JP2010113511A - Device for extracting and retrieving sensitivity information, method therefor and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To achieve sensitivity retrieval allowing acquisition of a retrieval result by only inputting a target, and not needing to impart metadata or the like related to the target via manpower. <P>SOLUTION: In a sensitivity information extraction part 10, a text document is language-analyzed, sensitivity expression inside the text document is extracted by use of text analysis information thereof, a sensitivity expression dictionary 20, and a sensitivity expression extraction rule 30, and sensitivity information including information related to at least the sensitivity expression in an extracted sensitivity expression unit is generated and is accumulated in a sensitivity information DB 40. In a sensitivity retrieval part 50, a sensitivity vector dictionary 60 is searched together with the sensitivity information DB 40 based on an input retrieval condition, a retrieval result thereof is totaled based on a tabulation condition to create a tabulation results, and the retrieval result of a predetermined format is output by use of the tabulation results. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a technique for extracting Kansei information from a text document into a database and performing information retrieval using the database.

  Most conventional information search apparatuses output a document including a keyword as a search result when the keyword is input. For this reason, for ambiguous and sensory requests such as wanting to know other objects that are similar in image to one object, a search is performed using an expression representing the image of the one object as a keyword. It has been necessary to take time and effort to check each document of the search result including a large amount of things that are far from the intention and to search for other objects that are actually similar in image to the one object. Further, when searching, it is difficult to enumerate all expressions representing the image of the one object, and it is difficult to find another object that is exactly similar in image to the one object.

  For example, if an actor recognizes that the image is "quiet and cool", search for "quiet and cool", check each document in the search results, I needed to find a person.

In order to meet such needs, a technology called “Kansei Search” has been put into practical use in which a product can be found from a product image or ambiguous language (see Non-Patent Document 1). However, in this technology, for example, in order to be able to search for another product related to an image of a product, metadata (accompanying information) representing various characteristics of the product on the coordinate axis of the image of each product There is a problem that it takes time and effort to input manually.
"Kansei search from marketing perspective-Chairman Alberto Yamakawa-People who challenge the recommended engine (3)", [online], May 29, 2008, (C) 2008 Nikkei Inc. / Nikkei Digital Media, Inc., [Heisei Search on October 21, 2008], Internet <URL: http://it.nikkei.co.jp/internet/news/index.aspx?n=MMITba000029052008>

  If you want to know other objects that are similar in image to the one, you can improve the convenience of the user if you can obtain the search result just by inputting the one, as in the case of normal information search. To do.

  Also, in order to enable such information retrieval, that is, sensuous information retrieval, it is laborious to provide metadata about each target manually, so that it is possible to save labor as much as possible. desirable.

  The present invention has been made in view of the above points, and extracts Kansei information from a text document to create a database. By using this database, an information retrieval is performed by obtaining the Kansei distance between a search object and its comparison object. By doing so, it is possible to obtain a search result only by inputting a target, and to add metadata related to the target manually, and to save labor and to reduce operational costs.

  The present invention is a sensitivity information extraction / retrieval device that extracts sensitivity information from a text document and creates a database, and performs information search using the database, and performs text analysis on the text document to obtain text analysis information. Registered the text analysis section to output, the text analysis information, the Kansei expression dictionary that registered the information of words that could be Kansei expressions that are expressions representing impressions and feelings about arbitrary objects, and the patterns of Kansei expressions to be extracted A sensitivity expression extraction unit that extracts a sensitivity expression in the text document using a sensitivity expression extraction rule; and a sensitivity information generation unit that generates and outputs sensitivity information including at least information related to the sensitivity expression in a sensitivity expression unit. The sentiment information extraction unit configured, the sentiment information database storing the sentiment information output from the sentiment information extraction unit, and the input A condition input unit that receives the search condition and the total condition, and at least a sensitivity vector dictionary in which the sensitivity vector for each sensitivity expression is registered together with the sensitivity information database based on the search condition, and the search result is based on the total condition And a sensibility search unit that includes a search / aggregation unit that aggregates and creates a tabulation result and a result output unit that outputs the search result in a predetermined format using the tabulation result. And

  According to the present invention, Kansei information is extracted from an arbitrary text document and converted into a database, and by searching for information on the Kansei distance between the search target and the comparison target using this database, It is possible to perform a sensuous search in which a search result can be obtained simply by inputting a target, and it is not necessary to manually add metadata or the like related to the target, so that labor can be saved and operation costs can be reduced accordingly.

  Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments.

  The sensitivity information extraction / retrieval device of the present invention comprises a computer device, and includes input means such as a keyboard, display means (output means) such as a monitor, storage means such as a hard disk and a memory, and a communication device that can be connected to an external network. (Both are not shown).

<First Embodiment>
FIG. 1 shows a first embodiment of a sensitivity information extraction / retrieval apparatus according to the present invention. The sensitivity information extraction / retrieval apparatus according to the present invention includes a sensitivity information extraction unit 10, a sensitivity expression dictionary 20, and a sensitivity expression extraction rule. 30, Kansei information database (DB) 40, Kansei search unit 50, Kansei vector dictionary 60, and tabulation result accumulation database (DB) 70. 2 is a flowchart of sensitivity information extraction in the sensitivity information extraction unit 10, FIG. 3 is a flowchart of sensitivity search processing in the sensitivity search unit 50, and FIG. 4 is an explanatory diagram showing an example of processing in the sensitivity information extraction unit 10. is there.

  As shown in FIG. 1, the sensibility information extraction unit 10 includes a text analysis unit 11, a sensibility expression extraction unit 12, and a sensibility information generation unit 13, and is directly input from an input unit (not shown) or read and input from a storage unit. Or a set of documents (text documents) input from another device or the like via a communication medium is input, and the sensitivity information is output to the sensitivity information DB 50.

  Here, the text document has at least a document ID for identifying the document and a text (text data). In addition, document meta information such as creation date / time, document type (eg, blog), author ID, and the like may be provided. FIG. 4A shows an example of a text document provided with document type, creation date and author ID as document meta information.

  The text analysis unit 11 performs at least a well-known morphological analysis process (generation of word information) on the text of the text document, and at least word information (consisting of word notation, part of speech, reading, etc.). The included text analysis information is output (step S1 in FIG. 2).

  As an example of text analysis information including at least word information, text analysis information (notation, reading, etc. is omitted) for the text document in FIG. 4A is shown in FIG. 4B (however, it will be described later in FIG. 4B). Although the underline indicating the result of the emotional expression extraction process, that is, the extracted emotional expression is included, it does not exist at this stage.

  The emotional expression extraction unit 12 extracts the emotional expression in the text document using the emotional expression dictionary 20 and the emotional expression extraction rule 30 together with the text analysis information output from the text analysis unit 11 (step S2 in FIG. 2). ).

  Here, the emotional expression is an expression such as “quiet”, “cool”, “solid”, “cool”, etc., representing an impression or feeling about an arbitrary object.

  The sensibility expression dictionary 20 is configured by registering category information together with the notation, part of speech, and reading of a word that can be a sensibility expression. FIG. 5 shows an example thereof.

  In addition, the Kansei Expression Extraction Rule 30 registers a Rule ID, a Kansei Expression Pattern composed of a regular expression of each word constituting the Kansei Expression, and a category of the Kansei Expression for the rule related to the description of the Kansei Expression. Is.

  FIG. 6 shows an example of the sensitivity expression extraction rule 30. In the Kansei expression pattern, <> is one (or word) regular expression, (?: <>) * Is zero or more (or word) regular expressions, (?: <>)? Is equivalent to 0 or 1 (word) regular expression, “e:” is a condition for emotional expression, “p:” is a condition for part of speech, and “h:” is a condition for notation. Further, the category “dic-category” represents that the category assigned in the emotional expression dictionary 20 is used as the category of the emotional expression as it is.

  Kansei expressions can be extracted by using the processing of an existing technique, for example, the evaluation expression extraction unit of Japanese Patent Application Laid-Open No. 2008-140359 “Evaluation Information Extraction Apparatus, Evaluation Information Extraction Method and Program”. In this case, the sensitivity expression dictionary 20 is used instead of the evaluation expression dictionary, and the sensitivity expression extraction rule 30 is used instead of the evaluation expression rule.

  Further, the sensitivity expression extracted by the sensitivity expression extraction rule 30 may be narrowed down by the information given during the extraction. For example, when the sensitivity expression extraction rule of FIG. 6 is used, only the expression whose rule ID starts with “Kansei” is extracted as the sensitivity expression.

  The processing result (extracted emotional expression) of the emotional expression extracting unit 12 for the text document of FIG. 4A is indicated by an underline in FIG. 4B.

  The sentiment information generating unit 13 generates and outputs the sentiment information including at least sentiment expression information that is information related to the sentiment expression in the sentiment expression unit extracted by the sentiment expression extracting unit 12 (step S3 in FIG. 2).

  The Kansei information is composed of Kansei expression unit records extracted by the Kansei expression extracting unit 12, and each record includes at least a record ID and Kansei expression information including notation, standard form, and category related to each Kansei expression. Furthermore, document meta information (document information) including a document ID, an author ID, and an author category related to a text document from which each sensitivity expression is extracted may be included. An example of sensitivity information for the text document in FIG. 4A is shown in FIG.

  The sensitivity information DB 40 is a database that stores sensitivity information output from the sensitivity information extraction unit 10, and a known information that can be searched for records by various search conditions such as SQL may be used.

  As shown in FIG. 1, the sensitivity search unit 50 includes a condition input unit 51, a search / aggregation unit 52, and a result output unit 53, and is directly input from an input unit (not shown) or read and input from a storage unit. The search condition and the total condition input from another device or the like via the communication medium are input, and the search result is output using the sensitivity information DB 40 and the sensitivity vector dictionary 60 or the total result storage DB 70 in addition to these.

  The sentiment vector dictionary 60 registers the sentiment vector for each sentiment expression together with the sentiment expression ID indicating each sentiment expression.

  The sensitivity expression ID is an ID for expressing a sensitivity expression having the sensitivity vector. For example, the Kansei expression standard form is used as the Kansei expression ID, or the Kansei information column of the Kansei information DB 40 stores the digitized Kansei expression ID, and the Kansei expression ID is used as the Kansei expression ID of the Kansei vector dictionary 60 as it is. Also good. Further, for example, if a sensitivity expression category is a sensitivity expression ID, a plurality of sensitivity expressions can be collectively expressed as one sensitivity vector. The sensitivity vector has a predetermined number of dimensions, and each dimension indicates a value with respect to a predetermined sensitivity axis.

  FIG. 7 shows an example of the sensitivity vector dictionary 60. In this example, the sensitivity expression standard form is used as the sensitivity expression ID. In addition, the sensitivity vector is defined such that the number of dimensions = 7, the first dimension axis = beauty (the value is larger as it is more beautiful), the second dimension axis = silence,..., The seventh dimension axis = strength. It shall be.

  The tabulation result accumulation DB 70 is for storing tabulation results for a specific search condition and tabulation conditions that have been searched and tabulated in advance, and an example is shown in FIG. Details of the contents will be described later.

  In addition, the text document input to the sensibility information extraction unit 10 is made into a database so that the text document can be searched using the document ID as a key, and the attributes of the author of the text document (author name, gender, age, hobby) , Occupations, etc.) may be prepared and used as a database in order to be searchable using the author ID as a key.

  FIG. 9 shows an example of processing in the sensitivity search unit 50, and here shows an example in the case of outputting a list of entertainers as search results in order of sensitivity to one author input with the author ID. Hereinafter, the operation of the sensitivity search unit 50 will be described in accordance with this.

  In this embodiment, to determine whether or not it is close to the sensibility, a sensitivity vector for each author is created, and the distance (in this embodiment, a known cosine distance is used. Using another distance measure) It is also good).

  The sensitivity vector of one author whose author ID is x is the result of searching all records with the author ID = x in the sensitivity information DB 40 or the sensitivity information DB 40 up to the maximum number of records specified by the author ID = x. However, for all the records, the sensitivity vectors corresponding to each sensitivity expression are obtained from the sensitivity vector dictionary 60, and the average of all the vectors for each dimension is set as the sensitivity vector of author ID = x.

  In the present embodiment, it is assumed that the author's occupation is held as the author category in the document information of the sensitivity information stored in the sensitivity information DB 40 without using the author DB. In addition, as shown in FIG. 8, it is assumed that at least the calculated sensitivity vector for each author whose author category is an entertainer is stored in the tabulation result storage DB 70 together with the author ID and the author category.

  The condition input unit 51 receives the input search condition and totaling condition and passes them to the search / totaling unit 52 (step S11 in FIG. 3).

  Here, various types of search conditions and tabulation conditions can be input. For example, all search conditions and aggregation conditions can be arbitrarily input. In addition, in order to perform a sensitivity search used for a specific application, a part of the search conditions and the totaling conditions are held in advance as fixed, and only the search conditions and the totaling conditions that can be arbitrarily input can be obtained via a graphical user interface or the like. You may make it input to.

  In the search conditions and tabulation conditions shown in FIG. 9A, only the author ID of the author to be searched can be arbitrarily entered as the search condition, and the author ID of the author whose profession is the entertainer, that is, the author category, is the other search condition. The author ID of the author who is the entertainer is fixed as a comparison target, and the sensitivity vector of each author ID that is the comparison target and the author's sensitivity vector of the entered search target author ID as the aggregation condition It is assumed that a fixed designation has been made such that the distances are calculated and ranked in order of increasing distance. In this example, it is assumed that “abc0005” is input as the search target author ID.

  Next, the search / aggregation unit 52 searches the sensitivity vector DB 60 and the aggregation result accumulation DB 70 together with the sensitivity information DB 40 based on the search conditions input from the condition input unit 51, and inputs the search results from the condition input unit 51. Based on the totaled condition, a total result is created (step S12 in FIG. 3).

  In the present embodiment, the sensitivity vector corresponding to the author ID to be compared is stored in the tabulation result accumulation DB 70 (FIG. 8), and therefore, there is no need to perform calculation, but the author ID = abc0005 that is the search target. It is necessary to calculate the sensitivity vector (the author category is not a celebrity). Here, when the document described by the author ID = abc0005 is only the document shown in FIG. 4A, when a record with the author ID = abc0005 is searched from the sensitivity information DB 60, the document shown in FIG. The sensitivity information shown is obtained as a result.

  The sensitivity vector dictionary 60 (FIG. 7) is searched using the sensitivity expression standard form of each sensitivity information shown in FIG. 4 (c-1) as a key, the sensitivity vector of each sensitivity expression is obtained, and the average is calculated. Author ID = abc0005 sensitivity vector. FIG. 9B shows the sensitivity vector of the author ID = abc0005 calculated in this way.

  Next, the sensitivity distance between the sensitivity vector of the author ID = abc0005 and the sensitivity vectors corresponding to each of the comparison-target author IDs searched from the tabulation result storage DB 70 (an upper limit may be set in advance) may be obtained. FIG. 9C shows the result of calculation, here calculated using a known cosine distance, and sorted in the order of the sensitivity distance.

  Finally, the result output unit 53 outputs the search result to a screen or the like in a predetermined format using the total result created by the search / total unit 52 (step S13 in FIG. 3).

  An example of the output result created from the tabulation result of FIG. 9C is shown in FIG. “Blog URL” and “I am writing this” (snippet including the emotional expression in the blog indicated by the blog URL) are stored in advance in the sensitivity information DB 40, or stored in the document DB. It can be obtained by searching the sensitivity information DB 40 with the author ID, specifying the document ID, and searching the document DB. The “entertainer name” can be obtained by providing a column in the sensitivity information DB 40 and storing it in advance, or by holding the author DB and searching the author DB with the author ID.

<Second Embodiment>
FIG. 10 shows a second embodiment of the sensitivity information extraction / retrieval apparatus according to the present invention. Here, in the first embodiment, a relationship extraction unit 14 is added to the sensitivity information extraction unit 10 and sensitivity is obtained as sensitivity information. Information about the object is output along with the expression.

  The relationship extraction unit 14 extracts, from the text document, those targeted by each emotional expression extracted by the emotional expression extraction unit 12. Extraction of related objects is possible by using techniques such as Hirano et al. “Extraction of semantic relations between proper expressions in text”, Natural Language Processing Society 13th Annual Conference, 2007, etc.

  Here, the text analysis information used for the relationship extraction includes the phrase / dependency information obtained by performing well-known dependency analysis in addition to the word information, and the unique information obtained by performing well-known specific expression extraction. Since the expression information is required, the text analysis unit 11 also performs dependency analysis and specific expression extraction processing, and outputs the phrase / dependency information and the specific expression information to the relationship extraction unit 14 as text analysis information. .

  In addition, in the sensitivity information generation unit 13, in the sensitivity expression unit extracted by the sensitivity expression extraction unit 12, the sensitivity information that is information related to the target extracted by the relationship extraction unit 14 together with the sensitivity expression information that is information related to the sensitivity expression. Kansei information including target information is generated and output.

  FIG. 4C-2 shows an example of sensitivity information including sensitivity object information as well as sensitivity expression information for the text document of FIG. Here, the sensibility expressions “chic” and “autumn” are expressed as “OO”.

  The sensibility search unit 20 is the same as that in the first embodiment. However, it differs from the case of the first embodiment in that sensitivity object information can also be used as a search condition and a totaling condition.

  According to the present embodiment, in addition to the author's sensitivity vector calculated in units of author ID, by taking the same method as that for calculating the author sensitivity vector in units of sensitivity object information (for example, sensitivity object standard form) By creating a target sensitivity vector and using the distance between the sensitivity targets, it is possible to search for a similar target.

The block diagram which shows 1st Embodiment of the sensitivity information extraction / retrieval apparatus of this invention Flow chart of Kansei information extraction process in Kansei information extraction unit Flow chart of search processing in Kansei search unit Explanatory drawing which shows an example of the process in the sensitivity information extraction part Explanatory drawing which shows an example of a sensitivity expression dictionary Explanatory drawing showing an example of Kansei expression extraction rules Explanatory drawing showing an example of a sensitivity vector dictionary Explanatory drawing which shows an example of totaling result accumulation DB Explanatory drawing which shows an example of the process in the sensitivity search part The block diagram which shows 2nd Embodiment of the sensitivity information extraction / retrieval apparatus of this invention

Explanation of symbols

  10: Kansei information extracting unit, 11: Text analyzing unit, 12: Kansei expression extracting unit, 13: Kansei information generating unit, 14: Relationship extracting unit, 20: Kansei expression dictionary, 30: Kansei expression extraction rule, 40: Kansei information Database (DB), 50: Kansei search unit, 51: Condition input unit, 52: Search / aggregation unit, 53: Result output unit, 60: Kansei vector dictionary, 70: Aggregation result accumulation database (DB).

Claims (9)

A sensitivity information extraction / retrieval device that extracts sensitivity information from a text document and creates a database, and performs information retrieval using this database,
A text analysis unit that performs text analysis on a text document and outputs text analysis information;
Using the text analysis information, a Kansei expression dictionary that registers information of words that can be Kansei expressions that are expressions representing impressions and feelings about an arbitrary object, and Kansei expression extraction rules that register patterns of Kansei expressions to be extracted A sensitivity expression extracting unit for extracting the sensitivity expression in the text document;
A sensitivity information extraction unit configured by a sensitivity information generation unit that generates and outputs sensitivity information including at least information related to the sensitivity expression in the sensitivity expression unit;
A sensitivity information database for storing sensitivity information output from the sensitivity information extraction unit;
A condition input part for receiving the input search condition and aggregation condition;
A search / aggregation unit that retrieves at least an emotion vector dictionary in which an emotion vector for each emotion expression is registered together with the sensitivity information database based on the search condition, and aggregates the search results based on the aggregation condition to create an aggregation result When,
A sensibility information extraction / retrieval device comprising: a sensibility search unit configured to include a result output unit that outputs a search result in a predetermined format using the aggregation result. Based on the search condition of the search target, search the sensitivity vector dictionary together with the sensitivity information database, obtain the sensitivity vector of each sensitivity expression related to the search target, calculate the average to find the sensitivity vector corresponding to the search target, and total The search / aggregation unit that calculates a sensitivity distance between a sensitivity vector corresponding to a comparison target based on a condition and a sensitivity vector corresponding to the search target using a cosine distance is provided. Kansei information extraction / retrieval device. In addition to the above
It has a summary result accumulation database that stores summary results for specific search conditions and summary conditions.
The sensibility information extraction / retrieval device according to claim 2, wherein the retrieval / aggregation unit reads out and uses the sensibility vector corresponding to the comparison target based on the aggregation condition from the aggregation result accumulation database. In addition to the above, the sensitivity information extraction unit
A relationship extraction unit that extracts a target of emotional expression from the text document;
The Kansei information generation / search unit according to any one of Claims 1 to 3, wherein the Kansei information generation unit generates and outputs Kansei information including at least the Kansei expression and information related to the target in the Kansei expression unit. apparatus. Kansei information extraction / retrieval method for extracting Kansei information from a text document into a database and performing information retrieval using this database,
A step in which the text analysis unit performs text analysis on the text document and outputs text analysis information;
Kansei expression extraction unit registered Kansei expression dictionary that registered the text analysis information, information of words that could be an expression of impressions and feelings related to any object, and Kansei expression patterns to be extracted Extracting emotional expressions in the text document using expression extraction rules;
A sentiment information generation unit that generates sentiment information including at least information related to the sentiment expression in the sentiment expression unit and outputs it to the sentiment information database;
A step in which the condition input unit receives the input search condition and aggregation condition;
The search / aggregation unit searches at least a sensitivity vector dictionary in which sensitivity vectors for each emotional expression are registered together with the sensitivity information database based on the search conditions, and totals the search results based on the aggregation conditions. A step to create,
A result output unit including a step of outputting a search result in a predetermined format using the tabulated result. The search / aggregation step searches the sensitivity vector dictionary together with the sensitivity information database based on the search condition of the search target, obtains the sensitivity vector of each sensitivity expression related to the search target, calculates the average, and corresponds to the search target The sensibility vector is obtained, and the sensibility distance between the sensibility vector corresponding to the comparison target based on the totaling condition and the sensibility vector corresponding to the search target is calculated using a cosine distance. Kansei information extraction / search method. The retrieval / aggregation step reads out and uses a sensitivity vector corresponding to a comparison target based on the aggregation condition from an aggregation result accumulation database that stores the aggregation result for the specific search condition and the aggregation condition. Kansei information extraction and retrieval method. In addition to the above
A relationship extraction unit including a step of extracting an object of emotional expression from the text document;
The Kansei information generation / retrieval step according to any one of Claims 5 to 7, wherein the Kansei information generation step generates and outputs Kansei information including at least the Kansei expression and information on the object in the Kansei expression unit. Method.   A program for causing a computer to execute each processing step of the sensitivity information extraction / retrieval method according to any one of claims 5 to 8.

Images