparatus that extracts an appropriate word (text) that explains the concept of a document to be searched.

  In recent years, microblogging using social networking services (SNS) such as Facebook (registered trademark) and Twitter (registered trademark) has become popular. Marketing research is also being conducted in which text posted on this social networking service is used as information for promoting product development / sales and services provided.

  In conventional text classification using TF-IDF or LDA as disclosed in Non-Patent Document 1, Non-Patent Document 2, and Patent Document 1, attention is paid to content expressions in sentences such as nouns, adjectives, and verbs. The feature quantity for classifying the target sentence (text) is extracted. That is, in the conventional text classification, classification is performed from the viewpoint of the content represented by the text (text) and the topic (topic). On the other hand, when using posted text (text) for marketing research, what kind of intention is included in the text (text), that is, the intention of the poster included in the text (hereinafter referred to as “speech intention”). ")"). For this reason, in text classification from the viewpoint of utterance intention, it is necessary to perform classification using information other than content expression, and conventional text classification techniques cannot be applied as they are. For this reason, conventionally, after classifying sentences (text) in terms of contents and topics (topics), a person in charge of marketing research has determined the intention of the classified sentences (text).

  For this reason, for example, Non-Patent Document 3 discloses a technique for estimating the utterance intention of a poster from a short sentence (short text) for Twitter (registered trademark). In the technique disclosed in Non-Patent Document 3, a full-text search technique using an n-gram method such as a unigram (uni-gram), a bigram (bi-gram), or a trigram (tri-gram) is used. The feature amount of the target sentence is determined, and the correspondence between the feature amount of the n-gram and the utterance intention is determined by subjective evaluation, thereby estimating the utterance intention of the posted short sentence.

JP 2012-221316 A

“Tf-idf”, [searched July 31, 2014], Internet, <http: // www. sophia-it. com / content / tf-idf> Blei, David M.M. , Ng, Andrew Y. and Jordan, Michael I .; , “Lent Dirichlet Allocation.”, Journal of Machine Learning Research 3, pp. 993-1022. , 3/1/2003 Renxian Zhang, Wenji Li, Deong Gao, You Ouyang, “Automatic Twitter Topical Summation Proceed Act Act.”, IEEE Transactions. 21, no. 3, pp. 649-658, 2013.3

  By the way, in the technique for estimating the utterance intention for the English version of Twitter (registered trademark) as disclosed in Non-Patent Document 3, from the combination of words (text) whose characteristic amount is determined by n-grams, A dictionary is created to estimate the utterance intention. However, it is difficult to apply a dictionary corresponding to English to Japanese as it is. This is because in Japanese, the number of patterns of word combinations is enormous compared to English, and it is difficult to generate a dictionary that covers various words (text) in Japanese.

  The present invention has been made on the basis of the above problem recognition, and in a text feature amount extraction apparatus for determining a feature amount of a sentence written in text, the utterance intention of input text corresponding to Japanese is set. An object of the present invention is to provide a text feature quantity extraction device, a text feature quantity extraction method, and a program for estimation.

  In order to solve the above problems, a text feature amount extraction device according to the present invention performs a morphological analysis on an input text, and extracts a particle and an auxiliary verb included in the text as a functional expression; The function expression extracted from the text belongs based on an intention category dictionary that represents a correspondence relationship between an intention category in which intentions represented by sentences are collected in advance and a function expression belonging to the intention category. A score calculation unit that determines an intention category and outputs a score obtained by counting the number of functional expressions included in the text for each intention category, and a ratio of each intention category represented by the text based on the score And calculating the intention represented by the sentence based on the calculated proportion of the intention category, And intention category estimation unit for outputting the estimation result associates the serial text, characterized in that it comprises a.

  In addition, the intention category estimation unit in the text feature quantity extraction device of the present invention, when there are a plurality of intention categories having the same score value, the sentence has the same proportion of intentions of the intention categories. It is characterized by estimating that it has.

  Further, the intention category dictionary in the text feature quantity extraction device of the present invention includes a correspondence relationship between the functional expression in which a weighting factor is set in advance and the intention category, and the score calculation unit is set in the functional expression. The number of functional expressions is counted for each intention category by multiplying the weighting factor.

  Further, the intention category estimation unit in the text feature quantity extraction device of the present invention ranks the calculated ratios of the intention categories in descending order, and sets the text belonging to a predetermined rank from the top to the ratio of the intention categories. The associated estimation result is output.

  In the text feature quantity extraction device of the present invention, the functional expression further includes a part of speech of either or both of a conjunction and a formal noun included in the text.

  The text feature amount extraction method of the present invention includes a text analysis step of performing morphological analysis on input text and extracting particles and auxiliary verbs included in the text as functional expressions, and a sentence generated in advance. Determining the intention category to which the functional expression extracted from the text belongs, based on an intention category dictionary representing a correspondence relationship between an intention category in which the intentions are represented and functional expressions belonging to the intention category. A score calculating step for outputting a score obtained by counting the number of the functional expressions included in the text for each intention category, and calculating a ratio of each intention category represented by the text based on the score. The intention represented by the sentence is estimated based on the ratio of the intention category, and the ratio of the intention category and the text are estimated. Characterized in that it comprises a contemplated category estimation step of outputting the estimation result and bets associates, the.

  The program of the present invention is represented by a text analysis step that performs morphological analysis on the input text and extracts particles and auxiliary verbs included in the text as functional expressions, and sentences generated in advance. The intention category to which the functional expression extracted from the text belongs is determined based on an intention category dictionary representing a correspondence relationship between the intention category in which the intention is summarized and the functional expression belonging to the intention category, and is included in the text A score calculation step of outputting a score obtained by counting the number of functional expressions for each intention category, and calculating a ratio of each intention category represented by the text based on the score; Estimate the intention represented by the sentence based on the ratio, and correspond the ratio of the intention category to the text And intent category estimation step of outputting the digit estimation result, and characterized by causing a computer to execute the.

  Advantageous Effects of Invention According to the present invention, the text feature amount extraction device that determines the feature amount of a sentence written in text can obtain an effect of being able to estimate the utterance intention of the input text corresponding to Japanese.

It is the block diagram which showed schematic structure of the functional block which comprises the text feature-value extraction apparatus in embodiment of this invention. It is the flowchart which showed the processing procedure of the intention category estimation in the text feature-value extraction apparatus of this embodiment. It is the figure which showed typically an example of the process of the text feature-value extraction in the text feature-value extraction apparatus of this embodiment. It is the figure which showed an example of the intention category dictionary memorize | stored in the intention category memory | storage part with which the text feature-value extraction apparatus of this embodiment was equipped. It is the figure which showed an example of the score memorize | stored in the score memory | storage part with which the text feature-value extraction apparatus of this embodiment was equipped. It is the figure which showed typically an example of the process of the text classification | category using the text feature-value extraction apparatus of this embodiment.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of functional blocks constituting the text feature quantity extraction device according to the present embodiment. The text feature quantity extraction device 10 is a system configured in, for example, a personal computer (hereinafter referred to as “PC”) and operating on the PC. The text feature quantity extraction device 10 estimates a speech intention (hereinafter, also simply referred to as “intention”) represented by a text by performing natural language processing on a short sentence (text) described in the text, and an estimation result (inputted) Output the ratio of each intention category of the text). Here, the text input to the text feature quantity extraction device 10 is classified by a keyword using, for example, a normal sentence (text) or a conventional text classification technique, that is, the sentence is viewed from the viewpoint of a topic (topic). It is a sentence (text) that is classified and includes the same topic, and any form can be used as long as it is a sentence.

  The text feature quantity extraction device 10 includes a text analysis unit 101, an intention category storage unit 102, a score calculation unit 103, a score storage unit 104, and an intention category estimation unit 105. Note that the text feature quantity extraction device 10 may be configured to function as an add-in software added to application software running on a PC or as a macro program, for example.

  The text analysis unit 101 performs morphological analysis on the input text, and extracts only Japanese functional expressions included in the text. Here, the functional expression is a word or word of a particle or auxiliary verb among words included in the text. In addition, the functional expression may include words (words) of conjunctions and formal nouns among the words included in the text. Then, the text analysis unit 101 sequentially outputs the extracted function expressions to the score calculation unit 103. At this time, the text analysis unit 101 outputs the extracted functional expression for each input text, that is, in association with the input text.

  The functional expression extracted and output by the text analysis unit 101 may be the functional expression word itself extracted from the text, or may be information representing the functional expression wording. Moreover, the information which represents the kind of wording of the functional expression contained in the input text may be sufficient.

  The intention category storage unit 102 includes a category (hereinafter referred to as an “intention category”) in which intentions represented by function expressions described in sentences (text) generated in advance are grouped and function expressions belonging to the intention category. An intention category dictionary representing the correspondence is stored. Here, the intention category included in the intention category dictionary is used to determine the intention of the poster expressed in text such as “description”, “suggestion”, “question”, “comment”, and “other”. It is an item. More specifically, the “description” corresponds to a text that conveys the contributor's own situation and the events in the world. In addition, “suggestion” corresponds to text in which the poster specifically proposes something. The “question” corresponds to a text in which the poster asks a question. The “comment” corresponds to a text that makes a subjective claim such as the emotion of the poster. In addition, “others” corresponds to text belonging to other than the above four intention categories.

  In the text feature quantity extraction device 10, the function expression in the input text is distributed to each intention category based on the intention category dictionary. A detailed description of the contents of the intention category dictionary will be described later.

  The score calculation unit 103 counts the number of function expressions sequentially input from the text analysis unit 101. That is, the score calculation unit 103 counts the number of appearances of functional expressions in each text. At this time, the score calculation unit 103 determines which intention category the input functional expression belongs to based on the intention category dictionary stored in the intention category storage unit 102, and the appearance of the function expression The number of times is counted for each intention category. And the score calculation part 103 memorize | stores in the score memory | storage part 104 by making into a score the appearance frequency of the function expression for every each intention category counted.

  The score storage unit 104 stores a score for each intention category counted by the score calculation unit 103 for each text. The score storage unit 104 stores each text in association with a score for each intention category of the functional expression included in the text.

  The intention category estimation unit 105 outputs the ratio of each intention category represented by the text based on the score stored in the score storage unit 104. Note that a detailed description of the intention category ratio output method in the intention category estimation unit 105 will be described later.

  Then, the intention category estimation unit 105 outputs the ratio (estimation result) of the intention category of each text to the outside of the text feature quantity extraction device 10. The estimation result output by the intention category estimation unit 105 is information in which a text is associated with a ratio of each intention category of the text. At this time, the intention category estimation unit 105 may output the results of all the texts whose intent is estimated as estimation results. For example, the intention category estimation unit 105 may calculate the ratio of the intention category specified by the user of the text feature quantity extraction device 10. The texts may be ranked in descending order (ranking), and the texts belonging to the N-th order from the top, that is, the texts belonging to the predetermined order, may be output together with the ratio of each intention category.

  Next, processing in the text feature quantity extraction device 10 will be described. FIG. 2 is a flowchart showing a processing procedure for intention category estimation in the text feature quantity extraction apparatus 10 of the present embodiment.

  In the processing in the text feature quantity extraction device 10, first, in step S10, the text analysis unit 101 performs morphological analysis on the input text, and decomposes the text into parts of speech. Here, not only the function expression for estimating the intention of the utterance, but also the words and phrases representing the content of the sentence and the topic, such as nouns and verbs, are decomposed as respective parts of speech.

  Subsequently, in step S20, the text analysis unit 101 extracts only the words (words) of the part of speech (particle, auxiliary verb, conjunction, and formal noun) of the functional expression included in the text decomposed into each part of speech. That is, the text analysis unit 101 does not extract words (words) of a part of speech of content expression representing the content of a sentence or a topic such as a noun or a verb. Then, the text analysis unit 101 sequentially outputs the extracted words (words) of the functional expression to the score calculation unit 103 for each text.

  Subsequently, in step S <b> 30, the score calculation unit 103 calculates the number of appearances of functional expression words (words) sequentially input from the text analysis unit 101 based on the intention category dictionary stored in the intention category storage unit 102. Count for each text. Then, the score calculation unit 103 outputs and stores the counted number of appearances of the functional expression word (word) in the score storage unit 104.

  Subsequently, in step S40, the intention category estimation unit 105 calculates the ratio of each intention category of each text based on the score stored in the score storage unit 104, and the result calculated here (estimation result). Are output to the outside for each text.

  With such a configuration and processing procedure, the text feature quantity extraction device 10 outputs the ratio of each intention category of the input text.

  Here, an example of each process in the text feature amount extraction apparatus 10 will be described. FIG. 3 is a diagram schematically illustrating an example of text feature amount extraction processing in the text feature amount extraction apparatus 10 of the present embodiment. In the example shown in FIG. 3, when two texts of text a and text b are input to the text feature quantity extraction device 10, the intention categories of the input text are estimated, and the ratio of the intention categories is The case where each text is classified in a high category is schematically shown.

  The text feature quantity extraction device 10 estimates the ratio of the intention category of each text. In the text feature amount extraction processing using the intention category, the text analysis unit 101 decomposes the input text into parts of speech (step S10), and further extracts a functional expression word (word) to obtain a score calculation unit. It outputs to 103 (step S20). Then, the score calculation unit 103 counts the number of appearances of the functional expression word (word) for each text based on the intention category dictionary stored in the intention category storage unit 102 (step S30).

  Here, an example of the intention category dictionary stored in the intention category storage unit 102 will be described. FIG. 4 is a diagram illustrating an example of an intention category dictionary stored in the intention category storage unit 102 provided in the text feature amount extraction apparatus 10 of the present embodiment. As described above, the intention category dictionary is a dictionary generated in advance that represents the correspondence between the functional expression and the intention category. Functional expressions include particles such as “ha”, “ga”, “ka”, “tai”, auxiliary verbs, conjunctions, and formal noun words (words). Each functional expression is classified into an intention category of “description”, “suggestion”, “question”, “comment”, and “other”. FIG. 4 shows an example of an intention category dictionary in which “ha” is classified as “declaration”, “ga” is classified as “declaration”, “ka” is classified as “question”, and “tai” is classified as “suggestion”. Is shown.

  Some functional expressions belong to a plurality of intention categories. In this case, an intention category dictionary in which the same function expression is associated with a plurality of intention categories may be generated. Also, based on the frequency of appearance of functional expressions in generally considered sentences, a weighting factor for associating the functional expression with each intention category may be set in advance, and an intention category dictionary including the weighting coefficient may be generated. Good. In this case, when calculating the number of appearances of the functional expression for each intention category, the score calculation unit 103 multiplies the weighting coefficient set for the functional expression.

  Note that the intention category dictionary is not only generated in advance and stored in the intention category storage unit 102. For example, the text analysis unit 101 learns based on the input text, and the intention category according to the learned content. The category dictionary may be updated.

  The score calculation unit 103 determines to which intention category the functional expressions sequentially input from the text analysis unit 101 belong, by referring to the intention category dictionary as shown in FIG. The number of appearances is counted for each intention category. Then, the score calculation unit 103 stores the counted number of appearances for each intention category in the score storage unit 104 as a score (step S30).

  Here, an example of the score stored in the score storage unit 104 will be described. FIG. 5 is a diagram illustrating an example of a score stored in the score storage unit 104 provided in the text feature amount extraction apparatus 10 of the present embodiment. The score storage unit 104 stores each input text and the score for each intention category of the functional expression included in the text in association with each other. FIG. 5 shows a case where the score of the functional expression included in the text x, the score of the functional expression included in the text y, and the score of the functional expression included in the text z are stored. More specifically, in FIG. 5, in the functional expression included in the text x, the number of appearances of “description” = 3, the number of appearances of “suggestion” = 4, the number of appearances of “question” = 2, and the “comment”. The number of times of appearance = 1 and the score of the number of appearances of “others” = 0 are stored. In addition, in the functional expression included in the text y, FIG. 5 shows that the number of appearances of “description” = 2, the number of appearances of “suggestion” = 2, the number of appearances of “question” = 4, and the number of appearances of “comment” = 2 shows a case where a score of “others” appearance count = 1 is stored. In addition, in the functional expression included in the text z, FIG. 5 shows that the number of appearances of “description” = 1, the number of appearances of “suggestion” = 1, the number of appearances of “question” = 1, and the number of appearances of “comment” = In this example, a score of 0, “others” appearance count = 0 is stored.

  The intention category estimation unit 105 determines each intention category from the viewpoint of the utterance intention of the person who posted the text, such as what ratio each text has, based on the score as shown in FIG. Is output (step S40). At this time, the ratio of the intention category increases in descending order of the number of appearances of the functional expressions included in the text. When the number of appearances of functional expressions included in the text is the same, the ratio of the intention category is the same. For example, in the case of the score shown in FIG. 5, the intention category estimation unit 105 determines that the ratio of the intention category of the text x is larger in the order of “suggestion”> “description”> “question”> “comment”> “others”. It is estimated that the proportion of the intention category of the text y is larger in the order of “question”> “description, proposal, comment (the same proportion is for the three categories)”> “other”. Further, in the case of the score shown in FIG. 5, it is estimated that the ratio of the intention category of the text z is larger in the order of “description, proposal, question (3 categories are in the same order)”> “comment, other”.

  In addition, when the score calculation unit 103 refers to the intention category dictionary including the weighting coefficient and counts the number of appearances of the functional expression, a plurality of scores such as the scores of the text y and the text z illustrated in FIG. Even if the number of appearances of functional expressions in the intention category is the same, the intention category estimation unit 105 can estimate the intention category more suited to the utterance intention. That is, for example, when the weighting coefficient of the functional expression included in any one intention category of “description”, “suggestion”, or “question” that has the same number of appearances of the functional expression in the text z is high (for example, When the statement weight coefficient is 2 and the proposal weight coefficient is 0.5), the ratio of the intention category of the text z is “statement”> “question”> “proposal” in which the functional expression with a high weight coefficient is counted ">" Comment, other "is estimated in order.

  In the example of the score shown in FIG. 5, a score obtained by summing up the number of appearances of all functional expressions is stored. For example, the score calculation unit 103 calculates the total number of appearances for each text in which the number of appearances is counted, and stores the total number of appearances of all functional expressions in the score storage unit 104. FIG. 5 stores the number of appearances of all function expressions included in text x = 10, the number of appearances of all function expressions included in text y = 11, and the number of appearances of all function expressions included in text z = 3. Shows the case.

  In this way, when the score of the number of appearances of all functional expressions is stored in the score storage unit 104, the intention category estimation unit 105 determines whether each intention category is classified when classifying each text from the viewpoint of utterance intention. The ratio of the intention category can be estimated based on a value obtained by normalizing the number of appearances with the number of appearances of all functional expressions. According to this method, for example, in the case where only text based on the number of appearances of functional expressions is selected from text having a large ratio of a specific intention category, the influence of the length of the input text is absorbed. be able to.

  And the intention category estimation part 105 outputs the ratio of each intention category of each text as an estimation result for every text. At this time, as described above, the intention category estimation unit 105 may output the ratio of the intention categories of all the texts as an estimation result. However, the intention category specified by the user of the text feature quantity extraction device 10 may be output. You may output the text which belongs to the top Nth with a high ratio as an estimation result. In addition, the intention category estimation unit 105 may output, as an estimation result, text having a score value equal to or higher than a threshold specified by the user of the text feature quantity extraction device 10 together with information on the intention category. The threshold value may be specified by the number of appearances, or may be specified by a normalized ratio. For example, when the threshold value designated by the user is “number of appearances is 3,” the intention category estimation unit 105 sets the number of occurrences of “description” = 3 and the number of occurrences of “suggestion” = 4. , The text y with the number of occurrences of “question” = 4 is output together with the ratio of each intention category.

  The text input to the text feature extraction device 10 is text classified by a conventional text classification method, that is, text obtained by classifying sentences from the viewpoint of a topic (topic) based on keywords included in the text. Also good.

  When texts classified in terms of topics (topics) are input to the text feature quantity extraction device 10, the text feature quantity extraction device 10 selects all text topics (topics), that is, all classified by keywords. The ratio of the intention category of the text may be output. However, for example, the estimation result of the ratio of the intended category may be output only for the text of the topic (topic) designated by the user of the text feature quantity extraction device 10, that is, the text including the keyword designated by the user. At this time, in step S20, the text analysis unit 101 extracts functional expressions only from the text of the topic (topic) specified by the user, and outputs the extracted functional expressions to the score calculation unit 103 sequentially for each text. The processing can reduce the processing load of the text feature quantity extraction apparatus 10.

  Here, an example of each process in the text feature amount extraction apparatus 10 will be described. FIG. 6 is a diagram schematically illustrating an example of text classification processing using the text feature amount extraction apparatus 10 of the present embodiment. FIG. 6A shows an example of processing for classifying text from the viewpoint of a topic, and FIG. 6B shows a case in which classification is performed by using text estimation results with a high proportion of intended categories. An example is shown.

  In the example of the text classification process shown in FIG. 6A, when the text obtained by classifying sentences from the viewpoint of a topic based on a keyword is used as the input of the text analysis unit 101, the estimation result of the intended category is obtained. The case where it classify | categorizes using is shown typically. More specifically, in the example illustrated in FIG. 6A, eight texts from text a to text h including any one of topics A to C are included in the text feature quantity extraction device 10. After the input text is classified from the viewpoint of topic A to topic C, the intention category of the text of topic A is estimated, and each text is classified in a category with a high ratio of the intention category. It shows.

  The text feature quantity extraction device 10 first classifies texts in terms of topics. In the example shown in FIG. 6A, text a, text b, text d, and text g are classified as topic A text, text c and text h are classified as topic B text, text e and text The case where f is classified as the text of topic C is shown. As described above, in the text classification from the viewpoint of a topic (topic), the text including the same topic is classified so as to be collected.

  Subsequently, the text feature quantity extraction device 10 estimates the ratio of the intention category of each text of the topic A, and classifies the ratio of the intention category of each text by category. In the text classification process using the intention category, the text analysis unit 101 decomposes the input text into parts of speech (step S10), further extracts a functional expression and outputs it to the score calculation unit 103 (step S20). ). Then, the score calculation unit 103 counts the number of appearances of functional expressions for each text based on the intention category dictionary stored in the intention category storage unit 102 (step S30).

  Also, in FIG. 6B, the topic A is the topic of “melon bread”, the ratio of the “proposal” intent category is high, “I want to buy a lot of that melon bread”, and “the melon bread becomes sweeter” An example is shown in the case of classification using the estimation result with the text “I'm fine”. Also, in FIG. 6B, the ratio of the intention category of “comment” is high, the text “I think this melon bread is very delicious”, and the text “Happy to eat melon bread and coffee together” An example in the case of classification using the estimation result is shown.

  Here, an example of the text classified into each intention category is shown. The text classified into the intention category of “statement” includes, for example, the text “I found that the store selling the melon bread is close to the house”. The text classified into the “suggestion” intention category includes, for example, a text “I want you to improve the flavor of melon bread”. The text classified into the “question” intention category includes, for example, the text “Why is the topic melon bread sold out at any store?”. The text classified into the intention category of “comment” includes, for example, the text “I feel happy because melon bread is too delicious”. The text classified into the “other” intention category includes, for example, the text “Both melon bread and bread are round”.

  As described above, according to the mode for carrying out the present invention, the text feature quantity extraction device performs a morphological analysis on the input text, and expresses a Japanese functional expression (particle) included in the text. , Auxiliary verbs, conjunctions, and formal nouns). And according to the form for implementing this invention, a text feature-value extraction apparatus is contained in the input text based on the intention category dictionary which matched the function expression and intention category which were memorize | stored beforehand. Count the number of appearances of functional expressions. And according to the form for implementing this invention, the ratio of each intention category in the input text is estimated. Thereby, in the text feature-value extraction apparatus of the form for implementing this invention, what kind of intention is contained in the text and what ratio can be estimated. In other words, the text feature quantity extraction device in the form for carrying out the present invention can automatically determine the intention of the text at the time of posting.

  As a result, a user who uses the text feature amount extraction apparatus of the embodiment for carrying out the present invention only classifies each text content or topic (topic) unit when analyzing a large amount of text. Instead, information (intentions) in a plurality of texts classified into specific contents or topics (topics) can be subdivided and analyzed. That is, the user selects a text of a desired intention category from among the texts separated by the text feature quantity extraction device, for example, a service for developing / selling or providing a product in marketing research or the like. Can be used as information to promote work, and work efficiency can be improved.

  Moreover, in the text feature quantity extraction device of the form for carrying out the present invention, the utterance intention is estimated from the functional expression of the text input based on the intention category dictionary generated in advance. Thereby, in the text feature quantity extraction device of the form for carrying out the present invention, the utterance intention is estimated from a combination of words (text) whose feature quantity is determined by n-gram as in the conventional case. It is possible to make the estimation of utterance intention correspond to Japanese without increasing the number of combination patterns.

  In the form for carrying out the present invention, the intention category storage unit and the score storage unit provided in the text feature quantity extraction device are replaced with other components (text analysis unit, score calculation unit, and intention category estimation unit 105). The case where it prepares in parallel with was demonstrated. However, the configuration including the intention category storage unit and the score storage unit is not limited to the configuration shown in the embodiment for carrying out the present invention, and is included in other components included in the text feature quantity extraction device. It may be a component. For example, the intention category storage unit and the score storage unit may be provided in the score calculation unit. In addition, for example, the intention category storage unit may be provided in the score calculation unit, and the score storage unit may be provided in the intention category estimation unit.

  In the embodiment for carrying out the present invention, the case has been described in which the input to the text feature quantity extraction device is a short sentence written in text. However, the input to the text feature quantity extraction device is not limited to the configuration shown in the embodiment for carrying out the present invention. For example, the structure may be such that voice data is input to the text feature quantity extraction device. In this case, the text analysis unit 101 converts the utterance included in the input voice data into text, and then executes the process for the text described above.

  In the embodiment for carrying out the present invention, the case has been described in which the text feature quantity extraction device is used for estimating the utterance intention of Japanese text. However, the range in which the text feature quantity extraction device of the present invention can be used, that is, the language that can estimate the utterance intention of the text using the text feature quantity extraction device of the present invention is limited to only Japanese. Is not to be done. For example, in the case of a language (for example, Chinese, Thai, etc.) in which the text can be decomposed into parts of speech by the morphological analysis executed by the text analysis unit 101, an intention category dictionary corresponding to the language is generated in advance. Thus, the present invention can be easily applied to estimation of the utterance intention of a short sentence (text) written in the text of the language.

  In the embodiment for carrying out the present invention, the case has been described in which text input by morphological analysis is decomposed by part of speech. However, the method of disassembling the text by part of speech is not limited to the morphological analysis shown in the embodiment for carrying out the present invention, and any other natural language processing that can decompose the text by part of speech. Similarly, when the text is decomposed by part of speech by the method, it is possible to estimate the utterance intention of the input text.

  Further, in the embodiment for carrying out the present invention, the case where the text feature amount extraction apparatus 10 is configured in a PC and operates as a function of add-in software added to application software operating on the PC or a macro program will be described. did. However, the configuration of the text feature quantity extraction device 10 is not limited to the mode for carrying out the present invention. For example, the text feature quantity extraction device 10 can be configured as a system that operates on a server computer shared by a plurality of users. Further, for example, the software is provided as software in which all the components of the text feature quantity extraction device 10 are incorporated, and the functions and operations of the text feature quantity extraction device 10 described above are performed by inputting and executing text in this software. It can also be configured to be realized.

  A program for realizing processing by each component in the text feature quantity extraction apparatus 10 shown in FIG. 1 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is stored in a computer system. The above-described various processes related to the text feature quantity extraction apparatus 10 may be performed by causing the text feature quantity extraction apparatus 10 to read and execute the process. Here, the “computer system” may include an OS and hardware such as peripheral devices. Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. The “computer-readable recording medium” means a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a CD-ROM, a hard disk built in a computer system, etc. This is a storage device.

  Further, the “computer-readable recording medium” refers to a volatile memory (for example, DRAM (Dynamic) in a computer system serving as a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. Random Access Memory)) that holds a program for a certain period of time is also included. The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what implement | achieves the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

  The embodiment of the present invention has been described above with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes various modifications within the scope of the present invention. It is.

DESCRIPTION OF SYMBOLS 10 ... Text feature-value extraction apparatus 101 ... Text analysis part 102 ... Intention category storage part 103 ... Score calculation part 104 ... Score storage part 105 ... Intention category estimation part

In order to solve the above problems, a text feature amount extraction device according to the present invention performs a morphological analysis on an input text, and extracts a particle and an auxiliary verb included in the text as a functional expression; The function expression extracted from the text belongs based on an intention category dictionary that represents a correspondence relationship between an intention category in which intentions represented by sentences are collected in advance and a function expression belonging to the intention category. A score calculation unit that determines an intention category and outputs a score obtained by counting the number of the functional expressions included in the text for each intention category; and at least one intention category represented by the text based on the score calculating a percentage of each intent represented by the sentence estimated based on the ratio of the intended category the calculated, the intended catheter It characterized in that it comprises the intended category estimation unit for outputting the estimation result that associates ratio of Li and said text, a.

The text feature amount extraction method of the present invention includes a text analysis step of performing morphological analysis on input text and extracting particles and auxiliary verbs included in the text as functional expressions, and a sentence generated in advance. Determining the intention category to which the functional expression extracted from the text belongs, based on an intention category dictionary representing a correspondence relationship between an intention category in which the intentions are represented and functional expressions belonging to the intention category. A score calculation step of outputting a score obtained by counting the number of functional expressions included in the text for each intention category, and calculating a ratio of each of at least one intention category represented by the text based on the score. Estimating the intention represented by the sentence based on the calculated proportion of the intention category, and assigning the intention category Characterized in that it comprises a and intent category estimation step of outputting the estimation result that associates with the text and.

The program of the present invention is represented by a text analysis step that performs morphological analysis on the input text and extracts particles and auxiliary verbs included in the text as functional expressions, and sentences generated in advance. The intention category to which the functional expression extracted from the text belongs is determined based on an intention category dictionary representing a correspondence relationship between the intention category in which the intention is summarized and the functional expression belonging to the intention category, and is included in the text A score calculation step for outputting a score obtained by counting the number of functional expressions for each intention category, and calculating a ratio of each of at least one intention category represented by the text based on the score; The intention represented by the sentence is estimated based on the ratio of the intention category, and the ratio of the intention category and the text are estimated. And intent category estimation step of outputting the estimation result and bets associates, that causes a computer to execute the features.

In order to solve the above problems, a text feature amount extraction device according to the present invention performs a morphological analysis on an input text, and extracts a particle and an auxiliary verb included in the text as a functional expression; The function expression extracted from the text belongs based on an intention category dictionary that represents a correspondence relationship between an intention category in which intentions represented by sentences are collected in advance and a function expression belonging to the intention category. A score calculation unit that determines an intention category and outputs a score obtained by counting the number of functional expressions included in the text for each intention category; and the score is normalized by the number of all the functional expressions included in the text based on the phased value, to calculate the percentage of each of the at least one intended categories the text represented, the intention that the calculated The intent represented by the sentence estimated based on the percentage of the categories, characterized in that it comprises a and intent category estimation unit for outputting the estimation result that associates said the ratio of the intended category text.

Furthermore, the intention category estimation unit in the text feature quantity extraction unit of the present invention, the calculated ratio of said intended category is ranks the plurality of the text in descending order, the said text belonging to up to a predetermined order from the upper The estimation result associated with the ratio of the intention category is output.

The text feature amount extraction method of the present invention includes a text analysis step of performing morphological analysis on input text and extracting particles and auxiliary verbs included in the text as functional expressions, and a sentence generated in advance. Determining the intention category to which the functional expression extracted from the text belongs, based on an intention category dictionary representing a correspondence relationship between an intention category in which the intentions are represented and functional expressions belonging to the intention category. A score calculation step of outputting a score obtained by counting the number of functional expressions included in the text for each intention category , and a value obtained by normalizing the scores by the number of all the functional expressions included in the text , Calculating a ratio of each of at least one intention category represented by the text, and calculating the intention category The intent represented by the sentence estimated based on the ratio, characterized in that it comprises a and intent category estimation step of outputting the estimation result that associates said the ratio of the intended category text.

The program of the present invention is represented by a text analysis step that performs morphological analysis on the input text and extracts particles and auxiliary verbs included in the text as functional expressions, and sentences generated in advance. The intention category to which the functional expression extracted from the text belongs is determined based on an intention category dictionary representing a correspondence relationship between the intention category in which the intention is summarized and the functional expression belonging to the intention category, and is included in the text A score calculation step of outputting a score obtained by counting the number of functional expressions for each intention category, and the text based on a value obtained by normalizing the scores by the number of all the functional expressions included in the text. Calculating a ratio of each of the at least one intention category to be expressed, and based on the calculated ratio of the intention category Wherein estimating the intention sentence represents, characterized in that to execute the said text and the ratio of the intended category and intended category estimation step of outputting the estimation result which associates, to the computer Te.

Claims (7)

A text analysis unit that performs morphological analysis on the input text and extracts particles and auxiliary verbs included in the text as functional expressions;
The intention to which the functional expression extracted from the text belongs based on an intention category dictionary that represents a correspondence relationship between an intention category in which intentions expressed in sentences are summarized and functional expressions belonging to the intention category. A score calculation unit that determines a category and outputs a score obtained by counting the number of functional expressions included in the text for each intention category;
A ratio of each intention category represented by the text is calculated based on the score, an intention represented by the sentence is estimated based on the calculated ratio of the intention category, and the ratio of the intention category and the text are calculated. An intention category estimation unit for outputting the associated estimation results;
A text feature quantity extraction apparatus comprising: The intention category estimation unit includes:
When there are a plurality of the intention categories having the same score value, it is estimated that the sentence has the same ratio of the intentions of the intention categories.
The text feature amount extraction apparatus according to claim 1, wherein: The intention category dictionary is
Including a correspondence relationship between the functional expression in which a weighting factor is set in advance and the intention category;
The score calculation unit
Multiplying the functional expression by a weighting factor set, and counting the number of functional expressions for each intention category;
The text feature amount extraction apparatus according to claim 2, wherein The intention category estimation unit includes:
Rank the calculated proportions of the intention categories in descending order, and output the estimation results in which the text belonging to the predetermined rank from the top is associated with the proportion of the intention categories;
The text feature quantity extraction apparatus according to claim 1, wherein the text feature quantity extraction apparatus is a text feature quantity extraction device. The functional expression is
It further includes the part of speech of either or both of a conjunction or a formal noun included in the text,
The text feature quantity extraction device according to claim 1, wherein the text feature quantity extraction device is a text feature quantity extraction device. A text analysis step of performing morphological analysis on the input text and extracting particles and auxiliary verbs included in the text as functional expressions;
The intention to which the functional expression extracted from the text belongs based on an intention category dictionary that represents a correspondence relationship between an intention category in which intentions expressed in sentences are summarized and functional expressions belonging to the intention category. A score calculation step of determining a category and outputting a score obtained by counting the number of functional expressions included in the text for each intention category;
A ratio of each intention category represented by the text is calculated based on the score, an intention represented by the sentence is estimated based on the calculated ratio of the intention category, and the ratio of the intention category and the text are calculated. An intention category estimation step for outputting a corresponding estimation result;
A text feature quantity extraction method characterized by including: A text analysis step of performing morphological analysis on the input text and extracting particles and auxiliary verbs included in the text as functional expressions;
The intention to which the functional expression extracted from the text belongs based on an intention category dictionary that represents a correspondence relationship between an intention category in which intentions expressed in sentences are summarized and functional expressions belonging to the intention category. A score calculation step of determining a category and outputting a score obtained by counting the number of functional expressions included in the text for each intention category;
A ratio of each intention category represented by the text is calculated based on the score, an intention represented by the sentence is estimated based on the calculated ratio of the intention category, and the ratio of the intention category and the text are calculated. An intention category estimation step for outputting a corresponding estimation result;
A program that causes a computer to execute.

Images