JP2001101222A - Data retrieving device and storage medium in which program for data retrieval is stored - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a retrieval result on which subjective similarity is sufficiently reflected from a conditional sentence subjectively expressed in a natural language by a user. SOLUTION: This data retrieving device holds a subjective evaluation information dictionary 5 formed by placing impression data for image information to be a retrieving object in a coordinate space to take a subjective similarity element as a coordinate axis and a subjective evaluation expression dictionary 4 formed by placing impression data of an impression expressing word to be a retrieval key in the coordinate space take the subjective similarity element as the coordinate axis. The impression expressing word is extracted from the conditional sentence for retrieval subjectively expressed by a natural language processing as a retrieval condition and corresponding impression data is extracted from the subjective evaluation expression dictionary 4. When plural impression expressing words exist, one or plural pieces of impression data integrated by an integration processing are generated as retrieval impression data. And the image information with the impression data with the closest Euclid distance to each piece of retrieval impression data is outputted as a retrieval result in the coordinate space to take the subjective similarity element as the coordinate axis.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] The present invention relates to a data search device,
In addition, the present invention relates to a storage medium storing a data search program, and relates to, for example, data search using a feeling expression.

[0002]

2. Description of the Related Art Various methods for searching for image information such as still images and moving images or contents including the same using impression expression words such as "soft" and "clear" have been conventionally proposed. The following two types of methods are mainly used as matching methods with information in a database or the like to be searched. As a first method, the impression given by each image information is measured in advance by an impression evaluation experiment or the like using an impression expression word pair, and impression data for each image information is created. Matching between the data and the input impression expression words is performed. On the other hand, as a second method, objective feature conditions such as the color and shape of image information that satisfy the impression word input at the time of the search are extracted, and objective feature data added to the image information in advance is extracted. And matching.

[0003]

When image information is matched by the first technique, impression expression words that can be used as keys in a search are limited to those used in an experiment, so that the It is necessary to remember the impression expression words that can be used in advance, or to search for and use the corresponding impression expression words from a separately created impression expression word table or the like. For this reason, it has not been possible to input image information to be searched for in an expression that matches the sense and sensibility felt by the user. Further, when performing matching using impression data, it is necessary to measure the impressions of all the image information in advance, which is extremely difficult when there is a huge number of image information items.

On the other hand, in the case of the second method in which matching is performed using objective feature data, even if a feature extraction technique for high-quality images or the like is employed, the extracted objective feature conditions are not necessarily intuitive. Subjective similarities due to impressions are not always sufficiently reflected, and the probability of obtaining search results satisfying the demand has been reduced. In addition, despite the high degree of subjective similarity, many pieces of image information that are not searched because of different objective feature data have been generated.

Accordingly, an object of the present invention is to provide a data search apparatus capable of obtaining a search result sufficiently reflecting subjective similarity from a conditional sentence subjectively expressed by a user in a natural language. I do.

[0006]

According to the first aspect of the present invention, there is provided a subjective evaluation information dictionary in which impression data for image information to be searched is placed in a coordinate space having a subjective similarity element as a coordinate axis; A subjective evaluation expression dictionary in which impression data for impression expression words serving as search keys are laid out in a coordinate space using a subjective similarity element as a coordinate axis, and a search condition sentence subjectively expressed in natural language as a search condition A retrieval condition sentence acquisition unit to be acquired, and an impression expression word extracted from the retrieval condition sentence acquired by the retrieval condition sentence acquisition unit, and the acquired impression expression words are stored in the subjective evaluation expression dictionary. Search impression data generating means for generating impression data for the search condition sentence, and impression data most similar to the impression data generated by the search impression data generating means. An image information retrieving means for retrieving image information from the subjective evaluation information dictionary, to achieve the above object and output means, the so provided in the data retrieval apparatus that outputs image information retrieved by the image information retrieval unit. In the invention according to claim 2, the impression data for each impression expression word stored in the subjective evaluation expression dictionary includes range data for specifying a range for each coordinate axis. According to the third aspect of the present invention, text data acquisition means for acquiring text data relating to new image information, and impression expression words are extracted from the text data acquired by the text data acquisition means, and stored in the subjective evaluation expression dictionary. Impression data generating means for generating impression data for the new image information from the impression data of the impression expression word, and the impression data generated by the impression data generating means as the impression data of the new image information. Image information adding means for registering in the evaluation information dictionary. In the invention according to claim 4, further, extraction means for extracting text data impression expression words, and impression expression words registered in the subjective evaluation expression dictionary and unregistered impression expression words in the impression expression words extracted by the extraction means. When an impression expression word exists, an impression expression word impression data generation unit that generates impression data for the unregistered impression expression word from the impression data for the registered impression expression word, and an impression expression word impression data generation unit. Means for registering the generated impression data in the subjective evaluation expression dictionary as impression data of the unregistered impression expression words. In the invention according to claim 5, the subjective evaluation information dictionary includes objective characteristic conditions for each image information, and the image information search unit obtains an objective characteristic condition from the search condition obtaining unit. , Image information is narrowed down using the objective characteristic condition as a constraint. In the invention according to claim 6, the subjective evaluation information dictionary includes objective characteristic conditions together with impression data for each image information, and is stored in the subjective evaluation information dictionary based on the objective characteristic conditions of new image information. A second image information adding unit that generates impression data for the new image information from the relationship between the obtained impression information of the image information and the objective feature information and stores the generated impression data in the subjective expression dictionary. In the invention according to claim 7, the search condition obtaining unit obtains the search condition from an external device connected to a network, and the output unit outputs image information to the external device connected to the network. In the invention according to claim 8, in a coordinate space having a subjective similarity element as a coordinate axis,
A function of creating a subjective evaluation information dictionary in which impression data for image information to be searched is placed, and impression data for impression expression words serving as search keys are placed in a coordinate space using a subjective similarity element as a coordinate axis. A function to create a subjective evaluation expression dictionary, a search condition sentence acquisition function to acquire a search condition sentence subjectively expressed in natural language as a search condition, and an impression from the search condition sentence acquired by this search condition sentence acquisition function A search impression data generating function for extracting expression words and generating impression data for the acquired search condition sentence from impression data of the impression expression words stored in the subjective evaluation expression dictionary; An image information search function for searching the subjective evaluation information dictionary for image information of impression data most similar to the impression data generated by The output function of outputting the image information retrieved by the image information retrieval unit, the data search program for realizing to a computer to record on a recording medium. A function for creating the subjective evaluation information dictionary, a function for creating the subjective evaluation expression dictionary, the search condition sentence acquisition function, and the search impression data generation function;
It may be a data search program, a data search program signal, a data search program transmission medium, or a data search program carrier for realizing the image information search function and the output function in a computer.

[0007]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, a data search device according to the present invention,
A preferred embodiment in a storage medium storing a data search program will be described with reference to FIGS. (1) Overview of Embodiment A subjective evaluation information dictionary 5 (see FIG. 6) in which impression data for image information to be searched is stored in a coordinate space having a subjective similarity element as a coordinate axis. Further, a subjective evaluation expression dictionary 4 (see FIG. 3) in which impression data of impression expression words serving as search keys are stored in a coordinate space having a subjective similarity element as a coordinate axis. The impression data of the subjective evaluation expression dictionary 4 includes a coordinate value for each coordinate axis and range data for specifying a range on each coordinate axis with the coordinate value as a center (see FIG. 4). Then, as a search condition, an impression expression word is extracted by natural language processing from a search condition sentence subjectively expressed by a natural language, and corresponding impression data is extracted from the subjective evaluation expression dictionary 4. When there are a plurality of impression expression words, one or a plurality of impression data integrated by the integration processing is generated as search impression data. Then, in a coordinate space having the subjective similarity element as a coordinate axis, image information having impression data whose Euclidean distance to each search impression data is the shortest is output as a search result. On the other hand, when adding new image information, an impression expression word is extracted from text data such as a description added to the image information to be added, and integrated into one impression data by an integration process. Impression data consisting of coordinate values obtained by removing the range data from the impression data is registered in the subjective evaluation information dictionary 5 as impression data of image information to be added. When an impression expression word that is not registered in the subjective evaluation expression dictionary 4 is extracted in the process of image information search and image information addition, if another impression expression word included in a sentence or text data used in the processing is extracted. One piece of impression data is generated by integrating the impression data, and is registered in the subjective evaluation expression dictionary 4 as impression data of the impression expression word.

(2) Details of Embodiment FIG. 1 conceptually shows the configuration of a data search apparatus according to this embodiment. As shown in the conceptual configuration diagram, the data search device includes an input unit 1, a search condition sentence acquisition unit 2,
Search impression data generation means 3, subjective evaluation expression dictionary 4, subjective evaluation information dictionary 5, image information search means 6, output means 7, image information addition means 8, and impression expression word addition means 9
It has. The input means 1 is for inputting a search condition sentence in a natural language desired by a user, and uses a keyboard or a voice recognition device. The search condition sentence obtaining means 2 obtains the search condition sentence input from the input means 1,
Alternatively, a search condition sentence transmitted from a wired or wirelessly connected external device is acquired using a communication unit. Examples of the external device include other personal computers connected to a LAN and an external personal computer connected to the Internet via the World Wide Web (WWW). It is configured to be able to acquire search condition sentences. The search impression data generating means extracts an impression expression word or the like from the search condition sentence acquired by the search condition sentence acquisition means and extracts impression data corresponding to the extracted impression expression word from the subjective evaluation expression dictionary 4 to obtain a search impression data. Generate data.

The subjective evaluation expression dictionary 4 stores, for each category, an impression expression word used to express an impression of a desired beer, such as an image or a sense, in a natural language when searching for beer. This is a dictionary classified and stored. Here, the impression expression word is not a physical feature of the image information to be searched, but a word (kansei word) expressing an intuitive impression by an adjective or an adjective verb such as "soft" or "clear". is there. Here, all the data formats that can express the relative relationship by subjective similarity can be handled as the image information to be searched.For example, in addition to images such as moving images and still images, music data, audio data, It also includes text data itself such as publications, other signal data, and content including these as elements. FIG. 2 conceptually shows the contents of the subjective evaluation expression dictionary 4. As shown in FIG. 2, the subjective evaluation expression dictionary includes beer, wine, sake,
It is classified into various categories for image information to be searched, such as movies, TV program information, travel information on the Web, and online shopping. Then, impression expression words used to express the impression of image information belonging to each category, specifically, expression words typified by adjectives and adjective verbs and degree adverbs (adverbs indicating the degree such as "somewhat" or "pretty") ), Nouns, proper nouns, and the like that are strongly associated with impression factors are collected in advance for each category and stored in the subjective evaluation expression dictionary.

On the other hand, for each category, an element representing a subjective similarity to the image information belonging to the category (a subjective similarity element) is extracted. For example, assuming that the category of the target image information is a movie, subjective similarity factors such as "thrill degree", "impression degree", and "comedy degree" are extracted as the properties of the work. Also, the category "Beer"
As shown in FIG. 2, “Koku” and “Kire”
Subjective similarity elements such as “umami”, “nodashi”, and “exhilaration” are extracted. Then, by designating the strength of the subjective similarity element for each impression expression word, the impression data for each impression expression word (a multi-dimensional Coordinate values (vectors) are determined, that is, for the collected impression expression words, each expression specifies the strength of a subjective similarity element, that is, indicates a specific range of the impression data space. The relative level value of each similarity element is given based on a hypothesis, for example, as shown in Fig. 2, the impression expression word "rich" for the category "beer" has a subjective similarity. Impression data is given in which the relative value of the sex element “Koku” is 3, the relative value of the subjective similarity element “Kire” is 0,.

FIG. 3 exemplifies a coordinate space having a subjective similarity element as a coordinate axis and impression data for each impression expression word laid out in the coordinate space. In FIG. 3, the coordinate space of beer is expressed using three subjective similarity elements of richness, sharpness, and umami as coordinate axes, and impression data for impression expression words “creamy”, “mellow”, and “dry” are placed. It was done. As shown in FIG. 3, each impression expression word uses a subjective similarity element extracted for image information belonging to the category as a coordinate axis,
It is represented by impression data (vector) having each coordinate value as an element.

As shown in FIG. 2, the actual impression expression word includes not only coordinate values indicating relative values of the respective subjective similarity elements but also range data indicating a range included in the impression expression word. Comes with. Therefore, an impression expression word is specified in an impression data space surrounded by a range centered on the coordinate value of each subjective similarity element (a fixed range centered on the coordinate value). FIG. 4 shows the impression data space of impression expression words as 2
It is represented by dimensional coordinates. As illustrated in FIG. 4, for example, the impression expression word “creamy” of the category “beer” has a range of plus or minus 2 around the coordinate value 4 with respect to the subjective similarity element “koku”, That is, it is specified in the range of 2 to 6, and plus or minus 2 with respect to the coordinate value −2 for the subjective similarity element “cut”.
, Ie, in the range of -4 to 0. As described above, the impression data of the impression expression word has a certain range with respect to each coordinate axis (subjective similarity element), so that the impression data is not represented by a point in the coordinate space but in a certain space in the coordinate space. (Volume), and image information included in the impression data space of the impression expression word can be listed as image information of a search candidate. By listing the image information in the impression data space of the impression expression words in this way, it is not necessary to determine the similarity of all the image information with the impression data, and the similarity of the listed image information with the impression data is eliminated. Can be determined, and the speed of the determination can be increased.

The subjective evaluation information dictionary 5 stores data relating to a search target, and stores impression data and auxiliary information for each image information, classified for each category. The category in the subjective evaluation information dictionary 5 is the same as the category in the subjective evaluation expression dictionary 4, and the search is performed within the same category. FIG.
Is a conceptual representation of the contents of the subjective evaluation information dictionary 5. As illustrated in FIG. 5, the impression data for each piece of image information is represented by a vector having a value on a coordinate axis composed of a common subjective similarity element as the impression data for the impression expression word. . In other words, the impression data of the image information “Asahi Super Dry” is a vector (3,4,2,3.5,3. 5,
…). Similarly, "Asahi Kuroshio", "Asahi First Lady", "Asahi Super Premium", "Asahi Draft Beer Dunk" (all are registered trademarks or trademarks of Asahi Breweries, Ltd.), "Kirin Lager Beer", "Kirin Ichiban Iris <Raw>"Beer","Kirin Ichiban Squeezer", "Kirin Brewery <Raw>", "Kirin Brow Meister" (all registered trademarks or trademarks of Kirin Brewery Co., Ltd.) are also vectors with the value of each subjective similarity element Is represented by Note that, unlike the case of impression expression words, each coordinate value of impression data for image information does not include range data. Therefore, image information is laid out in a multidimensional coordinate space having a subjective similarity element as a coordinate axis, and by specifying a region (impression data space) having a certain width in the coordinate space, the impression data space is obtained. It is possible to search (list up) the image information laid out in the inside. On the other hand, as additional information, objective feature conditions such as numerical data, target image data, and text data such as explanatory text are stored for each image information. In the attached information, various data that cause other impressions of the image information are stored as related information in various forms.

The objective characteristic condition stored in the subjective evaluation information dictionary 5 as additional information means physical characteristic data of image information, and refers to quantitative and measurable numerical data. Further, category information obtained by classifying image information according to its properties and uses also corresponds to the objective characteristic condition. For example, when an image is image information, colors and shapes correspond, and when a passenger car is image information, displacement, fuel consumption, weight, and the like correspond. The objective feature condition is used to narrow down the image information as a constraint when searching for the image information.

Various types of data that are factors that give an impression of image information include a person's name, a product name, etc., which have a strong impression of a certain kind, and the impression data of a primitive figure which is a component of an image. And so on. For example, when a movie work is used as image information, the impression of the leading actor may be more dominant than the impression of the work itself. In such a case, impression expression words extracted from the commentary of the movie work, In some cases, generating impression data of a work by quoting impression data assigned to a leading actor or the like may reflect subjective similarity more than impression data estimated from a scene image. , The name of the leading actor, and the like are stored as attached information.

As shown in FIGS. 2 and 5, the subjective evaluation expression dictionary 4 and the subjective evaluation information dictionary store category classification information such as nouns included in each category. For example, “beer” is stored as category classification information for determining the category “beer”, and “wine” is stored as category classification information of the category “wine”. The noun itself indicating each category is stored as category classification information. For example, for the category "Wine", "Bordeaux" or "Beaujolais Nouveau"
And conditional expressions such as “(grape + grape + grape) × brewing” are also stored as category classification words. Related terms and conditional expressions are also stored as category classification information for other categories. In this way, by storing the category classification information in the subjective evaluation expression dictionary 4, in the search processing, the image information addition processing, and the impression expression word addition processing, the matching nouns (or satisfying the conditional expressions) are included in the text data. If included, it can be automatically categorized into the corresponding category.

The image information search means 6 (FIG. 1) uses the search impression data (a set of subjective similarity element level values) generated by the search impression data generation means and objective feature conditions as search keys to perform a subjective evaluation. Various image information stored in the information dictionary 5 is searched. That is, image information registered in the subjective evaluation information dictionary 5 is narrowed down using objective characteristic conditions as constraints, and image information within the range indicated by the level value set is extracted therefrom. The similarity of the impression data is determined, and the image data is output as a search result in the order of image information having higher similarity.

The output means 7 outputs a search result searched by the image information search means 6. The output of the search result by the output unit 7 is performed by outputting the searched image information to a display device and printing it out to a printing device.

The image information adding means 8 generates impression data of the new image information and stores it in the subjective evaluation information dictionary 5 together with the accompanying information. Impression data of new image information is generated by extracting impression expression words stored in the subjective evaluation expression dictionary 4 from text data attached to the image information, and using the impression data of the impression expression words. Also. Impression data of registered image information is generated from objective characteristic conditions of image information registered in the subjective evaluation information dictionary 5 and objective characteristic conditions of new image information. As described above, since the image information can be newly added by the image information adding means 8 based on the text data and the objective characteristic condition, the subjective evaluation information dictionary can be easily enhanced.

The impression expression word adding means 9 generates impression data of the unregistered impression expression words and additionally registers them in the subjective evaluation expression dictionary 4. The addition of a new impression expression word by the impression expression word adding means 9 is performed in a sentence in a natural language input at the time of searching for image information, or when an impression expression word not registered in the image information adding means 8 is a text corresponding to image information. This is performed when an unregistered impression expression word exists in the data. As described above, the new impression expression words are automatically registered in the subjective evaluation expression dictionary 4 by the impression expression word adding means 9, so that the more the data search device is used, the more the subjective evaluation expression dictionary 4 becomes more complete. Become. In particular, the user's feelings and sensibilities for image information and the sentences expressing it are also diverse, so by collecting and adding new impression expression words in the search process, more user sensibilities and expressions can be obtained. The subjective evaluation expression dictionary 4 can be used.

FIG. 7 shows a specific system configuration of the data retrieval apparatus thus configured. As shown in FIG. 7, the data search device is configured by a computer system including a personal computer, a word processor, and the like. The data search device includes a control unit 11 for controlling the entire system as shown in FIG. The control unit 11 includes a keyboard 12 and a mouse 13 as input devices, a display device 14, a printing device 15, a storage device 16,
Storage medium drive 17, communication controller 18, input / output I / O
F19, a character recognition device 20, and a voice recognition device 21 are connected. The control unit 11 includes a CPU 111, a ROM 1
12, a RAM 113 is provided. ROM 112
It is a read-only memory in which various programs and data for the PU 111 to perform various controls and calculations are stored in advance.

The RAM 113 is a random access memory used as a working memory for the CPU 111. In the RAM 113, various areas for performing various processes such as a data search process, an image information addition process, and an impression expression word addition process according to the present embodiment can be secured.

The keyboard 12 constitutes an input means 1 and a search condition sentence obtaining means 2 for inputting a search condition sentence in a natural language within the own device, and includes a kana key, a numeric keypad for inputting a kana character, and various functions. Various keys such as a function key, a cursor key, and the like for executing the operation are arranged. The mouse 13 is a pointing device, and is an input device for designating a corresponding function by left-clicking a key, an icon, or the like displayed on the display device 14. As the display device 14, for example, a CRT or a liquid crystal display is used. The display device includes a keyboard 12 and a mouse 13
Is displayed, or the image information finally retrieved is displayed. The printing device 15 transmits the search condition sentence displayed on the display device 14,
This is for printing image information and the like as search results. Various printing devices such as a laser printer, a dot printer, an ink jet printer, a page printer, a thermal printer, and a thermal transfer printer are used as the printing device.

The storage device 16 is composed of a readable and writable storage medium and a drive device for reading and writing various information such as programs and data on the storage medium.
As a storage medium used for the storage device 16, a hard disk is mainly used, but a readable / writable storage medium among various storage media used in a storage medium driving device 17 described later may be used. Good. The storage device 16 stores the kana-kanji conversion dictionary 161 and the program storage unit 1
62, a data storage unit 163, a subjective evaluation information dictionary 5, a subjective evaluation expression dictionary 4, and other storage units (not shown) (for example,
A storage unit for backing up programs, data, and the like stored in the storage device 16. The program storage unit 162 stores various programs such as an image information search processing program, an image information addition processing program, and an impression expression word addition processing program according to the present embodiment, and kana characters input using the kana-kanji conversion dictionary 161. Various programs such as a kana-kanji conversion program for converting columns into mixed-kanji sentences are stored. The data storage unit 163 stores various data required by the system, such as data on the user.

The storage medium drive 17 is a drive for the CPU 111 to read a computer program or data including a document from an external storage medium. The computer programs and the like stored in the storage medium also include various processing programs such as image information search processing executed by the data search device of the present embodiment, and dictionaries and data used therein. Here, the storage medium refers to a storage medium in which a computer program, data, and the like are stored, and specifically, a magnetic storage medium such as a floppy disk, a hard disk, and a magnetic tape, and a semiconductor storage medium such as a memory chip and an IC card. , CD-ROM or M
O, PD (phase change rewritable optical disk) and other storage media from which information can be read optically, storage media using paper (and media having a function equivalent to paper) such as paper cards and paper tapes, and various other types A storage medium on which a computer program or the like is stored in the method is included. As a storage medium used in the data search device of the present embodiment, a storage medium such as a CD-ROM or a floppy disk is mainly used. The storage medium drive 17 can read data stored in the RAM 113 or the storage device 16 into a writable storage medium such as a floppy disk in addition to reading a computer program from these various storage media. It is.

When a search condition sentence stored in a floppy disk, a memory chip, an IC card, or the like is read through the storage medium driving device 17 and the image information search according to the present embodiment is performed, the storage medium driving device 17 It functions as search condition sentence acquisition means 2. In addition, when the text data of the image information to be added in the image information adding process is obtained from a predetermined storage medium using the storage medium driving device 17, the storage medium driving device 17 functions as a part of the image information adding unit 8. .

In the data search device of the present embodiment, the CPU 111 of the control unit 11 reads a computer program from an external storage medium set in the storage medium drive 17 and stores it in each unit of the storage device 16 (installation).
I do. When executing various processes such as the image information search process according to the present embodiment, the corresponding program is read from the storage device 16 into the RAM 113 and executed. However, instead of the storage device 16, the storage medium driving device 17 directly outputs the RAM 11 from an external storage medium.
It is also possible to read and execute the program in the third program. Further, depending on the data search device, the image information search processing program and the like of the present embodiment may be stored in the ROM 112 in advance, and the CPU 111 may execute the program. Further, various programs and data such as the document classification processing program of the present embodiment are transferred to the communication control device 18.
Alternatively, the program may be downloaded from another storage medium and executed.

The communication control device 18 is a control device for making a network connection between the data search device and various external electronic devices such as other personal computers and word processors. The communication control device 18 can access the data search device from these various external electronic devices, and can input a search condition sentence from the external electronic device. In this case, the communication control device 18 functions as the retrieval condition sentence acquisition unit 2. In addition, the communication control device 18 for acquiring text data and objective characteristic conditions for image information from an external electronic device via the communication control device 18 and newly registering the image information in the subjective evaluation information dictionary 5 includes: Functions as a part of the information adding means.

The input / output I / F 19 is an interface for connecting various devices such as a speaker for outputting voice and music. The character recognition device 20 is a device that recognizes characters written on paper or the like in various formats such as a text format or HTML, and is configured by an image scanner, a character recognition program, and the like. This character recognition device 20 functions as the search condition sentence obtaining means 2 and also functions as a part of the image information adding means. The voice recognition device 21 is connected to a microphone (not shown) that functions as voice capturing means for capturing voice, and recognizes voice captured from the microphone and converts the voice into text data. The speech recognition device 21 for recognizing the search condition sentence input from the microphone functions as the search condition sentence acquisition unit 2.

The data search device of the present embodiment is not only constituted by a computer system including a personal computer, a word processor and the like, but also by a LAN (local / local).
An area network), a computer (personal computer) communication host, a computer system connected to the Internet, and the like can also be used. Further, it is also possible to distribute the functions to the devices on the network and configure the data search device on the entire network. FIG. 8 shows a system configuration diagram when a data search device is configured by such a network. As shown in FIG. 8, the data search device includes a host device 30 for searching and adding image information and adding an impression expression word, and a plurality of client PCs 50 for inputting and transmitting a search condition sentence and transmitting to the host device 30. And a network 40 connecting the host device 30 and each client PC. The network 40 is mainly connected to the Internet, but can be connected to various networks 40 such as a LAN (local area network) and a computer network. A client PC 50 such as a personal computer is appropriately connected to the network 40.
0 is accessible.

The client PC 50 is constituted by a so-called personal computer system such as a personal computer, and is connected to the network 40 by WWW (World Wide Web) browser software.
(Internet) can be connected. On the other hand, the client PC 50 includes a control unit, a display unit, an input unit, an output unit, a communication control unit, a storage unit, and other devices.
The control unit of the client PC 50 processes and controls the entire device according to a predetermined program, and transmits a search condition sentence input from the input unit to the host device 30 via the communication control unit and the network 40. At the same time, the image information retrieved by the host device 30 is received and displayed on the display unit, stored in the storage unit, or printed out from the output unit.

On the other hand, the host device 30 includes a control unit 31 and a display unit 33, an input / output unit 34, a communication control unit 35, a storage unit 36, and other devices (not shown) via a bus line 32 such as a data bus. Is connected. The basic configuration of each of these units 31 to 36 is the same as that of the data search device shown in FIG.
The control unit 31 according to the present embodiment functions as a WWW server, and controls the control unit 11 of the data search device illustrated in FIG.
In addition, high-speed processing can be performed as compared with the control unit 11 of the client PC 50, and parallel processing can be performed to cope with access from a plurality of client PCs 50. Similarly, the communication control unit 35 can be connected to a plurality of ISDN lines and can perform processing at a higher speed than that of the client PC 50.

The processing operation of the data retrieval apparatus having the above-described configuration will now be described. FIG. 9 is a flowchart illustrating an image information search processing operation performed by the data search device. The image information search by the data search device is started when a user inputs a search condition sentence in a natural language. The user can also input a constraint condition such as an objective feature condition instead of or together with the search condition sentence. Examples of the conditional search sentence to be entered are "I want to find an elegant and refreshing beer.""I want to know a beer that is rich but less bitter." The user inputs a sentence that freely expresses the impression and sensibility of the image information that the user wants to search, such as "What is the end movie?" In addition, a user who cannot appropriately express an impression can also input a search condition sentence using a sequence of words such as “elegant refreshing beer”. Furthermore, it is also possible to express an objective feature condition and the like in the search condition sentence. The search condition sentence input by the user is obtained by the search condition sentence obtaining means 2 (step 11).

Next, an impression expression word, a degree adverb, an objective feature condition, and category classification information are extracted from the acquired search condition sentence by the search impression data generating means 3 (step 13). For example, from the search condition sentence “I want to find an elegant and refreshing beer”, the impression expression words “elegant” and “fresh” are extracted, and the category classification information “beer” is extracted. Then, it is confirmed whether or not the category classification information exists in the search condition sentence (step 15), and if it does not exist (; N), a category input screen is displayed (step 17). In other words, the search impression data generating means displays a list of each category name registered in the subjective evaluation expression dictionary 4, and urges the user to select a category. The search impression data generating means 3 specifies the category selected on the category input screen, and if the search condition sentence includes the category classification information (step 15; Y), the search impression data generation means 3 corresponds to the category classification information. The category to be specified is specified (step 19).

Next, the retrieval impression data generating means 3 extracts the impression data of the extracted impression expression words from the subjective evaluation expression dictionary 4 (step 21). That is, the level value and range data for each subjective similarity element are extracted from the subjective evaluation expression dictionary 4 for each impression expression word. If the impression expression words extracted from the search condition sentence include an impression expression word that has not been registered in the subjective evaluation expression dictionary 4, an impression expression word addition process described later is executed. If a degree adverb is present in the search condition sentence (step 23;
Y), the impression data of the corresponding impression expression word is corrected by the degree adverb so as to meet the required degree (step 2).
5). The correction of the impression data by the degree adverb is performed by changing the absolute value of the level value of the impression data so as to change. For example, if the search condition sentence is "very simple ..."
, The level value for the subjective similarity element “Koku” of the impression expression word “Simple” is corrected from “−3” to “−3.5” whose absolute value increases. In addition,
Level value correction by degree adverb is "very""very"
In the case of emphasized words such as "messy" or "super", correct so that the absolute value increases, and "slightly""slightly""slightly"
In the case of a word that weakens the tone, etc., correction is made so that the absolute value decreases.

In the present invention, the correction value in the present embodiment is uniformly changed by 0.5. However, the change value (correction width) may be changed according to each adverb. In this case, a correspondence table of each degree adverb and the change value is stored in the subjective evaluation expression dictionary 4.

Next, the search impression data generating means 3 generates search impression data from the impression data of each impression expression word (the corrected impression data if corrected) by an integration process described later (step 27). . In the integration processing in the image information search processing operation, as described later, a plurality of impression data to be integrated are aggregated into one or a plurality of (the number of integrated objects or less) impression data, and these are obtained as search impression data.

After the search impression data is generated, the image information search means 6 extracts image information candidates and searches for image information by narrowing down. In other words, the image information search means 6 searches the subjective evaluation information dictionary 5 for image information candidates included in the impression data space occupied by all the generated search impression data in the coordinate space having the subjective similarity element as a coordinate axis. (Step 29). In this way, not only is the search impression data represented at the point specified by the level value of each subjective similarity element, but also by using the impression data space consisting of the range given to the level value, Can be narrowed down in advance, the number of similarities to be calculated thereafter can be reduced, and the search can be speeded up.

Next, the image information search means 6 determines whether or not an objective characteristic condition exists in the condition search sentence acquired by the search condition sentence acquisition means 2 (step 31). When there is an objective feature condition (step 31; Y), the extracted image information candidates are narrowed down to image information candidates satisfying the objective feature condition as constraint conditions (step 33). For example, in the case of beer, "350m
If there is an objective characteristic condition of "1 and unit price of 150 yen or less",
By referring to the attached information of the extracted image information candidates, the image information candidates are narrowed down to image information candidates satisfying this condition.

Then, the subjective evaluation information dictionary 5 determines similarity from the impression data of each image information candidate and the level value of the search impression data. That is, the subjective evaluation information dictionary 5
Is the similarity S of the two vectors based on the Euclidean distance between the level value (vector) of the subjective similarity element excluding the data in the range of the entire search impression data and the impression data (vector) of each image information candidate. (Step 3
5).

The subjective evaluation information dictionary 5 uses the image information having the highest similarity S to the calculated search impression data as a search result, and displays it by the output means 7 (step 3).
7). Note that, in addition to displaying only the image information having the highest similarity S, a plurality of pieces of image information may be output in order of the image information having the highest similarity. Also,
Only the image information having the highest similarity S may be displayed so that the user can select to display another image information candidate.

By looking at the search results displayed on the display device, the user can give a request for improvement by additional input with respect to the searched image information. For example, an adverb and impression expression word representing a comparison such as “more a bit more dynamic” are input. If there is this correction request (step 3
9; Y), the correction request sentence is acquired by the search condition sentence acquisition means 2 (step 41), and the degree adverb is extracted by the search impression data generation means 3 (step 43). Then, Step 25
And the search is performed again by correcting the impression data using the degree adverb. In this way, by repeatedly performing the re-search using the corrected sentence, it is possible to reach the image information desired by the user.

On the other hand, when there is no correction request for the output image information (step 39; N), the image information search processing ends.

Next, the image information adding process for enriching the database of the image information stored in the subjective evaluation information dictionary 5 will be described with reference to the flowchart of FIG. When adding image information, a user or a system administrator inputs text data related to the image information or an objective feature condition of the image information. Based on the input, the image information adding means 8 acquires the input text data (step 51), and extracts an impression expression word, a degree adverb, an objective feature condition, and category classification information from the acquired text data. (Step 53). Then, it is checked whether or not the category classification information exists in the text data (step 5).
5) If not present (; N), a category input screen is displayed (step 57). That is, the image information adding unit 8 displays a list of each category name registered in the subjective evaluation information dictionary 5, and urges the user to select a category. The image information adding means 8 specifies the category selected on the category input screen, and if the search condition sentence includes category classification information (step 5).
5; Y) A category corresponding to the category classification information is specified (step 59).

Next, the image information adding means 8 checks whether at least one impression expression word registered in the subjective evaluation expression dictionary 4 exists in the text data (step 60). When an impression expression word exists (step 60;
Y), the image information adding means 8 extracts the impression data of the impression expression words from the subjective evaluation expression dictionary 4 (step 6).
1). When an impression expression word that has not been registered in the subjective evaluation expression dictionary 4 exists among the impression expression words extracted from the text data, an impression expression word addition process described later is executed. When the degree adverb is present in the text data (step 63; Y), the level value and the range data of the impression data of the corresponding impression expression word are expressed by the degree adverb.
Correction is made to meet the required degree (step 65). The correction using the degree adverb is performed in the same manner as the correction in the image information search processing.

Next, the image information adding means 8 generates impression data for the image information from each impression data (step 67). That is, the image information adding means 8
Is the level of impression data corresponding to the impression expression word (corrected impression data if corrected by the degree adverb) when only one impression expression word is extracted from the text data. The value is quoted as impression data for the image information. For example, the image information “Kirin Brewery Factory <Raw>” in the subjective evaluation information dictionary 5 shown in FIG. 5 is created corresponding to this case, and the impression expression word “light” in the text data is used. The respective level values of the impression data of the impression expression word “smooth” in the subjective evaluation expression dictionary 4 shown in FIG.
Impression data of "Kirin Brewery <Raw>" has been created.

On the other hand, when there are a plurality of impression expression words in the text data, impression data for the image information is generated by the integration processing described later. For example, suppose that the image information of the new additional contrast is "Kirin Half & Half <Raw>" and the accompanying text data is a sentence "Rich aroma unique to dark beer". From this text data, "unique to black beer", "rich" and "fragrant" are extracted as impression expression words.
Since "rich" and "fragrant" are registered in the subjective evaluation expression dictionary 4, the impression data of both of them is extracted. The level value is described immediately after each subjective similarity element, and the range data is described in parentheses after the level value.

"Rich"; (Rich 3 (2), sharp 1.
5 (1), umami 3 (2), throat 2 (2), exhilaration -1 (1), ...) "Savory"; , Throat 1 (1), Exhilaration 1 (1),
…)

By specializing and integrating the two impression data, the impression data based on “fragrant” (Koku 4 (2), Kire 0 (0), Umami 2 (2), Nodoshi 1
(1), one exhilaration 0 (0)) is generated. In the integration processing in the image information addition processing operation, the impression data based on one impression expression word is always aggregated from a plurality of impression data to be integrated, and each of the aggregated impression data except for the range data is removed. From the coordinate values using the similarity element as a coordinate axis, impression data (Koku 4, Kirei 0, Umami 2, Throat 1, Exhilaration 0, ...) of the image information "Kirin Half & Half <Raw>" is generated.

In step 60, if there is no impression expression word registered in the subjective evaluation expression dictionary 4 in the text data (N), the image information adding means 8 generates impression data from objective characteristic conditions. (Step 7
1). That is, the image information adding means 8 performs a multiple regression analysis of the impression data of the registered image information and the objective characteristic conditions attached thereto, and calculates the impression data when the objective characteristic conditions are determined. First, a multiple regression coefficient is obtained in advance. This multiple regression coefficient is obtained for each category. When the impression data of the image information is generated from the objective feature condition, the objective feature condition is used as an explanatory variable, and the coordinate value (level value) for each coordinate axis (subjective similarity element) of the impression data is used as an objective variable. Perform analysis. The multiple regression coefficient calculated here is a parameter.

The case of the category "passenger car" will be described as an example. Now, the subjective evaluation information dictionary 5 stores objective data such as "displacement", "fuel consumption", "weight", and impression data in each of the pre-registered passenger car information (image information). It is assumed that The image information adding means 8 obtains a multiple regression coefficient in advance so that impression data is calculated when objective characteristic conditions are determined by multiple regression analysis using the registered passenger car information. Specifically, objective characteristic conditions such as displacement, fuel consumption, and weight are used as explanatory variables, and coordinate axes of impression data (subjective similarity factors “compactness”, “operability”, “comfort”, “sporty degree”
The analysis is performed using the coordinate values (level values) for each of the above as target variables. Then, the image information adding means 8 calculates impression data for the newly added passenger car information (image information) from objective characteristic conditions such as displacement, fuel consumption, and weight.

The impression data for the new image information generated as described above is transferred to the image information adding means 8.
Is stored in the subjective evaluation information dictionary 5 together with the extracted auxiliary information such as the objective characteristic conditions, and the process is terminated.

As described above, new image information can be easily created from text data and objective characteristic conditions, and the impression data can be easily registered in the subjective evaluation expression dictionary 4, so that even new products can be easily created. Can be searched.

Next, the process of adding an impression expression word will be described with reference to the flowchart of FIG. This impression expression word addition processing is performed when an unregistered impression expression word exists in the condition search sentence placed in the image information search processing, and when an unregistered impression expression word exists in the text data in the image information addition processing. In this case, it is automatically executed based on a command from the search impression data generating means 3 or the image information adding means 8. The impression expression word adding means 9 acquires text data (a search condition statement or text data such as a catalog) and a category from the search impression data generating means 3 or the image information adding means 8 (step 81). Then, the impression expression word adding means 9 extracts impression expression words (both unregistered impression expression words and registered impression expression words) from the text data (step 83), and calculates the number of registered impression expression words. Check to determine whether there is only one word (step 8)
5). If the registered impression expression word is only one word (step 85; Y), the impression expression word adding means 9 reads the impression data of the registered impression expression word from the subjective evaluation expression dictionary 4, and outputs the unregistered impression expression word. It is quoted as impression data of the expression word (step 87).

On the other hand, when there are a plurality of registered impression expression words in the text data (step 85; Y), the impression expression word adding means 9 extracts the impression data for each impression expression word from the subjective evaluation expression dictionary 4. (Step 89). Then, the impression expression word adding means 9 generates new impression data from the extracted impression data by an integration process described later (step 91). In the integration processing in the impression expression word addition processing operation, a plurality of impression data to be integrated are always aggregated into impression data based on one impression expression word, and the aggregated impression data becomes new impression data.

The impression expression word adding means 9 converts the impression data (step 87) quoted from the unregistered impression expression words and the registered impression expression words or the new impression data (step 91) generated into a subjective evaluation expression dictionary. Then, the data is stored in the corresponding category of No. 4 and the process ends.

Next, a description will be given of an integration process for consolidating each impression data for a plurality of impression expression words into one or a plurality of representative impression data. In the present embodiment, two methods of specialized integration and generalized integration are prepared as a specific method of integrating a plurality of impression data. Specialized integration is an integration method that aims to combine multiple impression expressions in a direction to emphasize each evaluation element more and to add a special meaning to common parts. What is generalized integration?
This is an integration method that aims to aggregate multiple impression expressions in a direction that minimizes the prominence of each evaluation element and a direction that removes a special meaning and removes common parts. In general, the number of coordinate axes effective by specialized integration (Each element of impression data vector (Subjective similarity element)
Among them, the coordinate axes having range data> 0) are maintained as much as possible, and the number of effective coordinate axes is reduced due to the general integration.

In the present embodiment, of the two integration processes, specialized integration is used in each of the above-described image information search process, image information addition process, and impression expression word addition process. Then, it is possible to switch to generalized integration by setting options.

In the integration processing, (I) when retrieving image information by a sentence containing an impression expression word (image information retrieval processing), (II) when estimating impression data of image information accompanied by text data (image information addition processing) ), (I
II) Estimating the level value (coordinate value of impression data) of an unregistered impression expression word included in a sentence (impression expression word addition processing) is performed in three phases. The integration processing (specialized integration processing and generalized integration processing) includes primary integration processing performed in common to these three phases, and secondary integration processing that is a part of the integration processing excluding the primary integration processing. There is. In the secondary integration process, the integration result of the primary integration is used, and is performed for each phase. The primary integration process is a process of integrating (integrating) impression data of a pair of impression expression words, and includes a primary specialized integration process and a primary generalization process. The above (I) to (II)
In any case, the integration processing in each phase of I) is performed by integration means (not shown), and this integration means functions as search impression data generation means in phase (I),
In the phase (II), it functions as an image information adding means, and in the phase (III), it functions as an impression expression word adding means. Then, in the configuration shown in FIG. 7, the control unit 11 operates as an integrating unit, and in the configuration shown in FIG. 8, the control unit 31 operates as an integrating unit.

FIG. 12 is a flowchart showing the processing operation of the primary specialization integration processing. As shown in this figure, first, the integrating means acquires each impression data of two impression expression words to be integrated (a reference impression expression word and another impression expression word) from the subjective evaluation expression dictionary 4 (step 101). Here, the method of determining the reference impression word is different in each of the above aspects, and will be described in the later-described secondary integration processing.

Next, the integrating means extracts a comparison target axis from the acquired range data of each impression data (step 10).
3). That is, the integrating unit checks the value of the range data for each coordinate axis of each subjective similarity element of the two impression data to be integrated, and extracts a coordinate axis whose range data of both impression data is not 0 as a comparison target axis.

Then, the integrating means extracts the local axis from the extracted axes to be compared (step 105). That is, the integration unit checks the sign of the coordinate value of each comparison target axis of both impression data, and extracts a comparison target axis of the same station (the coordinate values are both positive or both negative) as the local axis. In addition,
A comparison axis whose sign of the coordinate value is a game (one of the coordinate values is plus and the other is minus) is a game axis.

Next, the integrating means obtains a value obtained by dividing the number of the local axes by the number of the axes to be compared. This value is calculated by dividing the value by a predetermined set value H (for example, 60%, 65%, 70%). , 75%,
It is determined whether it is 80%, 85%, n% (n is a variable that can be changed in the range of 1 to 100) or more (step 1).
07). If the ratio of the local axes is less than the set value H (step 107; N), the integrating means terminates the processing assuming that the specialization integration is not established (successful) because the contents of both impression data are too far apart. Return to routine.

On the other hand, if the ratio of the local axes is equal to or greater than the set value H (step 107; Y), the integrating means calculates the sum of the squares of the coordinate values of each axis to be compared for both impression expression words (impression data) ( Step 109). The integrating means determines whether the calculated values of the sum of the squares of the two coordinate values are the same (step 1).
11), if they are not the same (step 111; N),
The impression expression word having the larger coordinate value sum of squares is used as a base (step 113).

On the other hand, if the values of the sum of squares of both coordinate values are the same (step 111; Y), the integrating means checks the number of coordinate axes whose range data is not 0 for the impression data of both impression expression words (step 115). If the number of coordinate axes of the range data $ 0 is not the same (step 115; N), the impression expression word having the larger number of coordinate axes is used as the base (step 11).
7). If the number of coordinate axes of the range data # 0 is also the same (step 115; Y), the reference impression expression word is used as a base (step 119).

When the base impression expression word is obtained, the integrating means eliminates (clears) the value of the game axis from the impression data of the base impression expression word. That is, both the coordinate value of the game axis and the range data are set to 0. Then, the impression expression word serving as the base and the impression data with the value of the game axis set to 0 are integrated (step 121), and the process returns to the main routine.

Next, the primary specialization integration process described above will be described based on a specific example. Now, in the process of generating impression data (category: beer) from a plurality of impression expression words in the attached text data, the impression expression word “fresh” is used as a reference impression expression word, and another impression expression word “elegant” is used. Shall be integrated. First, the impression data of both impression expression words “fresh” and “elegant” are compared with the subjective evaluation expression dictionary 4 in FIG.
(Step 101). The acquired impression data are shown below. The numbers shown next to the coordinate axes (bodied, sharp, etc.) based on the subjective similarity elements are the coordinate values, and the numbers in parentheses () are the range data. In addition, the number of each subjective similarity element shown in FIG. 2, that is, the number of coordinate axes is determined by each category, and there are five or more coordinate axes of category; beer. The description will be made assuming that the five axes of feeling are provided. In addition, a description will be given assuming that the set value H of the local axis ratio that determines the success or failure of integration is 60%.

"Refreshing"; Koku-3 (2), Kirei-
1 (1), Umami 0 (0), Throat 1 (1), Exhilaration 3 (2) "Elegant"; ), Exhilaration 0 (0)

The range data of both impression data is examined, and for the coordinate axis “Koku”, the range data of the impression expression word “fresh” is 2, and the range data of the impression expression word “elegant” is 2; Since the range data is not 0, it is extracted as a comparison target axis. Similarly, coordinate axes "Kire""Nodoshi"
Is also extracted as the axis to be compared because both range data are not 0. On the other hand, for the coordinate axis "Umami",
Although the range data of the impression expression word “elegant” is 2 and is not 0, the range data of the impression expression word “refreshing” is 0, so that the axis becomes a non-comparison target. Similarly, the coordinate axis “exhilaration” is also an axis to be compared. As a result of examining the above range data, the number of comparative symmetry axes is three.

Next, the local axis is extracted from the comparison target axes "Koku", "Sharpness" and "Nodoshi". First, the target axis “Koku”
The coordinate value of the impression expression word "refreshing" is -3
Since the coordinate value of the impression expression word “elegant” is 3 and the signs of the two do not match, the comparison target axis “koku” is the game axis. Similarly, the coordinate value of the impression expression word “refreshing” is −1 and the “elegant” is 1.5
So it is a game axis. On the other hand, the comparison target axis "Nodoshi"
The coordinate value of the impression expression word "refreshing" is 1,
The coordinate value of “elegant” is 2, and the signs of both are coincident (both are plus), so that they are coaxial. Note that the coordinate value is-
The comparison target axis where n and -m (n, m> 0) are both co-located axes because their signs are both minus. As a result, the number of local axes is one.

Next, when the ratio of the local axis to the axis to be compared is examined, (1/3) × 100 = 33%, the set value H
= 60%, so integration is not possible. That is, the integration of the impression expression words "refreshing" and "elegant" was not established.

Next, a case where the impression expression words “elegant” and “savory” are integrated will be described. Note that “elegant” is used as a reference impression expression word. Impression data of both impression expression words is extracted from the subjective evaluation expression dictionary 4 of FIG. 2 as follows.

"Elegant"; rich 3 (2), sharp 1.5
(1), umami 3 (2), throat 2 (2), exhilaration
0 (0) "fragrant"; rich 4 (2), sharp 2 (2), umami 2 (2), throat 1 (1), exhilaration -1 (1)

When the impression data is examined, the range data of the impression expression word “elegant” is 0 for the coordinate axis “exhilaration”.
Therefore, it becomes a non-comparison target axis. Other coordinate axes "Koku"
As for “Kire”, “Umami”, and “Nodoshi”, since the range data of both the impression expression words “elegant” and “savory” are not 0, they become comparison axes. Therefore, the number of axes to be compared between the impression expression words “elegant” and “savory” is four.

Next, when examining the local axis, the comparison target axes "Koku", "Umami" and "Nodoshi" are the same local axis because the signs of the coordinate values of both impression expression words coincide with each other. Is the game axis because the signs of the two coordinate values do not match. Therefore, the number of co-axial axes of the impression expression words “elegant” and “savory” is three.

Then, the ratio of the local axis to the axis to be compared is (3/4) × 100 = 60%, and the set value H = 60%
Since the above is satisfied, integration becomes possible. That is,
The integration of the impression expression words “elegant” and “savory” is successful.

Next, when the sum of squares of the coordinate values of the axis to be compared with the impression expression word “elegant” is obtained, (3 × 3) + (1.5 × 1.
5) + (3 × 3) + (2 × 2) = 24.25 Similarly, the sum of the squares of the coordinate values of the impression expression word “fragrant” is (4
× 4) + ((-2) × (-2)) + (2 × 2) + (1 ×
1) = 25. Therefore, as a result of the specialized integration process, the impression expression word “savory” having the larger sum of the squares of the coordinate values becomes the base (step 113).

If the sum of the squares of the coordinate values of the two impression expression words “elegant” and “fragrant” is equal, the number of coordinate axes whose range data is not 0 is checked. Then, the impression expression word “elegant” is 4, and “savory” is 5. Therefore, the impression expression word "savory" having a large number of coordinate axes of the range data $ 0 is the basis (step 117). Further, when the sum of the squares of the coordinate values is equal and the number of coordinate axes of the range data $ 0 is the same, the reference impression expression word "elegant" is used as a base (step 119).

Then, from the impression data of the base impression expression word “fragrant”, the coordinate value and range data of the game axis “clear” are cleared to 0, which is the integrated result. That is, when the impression expression words “elegant” and “savory” are specialized and integrated, the impression expression words “savory” and the impression data (Koku 4
(2), sharpness 0 (0), taste 2 (2), throat 1
(1), exhilaration-1 (1)) is obtained as an integrated result.

Next, the first-order generalized integration will be described.
FIG. 13 is a flowchart showing the processing operation of the first generalized integration processing. It should be noted that many parts are common to the specialized integration described in FIG. 12, and the same processing parts are denoted by the same step numbers, and the different but corresponding processing parts are denoted by the same step numbers. A distinction is made by adding b to the end. Specifically, the method of determining the base impression expression word (steps 113, 117, and 119) is different from the method of determining the integration result (impression data) from the base impression expression word (step 121). Therefore, the description of the same processing portion will be omitted as appropriate, and different portions will be mainly described.

In the generalized integration, the integration means extracts the impression data of the impression expression word to be integrated (step 101), extracts the axes to be compared (step 103), Extraction (step 105) is performed, and when the ratio of the local axis is equal to or more than the set value H (step 10).
7; Y), the sum of the squares of the coordinate values of the axes to be compared is calculated (step 109), and it is determined whether or not the two sums of the squares of the coordinate values are the same (step 111). The processing up to this point is the same for generalized integration and specialized integration. Therefore, it is determined whether or not integration of the same impression expression word is established (Step 107 is Y → established,
Or N → not satisfied) is the same.

When the calculated values of the sum of the squares of the two coordinate values are not the same, the integrating means (step 111;
N), based on the impression expression word having the smaller sum of squares of coordinate values (step 113b). On the other hand, if the values of the sum of the squares of both coordinate values are the same (step 111; Y), the integrating means checks the number of coordinate axes whose range data is not 0 for the impression data of both impression expression words (step 115). If the number of coordinate axes of $ 0 is not the same (step 115;
N), based on the impression expression word having the smaller number of coordinate axes (step 117b). If the number of coordinate axes of the range data # 0 is also the same (step 115; Y), the impression expression word that is not the reference impression expression word is used as a base (step 1).
19b).

When the base impression expression word is obtained, the integrating means eliminates (clears) the values of the non-comparison target axis and the game axis from the impression data of the base impression expression word. That is, both the coordinate value and the range data are set to 0 for the comparison target axis and the game axis. Then, the base impression expression word and the impression data with the values of the non-comparison target axis and the game axis set to 0 are integrated (step 121b), and the process returns to the main routine.

As to a specific example of the generalized integration processing performed in this way, a case where the impression expression words “elegant” and “savory” that are integrated in the specific example of the specialized integration processing are generalized and integrated will be described. explain. When both impression expression words are generalized and integrated, by steps 101 to 111,
The sum of the squares of the coordinates of “elegant” is 24.5, and it is “savory”
Is smaller than 25, the impression expression word “elegant” is used as the base (step 113b). Then, from the impression data of the impression expression word “elegant”, the coordinate value and the range data of the non-comparative axis “exhilaration” and the game axis “cut” are set to 0 as an integrated result. That is, when the impression expression words “elegant” and “savory” are specially integrated, the impression expression word “elegant” and the impression data (Koku 3 (2), sharp 0
(0), taste 3 (2), throat 2 (2), exhilaration 0
(0)) is obtained as the integration result.

In the generalized integration, if the sum of the squares of the coordinate values of the two impression expression words “elegant” and “savory” is equal, the number of coordinate axes whose range data is not 0 is checked, and the impression expression word “elegant” is obtained. In step 4, since "savory" is 5, the impression expression word "elegant" having a small number of coordinate axes is used as a base (step 117b). Further, when the sum of squares of coordinate values is equal and the number of coordinate axes of the range data $ 0 is the same,
The impression expression word “fragrant” which is not the reference impression expression word is used as a base (step 119).

Next, using the result of the primary integration, the secondary integration in each of the above-described (I) image information search processing, (II) image information addition processing, and (III) impression expression word addition processing will be described. I do. This secondary integration is common to the case where the primary integration uses either the primary specialized integration or the primary generalized integration. When three or more impression expressions are present in the text to be processed, the integration process of the pair of impression expression words is repeated in the following procedure.

(1) In the integration process, primary integration is performed between the reference impression expression word and the impression expression word located closest to the reference impression expression word in the processing direction. Here, the reference impression expression word and the processing direction selected at the beginning of the integration processing vary depending on the above three aspects. That is,
(I) In the image information search process, an impression expression word that appears first from the beginning of a sentence to be a query is an initial reference impression word. The processing direction is from the beginning of the sentence to the end of the sentence. (II) In the image information adding process, the impression expression word that appears first from the head (text start) of the text data is the initial reference impression word. The processing direction is from the head of the text data to the end of the data. (III) In the impression expression word addition process, the impression expression words adjacent before and after the unregistered impression expression words are set as reference impression expression words.
Therefore, in the impression expression word addition processing, there are at most two reference impression expression words. Regarding the processing direction, the processing direction for the reference impression expression word before (at the beginning of the sentence) of the unregistered impression expression word is the head of the sentence, and the processing direction for the reference impression expression word after (at the end of the sentence) is at the end of the sentence.

(2) If the integration is successful in the primary integration process, the result of the integration (impression expression word as a base) is used as a new reference impression expression word. When the integration is not successful in the primary integration processing (when the local axis ratio <H), the impression expression word that appears next in the processing direction of the current reference impression expression word, that is, both unsuccessful integration words The impression expression word that is not the current reference impression expression word among the impression expression words is set as a new reference impression expression word. Then, integration of the new reference impression expression word and an impression expression word located at a position close to the new reference impression expression word in the processing direction (the next impression expression word in the processing direction) is performed as the next process. However, when the integration is not performed in the impression expression word adding process, the integration process ends there.

(3) Such processes (1) and (2) are repeatedly performed until a certain sentence break. The certain sentence segmentation is set to, for example, one sentence segmentation for each punctuation, two sentences one segmentation for two punctuation segments, or one segmentation for one paragraph to limit the range of the repetitive processing. In addition, the unit of the delimiter is usually set to one delimiter per sentence as an initial value, and can be changed to another delimiter. In the case of the image information search processing and the image information addition processing, the impression data itself of the impression expression word that has never been involved in the success of the primary integration, and the last if the primary integration of one or more consecutive times succeeds. The impression expression words and the impression data obtained by the primary integration are obtained as a result of the integration processing performed within a certain break. On the other hand, in the case of the impression expression word adding process, two integration processes of the unregistered impression expression word in the sentence head direction and the sentence end direction are performed within a certain break, and succeeded immediately before the unsuccessful primary integration. The impression expression words and impression data obtained by the next integration are obtained as a result of the integration processing in each direction.

(4) If there is a next sentence break, the above (1) to
The process of (3) is newly performed.

As described above, a plurality of integration results may be obtained in the process of repeating the integration. In such a case, in the image information search processing, all integrated results are used for the search in order to list many image information candidates.
On the other hand, in the image information information addition processing and the impression expression word addition processing, it is necessary to uniquely determine the impression data. For this reason, in the impression expression word adding process, primary integration is performed again on the integration result of the integration process 1 performed toward the beginning of the sentence and the integration result of the integration process 2 performed toward the end of the sentence.
Impression data obtained as a result of the integration is given as impression data of an unregistered impression expression word.

In the case where the primary integration of both the sentence head direction and the end of the sentence is not established in the impression expression word addition processing and in the case where a plurality of integration results exist in the image information addition processing, the next final selection Any one of the impression data is selected by the processing, and is set as a final integrated result.
That is, a plurality (two in the case of impression expression word addition processing)
For each impression expression word (if the impression expression word is obtained as a result of successful primary integration of one or more successive times,
In the next integration, the sum of the frequency of appearance of each of the impression expression words targeted for integration is compared, and the impression data of the larger impression expression word (the impression data generated by each primary integration; the same applies hereinafter) is finally integrated. Result. Here, in the present embodiment, the appearance frequency is sequentially counted from the input text. In addition, the appearance frequency may be continuously counted, registered and updated in the dictionary, and the appearance frequency may be determined from the past appearance results registered in the dictionary.

When there are a plurality of impression expression words having the highest frequency of appearance, the sum of the squares of the coordinate values of the corresponding impression expression words on all coordinate axes is compared. The impression data of the impression expression word is used as the final integration result, and in the case of generalized integration, the impression data that minimizes the sum of the squares of the coordinate values is used as the final integration result.

If there are a plurality of impression expression words that take the maximum value (in the case of specialized integration) or the minimum value (in generalized integration) of the sum of squares of coordinate values, the impression expression word addition process integrates the sentence toward the end of the sentence. Impression data obtained as an integrated result of the processing is adopted, and in the image information adding process, impression data of an integrated result present at a position closest to the beginning of a sentence is adopted.

Next, a specific procedure of the secondary integration will be described. FIG. 14 shows a specific procedure of the secondary integration in the image information search processing and the image information addition processing. It should be noted that the correction by the degree adverb is a process performed on the impression data, and therefore, this specific example and FIG.
In the specific example of 5, the description is omitted. Specific example 1 shown in FIG.
So, as shown at the top, "Clean,
Refreshing taste, yet sharp, sharp beer. An elegant body and refreshing throat. The following describes a case in which a message information search process is performed using two sentences such as "" as a query, or an image information addition process is performed as text data. Further, in this specific example, since each sentence is defined as a unit, each sentence is divided into a sentence of "Clearly-beer." The next integration is repeated. Therefore, in the first specific example, at least two integrated results are obtained, and in the image information adding process, the above-described final selection process is performed on these integrated results.

From the example sentence composed of the above two sentences, FIG.
Impression expression word "Clean as"
"Exhilarating""sharp""sharp""elegant"
"Refreshing" is extracted. Each of these impression expression words is denoted by A to F in order and is displayed above the corresponding impression expression word. Reference numerals A to F in the description of FIG. 14 indicate corresponding impression expression words (the same applies to FIG. 15). Then, the integration processing 1 and the integration processing 2 are repeated in the direction toward the end of the sentence, respectively, with the impression expression words “clean” and “elegant” as the initial reference impression words (indicated as (reference) in FIG. 14) in the respective delimiters. Will be.

First, as the integration process 1, the primary integration (1) of the initial reference impression expression word “clear” and the next “exhilarating” is performed. When the primary integration is successful, the integration result of the primary integration (1) is used as a new reference impression expression word, and the primary integration (2) with the next “sharp” is performed. If the primary integration (2) is also successful, the primary integration (2) integration result is used as a new reference impression expression word, and the primary integration (3) with the next “sharp” is performed. This primary integration (3) is also assumed to be successful. Since there is a break (punctuation mark) before the next impression expression word, the integration process 1 ends.

When the integration of the primary integration (1) is not successful, the other impression expression word “refreshing” in the primary integration processing (1) is used instead of the current reference impression expression word “clear”. The first integration (2) is performed with the next impression expression word “Kiriyaru” as the new reference impression expression word. If the primary integration (2) is successful and the primary integration (3) between the integration result and “sharp” is also successful, two impression data (“clear”) Impression data itself and impression data generated by primary integration of “exhilarating”, “sharp”, and “sharp”. In addition, when the primary integration (1) is not established, the primary integration (2) is established, and the primary integration (3) is not established, the integration result of the integration processing 1 is “clear”.
Impression data itself, impression data generated by the primary integration (2) of “exhilarating” and “sharp”, and three impression data itself of “sharp” impression data are obtained.

Next, as the integration processing 2 for the sentence after the break, the first impression expression word “elegant” viewed from the break in the processing direction (end of sentence) is set as the initial reference impression expression word, and this and the next impression expression word Primary integration with the word "fresh" (4)
I do. Here, it is assumed that the primary integration (4) has failed. Here, if there is the next impression expression word, the first integration is performed using “fresh” as a new reference impression expression word.
In the specific example of FIG. 14, since there is no next impression expression word, the integration process 2 ends.

Note that the integration process 2 also ends when the primary integration (4) succeeds, and in this case, as a result of the integration process 2, one impression data (primary integration of “elegant” and “fresh”) is obtained. Is obtained.

As described above, the primary integration (1) shown in FIG.
When (3) succeeds and the primary integration (4) fails, one impression data (the integration result of the primary integration (3)) is obtained by the integration processing 1, and two impression data are obtained by the integration processing 2. (The impression data of “elegant” and the “fresh”
Impression data itself) is obtained. As a result, the obtained three pieces of impression data are obtained as impression data for image information search or are subjected to final selection processing in the image information addition processing.

FIG. 15 is a flow chart showing the processing in the impression expression word adding process.
The following shows the specific procedure of integration. In the specific example 2 shown in FIG. 15, as shown at the top, “the refined body and refreshing throat unique to the first squeezing method, and the unique aroma and gorgeousness of black beer can be enjoyed.
Black draft beer that is easy to drink. "Will be described from the search impression data generating means 3 (see FIG. 1) as a query or from the image information adding means 8 as text data. Note that FIG.
In the specific example, it is assumed that the processing is performed as a single sentence.

From the above example sentences, the impression expression words “elegant”, “fresh”, “unique to black beer”, “fragrance”, “gorgeousness”, and “easy” that are outlined in FIG. 15 are extracted. Then, when the impression expression word “unique to black beer” is not registered in the subjective evaluation expression dictionary 4, impression data of an unregistered impression expression word “unique to black beer” is generated. First, the impression expression word “fresh” adjacent to the unregistered impression expression word “unique to black beer” is selected as the initial reference impression expression word (indicated as (reference 1) in FIG. 15), and integrated in the head direction. Processing 1 is repeated. Further, the impression expression word “fragrance” adjacent to “black beer peculiar” is also selected as the initial reference impression expression word (indicated as (reference 2) in FIG. 15), and the integration process 2 is repeated toward the end of the sentence.

First, as the integration process 1, the primary integration (1) of the initial reference impression expression word “refreshing” and the next impression expression word “elegant” in the processing direction (sentence direction) is performed.
If this primary integration (1) is successful, the integration process 1 ends because there is no next impression expression word in the processing direction.

Next, as the integration processing 2, one of the initial reference impression expression word “fragrance” and the next impression expression word “gorgeousness”
Next integration (2) is performed. When the primary integration (2) is successful, the integration result of the primary integration (2) is set as a new reference impression expression word, and the primary integration (3) with the next impression expression word “easy to drink” is performed. . It is assumed that the primary integration (3) has failed, and the integration processing 2 is terminated.

Next, as the integration process 3, the impression data obtained by the integration process 1 (the impression data generated by the primary integration (1) of “elegant” and “fresh”) and the impression data obtained by the integration process 2 Impression data (1
The first integration (4) with the impression data generated in the second integration (2) is performed. When the primary integration (4) is successful, the impression data generated here is newly registered in the subjective evaluation expression dictionary 4 as impression data of an unregistered impression expression word “unique to black beer”.

The process is also terminated when the primary integration (1) is not successful. In this case, the impression data of the current reference impression expression word (in the case of the specific example of FIG. Is the impression data of the initial reference impression expression word "fresh" because the primary integration has never been successful.)
Is obtained. If the primary integration (2) is not successful, the impression data itself of the current reference impression expression word “fragrance” is obtained as the integration result of the integration process 2.

Also, in the integration process 3, the primary integration (4)
Is unsuccessful, the sum of the appearance frequencies of all impression expression words that have been successfully integrated in the integration process 1 and the primary integration targets that have been successfully integrated in the integration process 2 And the sum of the frequencies of appearance of all impression expression words. That is, in the case of the specific example of FIG. 15, if the appearance frequency of the target impression expression word “elegant” of the primary integration (1) is 3 and the appearance frequency of “fresh” is 10, the appearance in the integration processing 1 The sum of the frequencies is 3 + 10 = 13. On the other hand, assuming that the appearance frequency of the target impression expression words “fragrance” of the primary integration (2) and (3) is 3 and the appearance frequency of “gorgeousness” is 2, the sum of the appearance frequencies is 3 + 2 = 5. Become. Integration processing 1
Since the appearance frequency sum 13> the appearance frequency 5 in the integration process 2, the impression data obtained in the integration process 1 is newly registered in the subjective evaluation expression dictionary 4 as impression data of the unregistered impression expression word “unique to black beer”. Is done.

[0109]

DESCRIPTION OF THE PREFERRED EMBODIMENTS (1) A data search device for searching for a program to be viewed from a multi-channel TV program A program to be viewed is searched by a natural expression including a subjective expression and constraints such as performers and genres. It is a data search device as described above. Here, information such as the genre of the channel and the performers of the program and the impression data thereof are generated and registered in the subjective evaluation information dictionary 5 in advance. Based on the user's natural language input, performers and genres are narrowed down as constraints, and the impression of the desired program is extracted from a subjective expression and searched, and presented in the order in which it is considered to be close to the desired program. At this time, T is regarded as having high subjective similarity (similarity S is large) even if it is out of the constraint condition.
The V program (image information) may be presented as a selection candidate. Therefore, the calculation of the similarity S is performed on all the image information extracted in step 29, and even if the similarity S is equal to or greater than the treatment value, even if the image information does not satisfy the constraint conditions, the user is determined as a candidate. The control to be presented.

(2) A data search apparatus for personalizing and providing travel information on the Web from the history of Web sites of travel destinations that have been visited in the travel information providing service on the Web. Travel destinations
Based on sightseeing spot information, travel trends and
Estimate preferences. Using this result, information close to the user's preference is searched for and provided preferentially. With this mechanism, even at the same destination, the tourist spot information can be personalized and provided according to the user's preference.

Specifically, the travel destination /
Impression data such as sightseeing spots is registered in the subjective evaluation information dictionary 5. Then, a plurality of image information included in the Web information or the like referred to by the user is extracted, and a common part is extracted from each impression data of the extracted image information by the above-described specialized integration (or generalized integration) technique. Then, new impression data is generated, and the image information (travel destination or the like) registered in the subjective evaluation information dictionary 5 is searched using the new impression data as the directional value of the user. For example, image information of the category "American travel destination" is "Hawaii""LasVegas"
It is assumed that “New York” or the like is registered in the subjective evaluation information dictionary 5. Further, it is assumed that “the Statue of Liberty”, “Carnegie Hall”, “Madison Square Garden” and the like are registered in the subjective evaluation information dictionary 5 as sightseeing spots of “New York”. When the user selects "Carnegie Hall" of "New York" among these pieces of information, the impression data given to "Carnegie Hall" is used as the directivity value of the user. Then, using the impression data of "Carnegie Hall", sightseeing spots and the like near "New York" with an impression similar to "Carnegie Hall" are searched from "American travel destinations" in the subjective evaluation information dictionary 5. .

(3) Device for Retrieving Data Using Subjective Expressions in Online Shopping In online shopping, various products are registered in the subjective evaluation information dictionary 5 as image information. The user not only searches by product name or product genre, but also inputs a subjective request by specifying a natural language or a specific product, brand, or the like. When specific products, brands, and the like are input, a subjective request represented by these inputs is estimated. Products are searched using this information, and presented in the order of products considered to be closer to the user's tendency and preference. Specifically, as image information, a product strongly associated with an impression such as a fashion / interior product or a music CD is registered in the subjective evaluation information dictionary 5.
Also, for brand names, product names, and the like that themselves have a certain strong impression, impression data is manually or automatically given in advance. When searching, do n’t just specify the impression you ’re looking for directly with impression words,
By specifying a brand name or a product name, the search is performed not only as a constraint but also as a target.

For example, when the user inputs "Bright suit of Armani" as a query, a search is first made from the suit information using impression data given to "Bright". Next, using the brand category “Armani” as a constraint, the search results are narrowed down to only image information in which this category is registered, and are shown as candidates. Finally, a search is performed using both the impression data assigned to “bright” and the impression data assigned to “Armani”, and the result is shown as a secondary candidate.

(4) A data search device for searching for a real estate property based on constraints such as price and subjective evaluation expressions such as appearance and habitability. Designate subjective requirements for preferred appearance, interior, comfort, etc. Using the information provided as real estate information, such as price and floor plan, as constraints, and presenting information in the order of information that is considered closer to the user's preference using impressions estimated from exterior and interior photographs or impressions actually surveyed I do.

Although the configuration and the processing operation of the present embodiment have been described above, the present invention is not limited to these embodiments, and various modifications are made within the scope of the invention described in the claims. It is possible. For example, at least one of the following two methods for searching for image information may be made searchable instead of or in addition to the search method described above.

The first method is a method of searching for image information using already registered image information as a query. For example, the subjective evaluation information dictionary 5 shown in FIG.
Searches for similar beers using the image information “Asahi Super Premium” registered in the database as a query. By designating “Asahi Super Premium”, the user reads out the impression data of “Asahi Super Premium” from the subjective evaluation information dictionary 5 and adds range data (for example, 2.0) to the impression data. Create search impression data consisting of a set of value / range data.
By searching for image information using the search impression data, "Asahi Super Dry" or the like is searched.

The second method is a method of retrieving passenger car information using objective characteristic condition data. For example, a search is performed using unregistered image information having no impression data. Specifically, when the image information has objective characteristic condition data such as displacement, fuel consumption, and weight, impression data is generated by the method described in step 71 of the image information adding process. The user adds range data (for example, 2.0) to the impression data to create search impression data composed of a set of level value / range data, and searches for image information using the search impression data. .

Further, in the image information search processing described with reference to the flowchart of FIG. 9, the case where step 31 and / or step 33 is performed after step 29 has been described. Step 29 may be performed later.

[0119]

Thus, according to the present invention, it is possible to obtain a search result sufficiently reflecting subjective similarity from a conditional sentence that is subjectively expressed by a user in a natural language.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a data search device according to an embodiment of the present invention.

FIG. 2 conceptually shows the contents of a subjective evaluation expression dictionary in the data search device.

FIG. 3 is an explanatory diagram exemplifying a coordinate space in the data search device using a subjective similarity element as a coordinate axis, and a conceptual state in which impression data for each impression expression word is laid out in the coordinate space;

FIG. 4 is an explanatory diagram showing two-dimensional coordinates of an impression data space of impression expression words in the data search device.

FIG. 5 conceptually shows the contents of a subjective evaluation information dictionary in the data search device.

FIG. 6 is an explanatory diagram exemplifying a coordinate space in the data search device using a subjective similarity element as a coordinate axis and a conceptual state in which impression data for each image information is laid out in the coordinate space;

FIG. 7 is a specific system configuration diagram of the data search device.

FIG. 8 is a system configuration diagram when the data search device is configured by a network.

FIG. 9 is a flowchart showing an image information search processing operation by the data search device.

FIG. 10 is a flowchart showing an image information addition processing operation by the data search device.

FIG. 11 is a flowchart showing an impression expression word adding process operation by the data search device.

FIG. 12 is a flowchart showing a processing operation of a primary specialization integration process in the data search device.

FIG. 13 is a flowchart showing a processing operation of a primary generalized integration process in the data search device.

FIG. 14 is an explanatory diagram showing a specific procedure of secondary integration in an image information search process and an image information addition process by the data search device.

FIG. 15 is an explanatory diagram showing a specific procedure of secondary integration in the impression expression word adding process by the data search device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Input means 2 Search condition sentence acquisition means 3 Search impression data generation means 4 Subjective evaluation expression dictionary 5 Subjective evaluation information dictionary 6 Image information search means 7 Output means 8 Image information addition means 9 Impression expression word addition means

Claims (8)

[Claims] 1. A subjective evaluation information dictionary in which impression data for image information to be searched is laid out in a coordinate space having a subjective similarity element as a coordinate axis, and a coordinate space having a subjective similarity element as a coordinate axis. A subjective evaluation expression dictionary in which impression data for impression expression words serving as search keys are laid out; a search condition sentence acquiring means for acquiring a search condition sentence subjectively expressed in a natural language as a search condition; Search impression data for extracting an impression expression word from the search condition sentence acquired by the acquisition means and generating impression data for the acquired search condition sentence from impression data of the impression expression word stored in the subjective evaluation expression dictionary And generating image information of impression data most similar to the impression data generated by the search impression data generating means from the subjective evaluation information dictionary. Data retrieval apparatus for an image information retrieval unit, and output means for outputting the image information retrieved by the image information retrieval unit, characterized by comprising the. 2. The data search device according to claim 1, wherein the impression data for each impression expression word stored in the subjective evaluation expression dictionary includes range data for specifying a range for each coordinate axis. 3. A text data obtaining means for obtaining text data relating to new image information, and an impression expression word is extracted from the text data obtained by the text data obtaining means, and said impression expression word is stored in the subjective evaluation expression dictionary. An impression data generating means for generating impression data for the new image information from impression data of an impression expression word; and the subjective evaluation information dictionary including the impression data generated by the impression data generating means as impression data of the new image information. 3. The data search device according to claim 1, further comprising: an image information adding unit to be registered. 4. Extraction means for extracting text data impression expression words, and the impression expression words extracted by the extraction means include impression expression words registered in the subjective evaluation expression dictionary and unregistered impression expression words. An impression expression word impression data generating means for generating impression data for the unregistered impression expression word from impression data for the registered impression expression words; and 4. The data retrieval apparatus according to claim 1, further comprising: an impression expression word adding unit that registers the impression expression words of unregistered impression expression words in the subjective evaluation expression dictionary. apparatus. 5. The subjective evaluation information dictionary includes objective characteristic conditions for each image information, and the image information search means, when the objective characteristic conditions are obtained from the search condition obtaining means, The data search apparatus according to any one of claims 1 to 4, wherein the image information is narrowed down using an objective feature condition as a constraint condition. 6. The subjective evaluation information dictionary has objective characteristic conditions together with impression data for each image information, and the image information stored in the subjective evaluation information dictionary based on the objective characteristic conditions of new image information. And a second image information adding unit that generates impression data for the new image information from a relationship between the impression data and the objective characteristic information and stores the impression data in the subjective expression dictionary. The data search device according to any one of claims 1 to 5. 7. The search condition obtaining unit obtains the search condition from an external device connected to a network, and the output unit outputs image information to the external device connected to the network. The data search device according to any one of claims 1 to 6, wherein the data search device is a data search device. 8. A function for creating a subjective evaluation information dictionary in which impression data for image information to be searched is laid out in a coordinate space having a subjective similarity element as a coordinate axis, and using a subjective similarity element as a coordinate axis. A function to create a subjective evaluation expression dictionary in which impression data for impression expression words serving as search keys are laid out in a coordinate space, and a search condition sentence that obtains a search condition sentence expressed subjectively in natural language as a search condition An acquisition function, an impression expression word is extracted from the search condition sentence acquired by the search condition sentence acquisition function, and an impression data for the acquired search condition sentence is extracted from the impression data of the impression expression word stored in the subjective evaluation expression dictionary. A search impression data generation function for generating impression data; and image information of impression data most similar to the impression data generated by the search impression data generation means. Wherein an image information retrieving function of retrieving from the subjective evaluation information dictionary, recording medium on which data search program is recorded for implementing an output function, to a computer for outputting the image information retrieved by the image information retrieval unit.