JP2005107718A - Database retrieval device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a database retrieval device realizing retrieval which is much more suitable for an individual. <P>SOLUTION: The device is provided with a first means for inputting user information for specifying a classification pattern to a user in addition to an impression value at every sensitivity word pair at the time of retrieval, a second means calculating a sensitivity spatial coordinate value corresponding to the impression value at every sensitivity word pair, which are inputted by the first means, a third means calculating a spatial distance between the sensitivity spatial coordinate value calculated by the second means and the sensitivity spatial coordinate value registered in a database as an index, and a fourth means selecting a content whose spatial distance calculated by the third means is the shortest. The third means uses only index information corresponding to the classification pattern specified by user information inputted by the first means as an object of calculation of the spatial distance. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a database search apparatus for music, video and the like.

  A system for searching music by sensitivity words has been proposed ("Music database search system by sensitivity words using music sensitivity space", Information Processing Society of Japan Vol.42 No12 2001). In this conventional search system, a search space for mapping songs in a database is generated by the SD method and factor analysis. In addition, when a new song is registered in the database, mapping to a search space is performed by an automatic indexing system constituted by a GA and a neural network. At the time of search, by inputting the degree (1-7) of eight sensitivity word pairs to the neural network, the neural network outputs the coordinates in the sensitivity space corresponding to those inputs. The search candidates are presented to the user in order from the song with the closest Euclidean distance from the output coordinate value.

  The sensitivity word pairs are “bright-dark”, “heavy-light”, “hard-soft”, “stable-unstable”, “clear-cloudy”, “smooth-crisp”, “ Eight types are used: severe-mild and thick-thin.

By the way, although it is thought that the user's sensitivity with respect to a music and an image also changes with sex and age groups, this point is not considered in the said conventional search system.
JP 2001-306580 A "Music database retrieval system by sensitivity word using music sensitivity space" IPSJ Journal Vol.42 No12

  An object of the present invention is to provide a database search device that enables a search that is more personalized.

  According to the first aspect of the present invention, for each content registered in the database, the Kansei space coordinate value obtained for the content is registered in the database as an index of the content, and is input by the user during the search. Based on the impression value for each sensitivity word pair, the sensitivity space coordinate value corresponding to the input impression value for each sensitivity word pair is calculated, and the spatial distance from the calculated sensitivity space coordinate value is the shortest In a database search device for searching for content, the sensitivity space coordinate value of each content is obtained for each classification pattern classified by gender and / or age, and is registered in the database as index information. In order to make the user input user information for specifying the classification pattern in addition to the impression value for each sensitivity word pair The first means, the second means for calculating the sensitivity space coordinate value corresponding to the impression value for each sensitivity word pair input by the first means, and the sensitivity space coordinate value calculated by the second means and registered in the database as an index A third means for calculating a spatial distance between the sentimental spatial coordinate values, and a fourth means for selecting the content having the shortest spatial distance calculated by the third means. The third means comprises: When calculating the spatial distance, only index information corresponding to the classification pattern specified by the user information input by the first means is used as a spatial distance calculation target.

  According to the second aspect of the present invention, for each content registered in the database, the Kansei space coordinate value obtained for the content is registered in the database as an index of the content, and is specified by the user at the time of search. In a database search device that searches for content having the shortest spatial distance from the emotional space coordinate value, the emotional space coordinate value of each content is obtained for each classification pattern classified by sex and / or age, and index information is obtained. And is input by the first means and the first means for allowing the user to input user information for specifying the classification pattern in addition to the sensitivity space coordinate value at the time of the search. Kansei space coordinate values and Kansei space coordinate values registered as indexes in the database A second means for calculating the spatial distance between the second means, and a third means for selecting the content having the shortest spatial distance calculated by the second means. When the second means calculates the spatial distance, Only the index information corresponding to the classification pattern specified by the user information input by the first means is used as a spatial distance calculation target.

  According to the present invention, it is possible to search more suited to an individual.

  DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention applied to a music database search system will be described below with reference to the drawings.

[1] Explanation of the configuration of the music database search system

  FIG. 1 shows the configuration of a music database search system.

  The music database search system includes a music sensitivity space generation unit 100 and a search system unit 200. The music sensitivity space generation unit 100 generates a music sensitivity space as a search space. The search system unit 200 searches the music database for a song that matches the input song impression, and outputs the song.

[2] Description of the music sensitivity space generation unit 100

  As shown in FIG. 1, the music sensitivity space generation unit 100 automatically calculates an initial music sensitivity space generation unit 101 that generates a music sensitivity space by a listening experiment, and music sensitivity space coordinate values of a song to be added to the database. And an automatic indexing unit 102.

[2-1] Description of the initial music sensitivity space generation unit 101

  The initial music sensation space generation unit 101 includes a factor analysis unit 111 that performs factor analysis using a song impression value obtained by a listening experiment, and a database unit 112.

  A listening experiment will be described. As subjects, 100 pops were presented to 10 males under 30 years old, males over 30 years old, females under 30 years old, females over 30 years old, each with 7 levels for each sensitivity word pair. We had you evaluate.

The following four types of sensitivity word pairs were used.
(1) bright-dark (2) heavy-light (3) stable-unstable (4) powerful-weak

  From this experimental result, for each song, for each sex word pair, for each group of 4 groups by sex and age (male under 30 years old, male over 30 years old, female under 30 years old, female over 30 years old). The song impression value was calculated. The music impression value for each sensitivity word pair is represented by an average value of the evaluation values of the sensitivity word pairs.

  The factor analysis unit 111 performs, for each group, a factor analysis in which the song impression value for each sensitivity word pair of each song is multivariate, and for each song, the psychological impression of the song impression with each factor as an axis. Coordinate data of a space (musical sensation space) representing a strong correlation is obtained. Thus, the music sensitivity space is a space having a plurality of factor axes. In this example, it is assumed that there are three types of factors and three types of factor axes.

  In the database unit 112, music data of each music is registered, and coordinate data in the sensitivity space of each music is registered as index information of each music for each group.

[2-2] Description of the automatic indexing unit 102

  The automatic indexing unit 102 analyzes a newly registered song and extracts physical feature amounts such as tempo and rhythm, and a physical feature amount extracted by the physical feature amount extraction unit 121. Is converted into a song impression value for each sensitivity word pair, and a music impression value for each sensitivity word pair obtained by the physical feature value / music impression value conversion unit 122. Is converted into a music sensitivity space coordinate value, and a music impression space / music sensitivity space coordinate conversion unit 123 is provided.

  The music sensitivity space coordinate value of the newly registered song obtained by the automatic indexing unit 102 is registered in the database unit 112 together with the song data of the newly registered song.

  FIG. 2 shows the configuration of the automatic indexing unit 102 more specifically.

  The physical feature quantity / musical impression value conversion unit 122 includes regression analysis units 122a to 122d corresponding to four types of sensitivity word pairs. The music impression value / musical sensitivity space coordinate conversion unit 123 includes a neural network.

  A plurality of physical feature quantities related to the corresponding sensitivity word pairs are input to the regression analysis units 122 a to 122 d in the physical feature quantity / musical impression value conversion unit 122. The types of physical feature values input to the regression analysis units 122a to 122d are determined in advance.

  Each regression analysis part 122a-122d converts a physical feature-value into the music impression value of the corresponding sensitivity word pair based on following Formula (1).

  In the above equation (1), OUT is an output value of the regression analysis unit, and represents an estimated value of the music impression value related to the sensitivity word pair corresponding to the regression analysis unit. In [k] represents the k-th physical feature amount input to the regression analysis unit. W [k] is a weighting coefficient for In [k]. W [0] is a weighting coefficient.

  The weighting factors W [k] and W [0] in the above formula (1) are based on the results of the above-mentioned listening experiment and the physical features of each song used for each group of 4 groups by sex and age. Is required in advance. The weighting factors W [k] and W [0] for one group are determined as follows. A listening experiment result for the group, that is, a song impression value for each sensitivity word pair obtained for each song is used as a teacher signal (output value OUT). Further, the physical feature amount extracted by the physical feature amount extraction unit 121 for each song is set as the input value In. Then, the weighting coefficients W [k] and W [0] are calculated so that the regression analysis results for all the songs become values closest to the teacher signal for the songs.

  Learning of the neural network constituting the music impression value / musical sensitivity space coordinate conversion unit 123 will be described. In the neural network, learning is performed for every four groups according to sex and age, and the internal state is determined for every four groups according to sex and age. The input signal at the time of learning is the result of the above-described listening experiment, that is, the music impression value for each sensitivity word pair obtained for each music. The teacher signal at the time of learning is a music sensitivity space coordinate value (coordinate value of each factor axis) of each song obtained by the factor analysis unit 111.

[3] Description of the search system unit 200

  As shown in FIG. 1, the search system unit 200 includes a user input unit 201, a song impression value / musical sensitivity space coordinate conversion unit 202, a search unit 203, and a search result output unit 204.

  The user inputs user data such as gender and age, song impression values for each sensitivity word pair (seven levels), importance level information of each sensitivity word pair, and the like on the user input unit 201. User data such as sex and age input to the user input unit 201 and the song impression value for each sensitivity word pair regarding the song to be selected are given to the song impression value / music sensitivity space coordinate conversion unit 202. As the music impression value / musical sensitivity space coordinate conversion unit 202, a neural network constituting the music impression value / musical sensitivity space coordinate conversion unit 123 in the automatic indexing unit 102 is used as it is.

  The music impression value / musical sensitivity space coordinate conversion unit 202 sets the internal state of the neural network to the internal state of the group corresponding to the user data given from the user input unit 201, and then each of the given values given from the user input unit 201. The music impression space coordinate value corresponding to the sensitivity word pair is acquired by inputting the music impression value for each sensitivity word pair to the neural network. The music impression value / music sensitivity space coordinate conversion unit 202 gives the acquired music sensitivity space coordinate value to the search unit 203.

  The search unit 203 is also given importance level information of each sensitivity word pair input to the user input unit 201. First, the processing of the search unit 203 when importance level information of each sensitivity word pair is not given will be described.

  The music sensitivity space coordinate values corresponding to the group specified by the user data of the music registered in the database unit 112 are defined as (X1 [i], X2 [i], X3 [i]). i is a song number. Also, values obtained by converting the music impression value for each sensitivity word pair input by the user into music sensitivity space coordinates by the music impression value / music sensitivity space coordinate conversion unit 202 are (a1, a2, a3). ).

  The search unit 203 calculates a spatial distance L for each registered song based on the following formula (2), and selects a song having the smallest spatial distance L. The search result output unit 204 reads the song data of the song selected by the search unit 203 from the database unit 112 and outputs it.

L = (X1 [i] -a1) ² + (X2 [i] -a2) ² + (X3 [i] -a3) ² (2)

  Next, processing of the search unit 203 when importance level information of each sensitivity word pair is given will be described.

The user can input importance information for each sensitivity word pair in consideration of the importance for the four types of sensitivity word pairs (bright-dark, heavy-light, stable-unstable, strong-weak). Is possible. The importance level information of each sensitivity word pair (bright-dark, heavy-light, stable-unstable, strong-weak) is represented by {p1, p2, p3, p4}. p _n (n = 1, 2, 3, 4) is a value in the range of 1 ≦ p _n ≦ 4, with 1 being the most important, and the greater the numerical value, the lower the importance. When all sensitivity word pairs are handled equally, importance level information for all sensitivity word pairs is set to 1.

  Table 1 shows the contribution rate (factor load) of each sensitivity word pair to each factor (factor 1, factor 2, factor 3).

  When the contribution rate of each sensitivity word pair to each factor (factor 1, factor 2, factor 3) is as shown in Table 1, the weights W1, W2, and weights of each factor (factor 1, factor 2, factor 3) W3 is calculated based on the following equation (3). The weights W1, W2, and W3 of each factor (factor 1, factor 2, factor 3) are calculated in advance and held in the search unit 203.

W1 = (0.14 +0.17 +0.96 +0.34)
÷ (0.14 x p1 + 0.17 x p2 + 0.96 x p3 + 0.34 x p4)
W2 = (0.69 +0.01 +0.12 +0.91)
÷ (0.69 x p1 + 0.01 x p2 + 0.12 x p3 + 0.91 x p4)
W3 = (0.64 + 0.96 + 0.00 + 0.01)
÷ (0.64 x p1 + 0.96 x p2 + 0.00 x p3 + 0.01 x p4) (3)

  The weights W1, W2, and W3 of each factor (factor 1, factor 2, factor 3) may be calculated based on the following equation (4).

W1 = 0.14 x (1 / p1) + 0.17 x (1 / p2) + 0.96 x (1 / p3) + 0.34 x (1 / p4)
W2 = 0.69 x (1 / p1) + 0.01 x (1 / p2) + 0.12 x (1 / p3) + 0.91 x (1 / p4)
W3 = 0.64 x (1 / p1) + 0.96 x (1 / p2) + 0.00 x (1 / p3) + 0.01 x (1 / p4)… (4)

  The music sensitivity space coordinate values corresponding to the group specified by the user data of the music registered in the database unit 112 are defined as (X1 [i], X2 [i], X3 [i]). i is a song number. Also, values obtained by converting the music impression value for each sensitivity word pair input by the user into music sensitivity space coordinates by the music impression value / music sensitivity space coordinate conversion unit 202 are (a1, a2, a3). ).

  The search unit 203 calculates a spatial distance L for each registered song based on the following equation (5), and selects a song having the smallest spatial distance L. The search result output unit 204 reads the song data of the song selected by the search unit 203 from the database unit 112 and outputs it.

L = W1 × (X1 [i] −a1) ² + W2 × (X2 [i] −a2) ² + W3 × (X3 [i] −a3) ² (5)

  The search unit 203 may calculate the spatial distance L for each registered song based on the following equation (6), and may select a song having the smallest spatial distance L.

L = W1 × α1 × (X1 [i] −a1) ² + W2 × α2 × (X2 [i] −a2) ²
+ W3 × α3 × (X3 [i] −a3) ² (6)

 In the above equation (6), weights α1, α2, and α3 are added to each factor in addition to the weights W1, W2, and W3. αn (where n = 1, 2, 3) is k1 or k2 depending on the sign of Xn [i] (where n = 1, 2, 3) and an (where n = 1, 2, 3). I take the. k1 ≧ k2> 0.

  That is, when Xn [i] and an have different signs (when Xn [i] and an are on opposite sides of the origin of the factor axis), the distance between them is increased. Therefore, αn = k1. When Xn [i] and an have the same sign, αn = k2.

  In the above embodiment, music sensitivity space coordinates of each registered song are obtained for each of a plurality of groups divided by sex and age, but for each of a plurality of patterns divided by sex, age, and music genre. In addition, the music sensitivity space coordinates of each registered song may be obtained.

  In the above-described embodiment, the music impression value for each sensitivity word pair is input to the user input unit 201, but the sensitivity space coordinate values (coordinate values for each factor) may be directly input. In this case, importance information for each factor may be directly input as importance information.

  The present invention can be used for music playback devices such as car stereos, home audio, and PCs. The present invention can also be used for search software such as a music distribution service.

  The present invention can also be used for searching video databases as well as music databases. That is, it can be used as a search device for databases such as pictures and catalogs.

It is a block diagram which shows the structure of a music database search system. 2 is a block diagram showing more specifically the configuration of an automatic indexing unit 102. FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Music sensitivity space generation part 200 Search system part 101 Initial music sensitivity space generation part 102 Automatic indexing part 111 Factor analysis part 112 Database part 121 Physical feature-value extraction part 122 Physical feature-value / music impression value conversion part 123 Music impression value / Music sensitivity space coordinate conversion unit 201 User input unit 202 Music impression value / Music sensitivity space coordinate conversion unit 203 Search unit 204 Search result output unit

Claims (2)

For each content registered in the database, the Kansei space coordinate value obtained for the content is registered in the database as an index of the content, and the impression value for each sensitivity word pair input by the user at the time of search In the database search device for searching the content having the shortest spatial distance from the calculated emotional space coordinate value, the emotional space coordinate value corresponding to the input impression value for each sensitivity word pair is calculated based on
For each classification pattern classified by gender and / or age, the Kansei space coordinate value of each content is obtained and registered in the database as index information.
A first means for allowing a user to input user information for specifying a classification pattern in addition to an impression value for each sensitivity word pair at the time of search;
Second means for calculating a sensitivity space coordinate value corresponding to an impression value for each sensitivity word pair input by the first means;
A third means for calculating a spatial distance between the emotional space coordinate value calculated by the second means and the emotional space coordinate value registered as an index in the database; and the spatial distance calculated by the third means is the shortest. A fourth means of selecting content,
The third means uses only the index information corresponding to the classification pattern specified by the user information input by the first means as the spatial distance calculation target when calculating the spatial distance. apparatus. For each content registered in the database, the Kansei space coordinate value obtained for the content is registered in the database as an index of the content, and the spatial space with the Kansei space coordinate value specified by the user at the time of search In a database search device that searches for content with the shortest distance,
For each classification pattern classified by gender and / or age, the Kansei space coordinate value of each content is obtained and registered in the database as index information.
A first means for allowing a user to input user information for specifying a classification pattern in addition to the sensitivity space coordinate value at the time of retrieval;
A second means for calculating a spatial distance between the emotional space coordinate value input by the first means and the emotional space coordinate value registered as an index in the database; and the spatial distance calculated by the second means is the shortest. A third means for selecting content,
The second means uses the index information corresponding to the classification pattern specified by the user information input by the first means as the spatial distance calculation target when calculating the spatial distance. apparatus.

Images