JP5499361B2 - Semi-teacher topic model learning apparatus, semi-teacher topic model learning method and program - Google Patents

Description

  The present invention relates to a semi-teacher topic model learning device, a semi-teacher topic model learning method, and a program.

  In recent years, an image retrieval technique for retrieving a desired image based on given language information, and an image recognition technique for automatically assigning language information that describes the image to a given image are digital cameras, With the spread of imaging devices such as mobile phones and the generalization of image sharing on the Internet, it has become a very important technology. In addition, many image recognition search technologies that realize image search and image recognition under the same framework have been developed (see, for example, Non-Patent Document 1).

Nakayama, Harada, Kuniyoshi, Otsu “Ultra-high-speed image recognition / retrieval method using stochastic structure learning corresponding to image / word concept”, IEICE Technical Report, PRMU2007-147, December 2007

  However, in the technique described in Non-Patent Document 1, it is necessary to learn the relationship between an image and language information before actually performing search or recognition. Many pairs of images and language information are required. However, manually adding linguistic information to a large number of images involves a lot of labor, and it is difficult to summon a large number of images associated with linguistic information. If the image associated with the language information cannot be prepared sufficiently, the relationship between the image and the language information cannot be learned with high accuracy, and as a result, the reliability of image recognition and image search is impaired. There's a problem.

  The present invention has been made in view of the above points, and an object of the present invention is to provide an image (signal) and language information (in the case where only a small amount of an image (signal) associated with language information (additional information) is available. The object of the present invention is to provide a semi-teacher topic model learning device, a semi-teacher topic model learning method, and a program capable of accurately learning the relationship of (additional information).

The present invention has been made to solve the above problems, and one aspect of the present invention is a semi-teacher topic for learning the relationship between a given signal and additional information that is information describing the signal. A model learning apparatus, comprising: a completely accumulated signal that is an element of a completely accumulated signal set that is a set of signals to which additional information is given in advance; and an incompletely accumulated signal set that is a set of signals to which no additional information is given. A storage signal feature extraction unit that extracts a stored signal feature that is a vector representing the characteristics of each signal from each of the incomplete storage signals that are elements, and a storage addition that is a set of additional information added to the complete storage signal A storage additional information feature extraction unit that extracts a storage additional information feature that is a vector expressing characteristics of the storage additional information from each of the storage additional information that is an element of the information set; and the complete storage signal A complete stored signal feature set that is a set of stored signal features; an incomplete stored signal feature set that is a set of stored signal features of the incomplete stored signal; and a stored additional information feature set that is a set of stored additional information features; A storage latent variable extraction unit that extracts a storage latent variable that is a variable for describing the relationship between a signal and additional information, the complete storage signal feature set, the incomplete storage signal feature set, and the storage addition and information feature set, from a set a is accumulated latent variable set of the storage latent variables, and a topic model learning unit for extracting topic model is a model that describes the relationship between the signal and the additional information, the storage potential The variable extraction unit is a multidimensional vector space in which accumulated latent variables exist from the complete accumulated signal feature set, the incomplete accumulated signal feature set, and the accumulated additional information feature set. A latent variable space generating unit that generates a latent variable space, the complete accumulated signal feature set, the accumulated additional information feature set, and the latent variable space are used to store an accumulated latent variable set corresponding to the complete accumulated signal feature set. Corresponding to an incompletely stored signal feature set using a fully stored latent variable set extraction unit that extracts a set of completely stored latent variables, the fully stored signal feature set, the incompletely stored signal feature set, and the latent variable space A storage latent variable set extraction unit that extracts an incomplete storage latent variable set that is a storage latent variable set to be formed and combines the incomplete storage latent variable set and the complete storage latent variable set to form an entire storage latent variable set; is a semi-supervised topic model learning device which is characterized in that said Rukoto equipped with.

  Further, according to one aspect of the present invention, in the above semi-teacher topic model learning apparatus, the topic model learning unit is a latent variable model that is a model that describes a stochastic structure of the accumulated latent variable itself from the accumulated latent variable set. A latent variable model learning unit that extracts signals, and a signal / latency model that describes a relationship between a signal and a latent variable from the complete accumulated signal feature set, the incomplete accumulated signal feature set, and the accumulated latent variable set. Additional information / latency is a model that describes the relationship between additional information and latent variables from the signal / latent variable relationship model learning unit for extracting a variable relationship model, the accumulated additional information feature set, and the accumulated latent variable set. An additional information / latent variable relationship model learning unit for extracting a variable relationship model, the latent variable model, the signal / latent variable relationship model, and the additional information. - to integrate the latent variable relationship model is characterized in that the topic model.

  Further, according to one aspect of the present invention, in the above semi-teacher topic model learning device, the topic model learning unit includes the additional information / latent variable relationship model, the accumulated additional information feature set, and the accumulated latent variable set. An additional information / latent variable relationship model update unit that updates the additional information / latent variable relationship model in consideration of the distribution of latent variables corresponding to the incompletely stored signal feature set, and the additional information / latent variable relationship model is updated. The additional information / latent variable relationship model for correcting the additional information / latent variable relationship model so that the information of the stored additional information feature set is retained to a certain extent from the additional information / latent variable relationship model and the stored additional information feature set. And a correction unit.

  Further, according to one aspect of the present invention, in the above semi-teacher topic model learning device, the latent variable model learning unit extracts a latent variable model by superimposing a probability density function using each accumulated latent variable as a parameter, The signal / latent variable relation model learning unit extracts a signal / latent variable relation model individually for each accumulated latent variable and based on a probability density function using the accumulated signal feature as a parameter, and the additional information / latent model is extracted. The variable relationship model learning unit is characterized by extracting an additional information / latent variable relationship model individually for each accumulated latent variable and based on a probability density function having the additional information feature as a parameter.

Another aspect of the present invention is a semi-teacher topic model learning method used in a semi-teacher topic model learning apparatus that learns the relationship between a given signal and additional information that is information describing the signal. The accumulated signal feature extraction unit is a complete accumulated signal that is an element of a complete accumulated signal set that is a set of signals to which additional information is given in advance, and an incomplete accumulated signal set that is a set of signals to which no additional information is given A step of extracting a stored signal feature that is a vector expressing the characteristics of each signal from each of the incompletely stored signals that are elements of the stored information, and a stored additional information feature extracting unit A step of extracting a stored additional information feature, which is a vector expressing the characteristics of the stored additional information, from each of the stored additional information that is an element of the stored additional information set that is a set. And an accumulated latent variable extraction unit includes a complete accumulated signal feature set that is a set of accumulated signal features of the complete accumulated signal, an incomplete accumulated signal feature set that is a set of accumulated signal features of the incomplete accumulated signal, and A step of extracting a storage latent variable that is a variable for describing a relationship between a signal and additional information from a stored additional information feature set that is a set of stored additional information features, and a topic model learning unit includes the complete stored signal It is a model that describes the relationship between a signal and additional information from a feature set, the incompletely stored signal feature set, the stored additional information feature set, and a stored latent variable set that is a set of the stored latent variables. see containing and extracting the topic model, extracting the accumulated latent variables, latent variable space generating section, the full accumulation signal feature set and the incomplete stored signal characteristics collection Generating a latent variable space that is a multi-dimensional vector space in which accumulated latent variables exist, and a fully accumulated latent variable set extracting unit comprising the complete accumulated signal feature set and the accumulated addition Using the information feature set and the latent variable space to extract a fully stored latent variable set that is a stored latent variable set corresponding to the fully stored signal feature set; and a stored latent variable set extracting unit includes the complete storage Using the signal feature set, the incompletely stored signal feature set, and the latent variable space, an incompletely stored latent variable set that is a stored latent variable set corresponding to the incompletely stored signal feature set is extracted. semi-supervised topic model learning, wherein the free Mukoto and forming the entire storage latent variable set together with the full storage latent variable set and latent variable set Is the method.

Further, according to one aspect of the present invention, in the semi-teacher topic model learning method described above, the step of extracting the topic model includes: a latent variable model learning unit including a stochastic structure of the accumulated latent variable itself from the accumulated latent variable set; Extracting a latent variable model, which is a model describing a signal, and a signal / latent variable relation model learning unit, from the complete accumulated signal feature set, the incompletely accumulated signal feature set, and the accumulated latent variable set, A step of extracting a signal / latent variable relationship model, which is a model describing a relationship with a latent variable, and an additional information / latent variable relationship model learning unit, from the accumulated additional information feature set and the accumulated latent variable set, Extracting the additional information / latent variable relationship model, which is a model describing the relationship between the additional information and the latent variable, the latent variable Characterized by a topic model integrates Dell and said signal-latent variables relationship model and the additional information-latent variables relations model.

In the step of extracting the topic model, the additional information / latent variable relationship model update unit includes the incompletely stored signal feature from the additional information / latent variable relationship model, the stored additional information feature set, and the stored latent variable set. A step of updating the additional information / latent variable relationship model in consideration of a distribution of latent variables corresponding to the set, and an additional information / latent variable relationship model correcting unit updating the additional information / latent variable relationship model; The method further includes the step of correcting the additional information / latent variable relationship model so as to hold the information of the stored additional information feature set to a certain extent from the latent variable relationship model and the stored additional information feature set. .

The step of extracting the latent variable model extracts a latent variable model by superimposing a probability density function using each accumulated latent variable as a parameter, and the step of extracting the signal / latent variable relationship model includes adding each latent latent variable to each accumulated latent variable. In contrast, the step of extracting the signal / latent variable relationship model individually and based on the probability density function having the stored signal feature as a parameter and extracting the additional information / latent variable relationship model is performed for each stored latent variable. The additional information / latent variable relationship model is extracted individually based on a probability density function having the additional information feature as a parameter.

One embodiment of the present invention is a computer as a semi-teacher topic model learning device that learns the relationship between a given signal and additional information that is information describing the signal. A complete accumulation signal that is an element of a complete accumulation signal set that is a set of signals to which additional information is given in advance, and an incomplete accumulation signal that is an element of an incomplete accumulation signal set that is a set of signals to which no additional information is given Extracting a stored signal feature that is a vector expressing the characteristics of each signal from each of the above, and a stored additional information feature extracting unit of a stored additional information set that is a set of additional information given to the complete stored signal A step of extracting a stored additional information feature that is a vector expressing characteristics of the stored additional information from each of the stored additional information that is an element, and a stored latent variable extracting unit, A complete accumulated signal feature set that is a set of accumulated signal features of all accumulated signals, an incomplete accumulated signal feature set that is a set of accumulated signal features of the incomplete accumulated signal, and an accumulation addition that is a set of accumulated additional information features. Extracting an accumulated latent variable that is a variable for describing the relationship between a signal and additional information from the information feature set; and a topic model learning unit comprising the complete accumulated signal feature set and the incompletely accumulated signal feature. Executing a step of extracting a topic model, which is a model describing a relationship between a signal and additional information, from a set, the stored additional information feature set, and a stored latent variable set that is a set of the stored latent variables. The step of extracting the accumulated latent variable includes a step in which a latent variable space generation unit includes the complete accumulated signal feature set, the incomplete accumulated signal feature set, and the accumulated additional information feature set. Generating a latent variable space that is a multidimensional vector space in which accumulated latent variables exist, and a complete accumulated latent variable set extracting unit includes the complete accumulated signal feature set, the accumulated additional information feature set, and the latent variable space. And extracting a fully stored latent variable set that is a stored latent variable set corresponding to the complete stored signal feature set, and a stored latent variable set extracting unit includes the complete stored signal feature set and the incomplete storage. Using the signal feature set and the latent variable space, an incomplete storage latent variable set that is a storage latent variable set corresponding to the incomplete storage signal feature set is extracted, and the incomplete storage latent variable set and the complete storage latent is a program characterized Rukoto to execute the steps of forming the entire storage latent variable set by combining the variable set.

  According to the present invention, the accumulated latent variable extraction unit describes the relationship between the signal and the additional information from both the signal associated with the additional information (language information) and the signal not associated with the additional information. By extracting the accumulated latent variable, the relationship between the signal and the additional information can be accurately learned even when only a small amount of the signal associated with the additional information can be used. That is, when learning the relationship between a signal and additional information, a signal that is not associated with additional information is used at the same time, which is higher than when only a small number of signals associated with additional information are used. The relationship between the signal and the additional information can be learned with accuracy.

It is a block diagram which shows the function structure of the semi-teacher topic model learning apparatus by this embodiment. It is a block diagram which shows the function structure of the accumulation | storage latent variable extraction part by this embodiment. It is a block diagram which shows the function structure of the topic model learning part by this embodiment. It is a flowchart which shows the procedure of the semi-teacher topic model learning process by this embodiment. It is a flowchart which shows the procedure of the accumulation latent variable extraction process in the accumulation latent variable extraction part by this embodiment. It is a flowchart which shows the procedure of the topic model learning process in the topic model learning part by this embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
First, the words used below are defined as follows.
The additional information is information (text label) describing the target signal, and generally refers to a signal of a different type from the target signal. The types of signals used here include, for example, still images, moving images (video), acoustic signals (speech, music, environmental sounds, etc.), text, biological signals (electroencephalograms, myoelectric signals, etc.), geographic information (GPS acquisition) Information, location name, etc.) and device information (camera model number, encoding codec, etc.). In the following, the expression “stored image” is used to explain the case where an image is used as a target signal and text is used as additional information. However, since it can be replaced with a signal as described above, It may be read as “accumulated signal”. A completely accumulated image (signal) is an image (signal) to which additional information is given in advance. An incompletely stored image (signal) is an image (signal) to which no additional information is given. Further, a completely accumulated image (signal) set G _c = {g ₁ , g ₂ ,..., G _N } is a set of N (N is an integer of 1 or more) sheets of completely accumulated images. The incompletely stored image (signal) set G _I = {g _{N + 1} , g _{N + 2} ,..., G _{N + Nx} } is a set of Nx (Nx is an integer of 1 or more) sheets of incompletely stored images. In many cases, N << Nx (N is sufficiently smaller than Nx). Also, full storage image set G _c and imperfect accumulated image set G _I and the combined accumulated image (signal) to the set G. The accumulated additional information set W1 = {w ₁ , w ₂ ,..., W _N } is a set of additional information given to the image (signal).

FIG. 1 is a block diagram showing a functional configuration of a semi-teacher topic model learning device 10 according to the present embodiment.
Semi-supervised topic model learning unit 10 is complete and the storage image set G _c incomplete accumulation image set G _I and accumulating additional information set W1 as input, describe the relationship between the image and the additional information model image ( Signal) / additional information relation model (topic model) is output (learned), for example, an apparatus constructed from a computer or the like. The semi-teacher topic model learning device 10 includes an accumulated signal feature extracting unit 1, an accumulated additional information feature extracting unit 2, an accumulated latent variable extracting unit 3, and a topic model learning unit 4.

Stored signal feature extraction unit 1, a complete and a stored image set G _c and imperfect accumulated image set G _I as an input, a full storage image set G _c elements in a complete stored image and an incomplete storage image set G _I elements From each of the incompletely stored images, stored image (signal) features, which are vectors representing the characteristics of each image, are extracted. The accumulated signal feature extraction unit 1 then performs a completely accumulated image (signal) feature set that is a set of completely accumulated image (signal) features and an incompletely accumulated image (signal) feature that is a set of incompletely accumulated image (signal) features. The set is output to the accumulated latent variable extraction unit 3. The completely stored image feature is a stored image feature of the completely stored image, and the incompletely stored image feature is a stored image feature of the incompletely stored image.

Specifically, if the accumulated image is a still image, the accumulated signal feature extraction unit 1 first performs SURF (Speed-up robust features; for example, H. Bay, A. Ess, T. Tuytelaars, and L. Van Gool, “Speeded-up robust features (SURF),” CVIU, vol. 110, no. 3, pp. 346-359, 2008.) is used to detect image feature points. Subsequently, the accumulated signal feature extraction unit 1 quantizes the local feature amount calculated for the feature point, and expresses each quantized feature amount as a histogram. As a result, the accumulated signal feature extraction unit 1 performs the complete accumulated image feature set X _c = {x ₁ , x ₂ ,..., X _N } and the incomplete accumulated image feature set X _I = {x _{N + 1} , x _{N + 2} ,. xN _{+ Nx} } is obtained. Elements x _n (n = 1, 2, 3,..., Nx) of the image feature set are histograms.

  If the stored image is a moving image, a histogram generated from a time-series signal as in the “time-series active search (TAS) method (Japanese Patent No. 30653314)” may be used as a feature amount. Further, when an acoustic signal is used instead of the stored image, MFCC (Mel-frequency cepstral coefficients), chroma feature amount, or the like can be used. Further, in the case of geographic information, information such as longitude, latitude, and time acquired by GPS can be used as a feature amount as it is. In the case of text, it will be described in the explanation of the accumulated load information feature extraction unit 2 described later.

  In the present embodiment, the accumulated signal feature extraction unit 1 obtains the accumulated image feature from the quantized histogram of the local feature amount, but instead, for example, a color histogram, digital of each small region in the image Low frequency component of cosine transform, low frequency component or high frequency component of Haar wavelet, high-order local autocorrelation feature (for example, N. Otsu and T. Kurita “A new scheme for practical flexible and intelligent vision systems,” Proc. IAPR Workshop on Computer Vision, pp.431-435, 1988) can also be used. In addition, the above-described SURF is used for feature point extraction and local feature amount calculation. Instead, for example, SIFT (for example, D. Lowe, “Distinctive image features from scale-invariant keypoints,” International Journal other methods such as of Computer Vision, Vol. 60, No. 2, pp. 91-110, 2004) can also be used.

  The accumulated additional information feature extraction unit 2 receives the accumulated additional information set W1 as input, and stores accumulated additional information features that are vectors representing the characteristics of the accumulated additional information from each of the accumulated additional information that is an element of the accumulated additional information set W1. The stored additional information feature set Y, which is a set of stored additional information features, is output to the stored latent variable extraction unit 3. The method of extracting the accumulated additional information feature set Y is not particularly limited, but in the present embodiment, the following three examples will be described.

[First embodiment]
The accumulated additional information feature is a binary vector that represents the presence or absence of a word included in the text label. That is, the accumulated additional information feature is a vector having the same number of dimensions as the total number of words to be considered, and each dimension of the vector corresponds to a word. For convenience, the word is expressed using the corresponding dimension index in the vector. The accumulated additional information feature extraction unit 2 sets the i-th dimension of the accumulated additional information feature to 1 if the word i is included in the accumulated additional information, and 0 otherwise.

[Second Embodiment]
The word co-occurrence vector Y = {y ₁ , y ₂ ,..., Y _N } is used as the accumulated additional information feature. Storing the additional information feature extraction unit 2, the d-th element y _n of word cooccurrence vector y _n, the _d, and d-th number of words included in the text labels w _n.

[Third embodiment]
Use LSA (Lentent Semantic Analysis). Considering a matrix W in which the vector derived by the method of the first or second embodiment is a column vector, the accumulated additional information feature extraction unit 2 performs singular value decomposition on this matrix W, so that W = UΣV ^T ( Obtain orthonormal matrices U and V and diagonal matrix Σ satisfying V ^T (transpose matrix of V). The singular values of W are arranged in the diagonal component of Σ. At this time, it is assumed that the U and V columns are appropriately rearranged so that the value of Σ is in descending order from the upper left (in descending order). Next, the accumulated additional information feature extraction unit 2 uses a submatrix ΣDy of Σ upper left Dy × _Dy and a submatrix U _Dy extracted up to the left Dy column of U for a predetermined _Dy , storing the additional information, wherein _{y n (n = 1,2, ···} , n) and obtained by _{^{y n = U Dy T Σ Dy}} w n.

The accumulated latent variable extraction unit 3 receives the complete accumulated image feature set _Xc , the incompletely accumulated image feature set _XI, and the accumulated additional information feature set Y as inputs, and from these feature sets, an image (signal) and an additional An accumulated latent variable set Z = {z ₁ , z ₂ ,..., Z _{Nx + N} }, which is a set of variables for describing information relationships, is extracted, and this accumulated latent variable set is extracted to the topic model learning unit 4. Output. Details of the processing in the accumulated latent variable extraction unit 3 will be described later.

Topic model learning unit 4, a full storage image feature set X _C, and an incomplete storage image feature set X _I, and the accumulated additional information feature set Y, and inputs the accumulation latent variable set Z, from these sets, images A topic model, which is a model that describes the relationship between and the additional information, is learned (extracted), and the topic model is output. Details of the processing in the topic model learning unit 4 will be described later.

Next, details of the processing in the accumulated latent variable extraction unit 3 will be described. FIG. 2 is a block diagram showing a functional configuration of the accumulated latent variable extraction unit 3 according to the present embodiment.
The accumulated latent variable extraction unit 3 includes a function f _x : R ^Dx → R ^Dz (a function that generates a set of Dz-dimensional real vectors from a set of Dx-dimensional real vectors), and a function f _xy : R ^Dx × R ^Dy → R ^Dz A set (f _x , f _xy ) of (a function that generates a set of Dz-dimensional real vectors from a set of Dx-dimensional real vectors and a set of Dy-dimensional real vectors) is obtained from (X, Y), and these functions Is used to generate Z from (X, Y). Here, X is the accumulation image feature set that matches the full storage image feature set X _c and incomplete accumulated image feature set X _I. Dx is the dimension of the stored image feature, Dy is the dimension of the stored additional information feature, and Dz is the dimension of the stored latent variable. In the present embodiment, the accumulated latent variable set is extracted using a method obtained by improving canonical correlation analysis, which is a kind of multivariate analysis. The accumulated latent variable extracting unit 3 includes a latent variable space generating unit 31, a complete accumulated latent variable set extracting unit 32, and an accumulated latent variable set extracting unit 33.

The latent variable space generation unit 31 is a latent variable that is a multidimensional vector space in which accumulated latent variables exist from the completely accumulated image feature set _Xc , the incompletely accumulated image feature set _XI, and the accumulated additional information feature set Y. Create a space. The generation method of the latent variable space is not particularly limited, but in the present embodiment, the latent variable space generation unit 31 is derived by solving the generalized eigenvalue problem given by the following equation (1). The generalized eigenvalue problem is a problem for obtaining an eigenvalue λ and an eigenvector a that satisfy the equation (1).

  Here, a is a (Dx + Dy) -dimensional vector. However, B in Formula (1) is represented by the following Formula (2).

  Moreover, C in Formula (1) is represented by following Formula (3).

Where ^T means transpose (vector or matrix). I _d represents a d × d unit matrix. In addition, the character which attached | subjected the hat (^) to the character upper part in a type | formula shows and describes a hat (^) before a character in a sentence. B and each element hat constituting the _{^{C (^) S (C)}} xx, hat _{^{(^) S (C) yy}} , hat _{^{(^) S (C) xy}} , hat _{^{(^) S (C) yx}} , hat ( ^) S _xx and hat (^) S _yy are expressed by the following equations (4) to (9).

  Here, β is a real number given in advance. For the solution of the generalized eigenvalue problem, as a reference, Hardon, Szedmak and Shawe-Taylor “Canonical correlation analysis: An overview with applications to learning methods,” Neural Computation, Vol.16, Issue 12, pp.2639- 2664, December 2004.

  Solving the generalized eigenvalue problem shown in Equation (1) to obtain a set of a predetermined number of eigenvalues and eigenvectors, or a set of a minimum number of eigenvalues and eigenvectors whose sum of eigenvalues exceeds a predetermined threshold. Alternatively, the latent variable space can be determined by obtaining the number of eigenvalues and eigenvectors determined by the method using the Bartlett test. Regarding the Bartlett test, there is M.S. Bartlett “THE STATISTICAL SIGNIFICANCE OF CANONICAL CORRELATIONS,” Biometrika, Vol. 32 pp. 29-37, 1941.

However, Dz is the number of extracted eigenvalues and eigenvectors and satisfies Dz ≦ min (Dx, Dy). Each eigenvector a _d (d = 1, 2,..., Dz) can be decomposed into a leading Dx-dimensional vector a _{x, d} and a subsequent Dy-dimensional vector a _{y, d} . Thereby, the latent variable space generation unit 31 outputs a matrix (A _x , A _y ) having eigenvectors obtained by decomposing the latent variable space as columns. Here, A _x = {a _{x, 1} , a _{x, 2} , a _{x, 3} ,..., A _{x, Dx} }, and A _y = {a _{y, 1} , a _{y, 2} , a _{y, 3} ,..., A _{y, Dy} }.

The complete storage latent variable set extraction unit 32 uses the complete storage image feature set _Xc , the storage additional information feature set Y, and the latent variable space A to store a storage latent variable set corresponding to the complete storage image feature set _{Xc. Is} extracted as a fully accumulated latent variable set Zc. Specifically, first, the complete accumulation latent variable set extraction unit 32 determines f _xy by the following equation (10). In addition, a character with “˜” in the upper part of the character in the formula is indicated by “˜” before the character in the sentence.

Here, Λ is a diagonal matrix having the d-th eigenvalue λ _d (d = 1, 2,..., Dz) as the d-th diagonal component in the solution of Equation (1). M _x and M _y are Dz-dimensional square matrices that satisfy M _x M _y ^T = Λ and whose spectral norm is smaller than 1. ˜Λ is expressed by the following equation (11).

Then, the complete accumulation latent variable set extraction unit 32 calculates each element z _n (n = 1, 2,... N) of the complete accumulation latent variable set Z _c from the following equation (12), whereby the complete accumulation latent variable is calculated. A set _Zc is extracted.

Accumulated latent variable set extraction unit 33, a full storage image feature set X _c, and an incomplete storage image feature set X _I, by using the latent variable space A, the storage potential corresponding to the incompletely stored image feature set X _I extracting the incomplete storage latent variable set Z _I is a variable set. Then, the accumulated latent variable set extraction unit 33 combines the complete accumulated latent variable set Z _c and the incomplete accumulated latent variable set Z _I to form the entire accumulated latent variable set Z. Specifically, first, accumulated latent variable set extraction unit 33 defines the _{f x} by the following equation (13).

Then, the accumulated latent variable set extraction unit 33 calculates each element z _n (n = N + 1, N + 2,... N + Nx) of the incomplete accumulated latent variable set Z _I from the following equation (14), thereby obtaining the incomplete accumulated latent potential. to extract a set of variables _{Z I.}

The accumulated latent variable set extraction unit 33, a storage latent variable set Z together with full accumulation latent variable set Z _c and incomplete accumulated latent variable set Z _I, and outputs the accumulated latent variable set Z.

Next, details of processing in the topic model learning unit 4 will be described. FIG. 3 is a block diagram showing a functional configuration of the topic model learning unit 4 according to the present embodiment.
The topic model learning unit 4 includes a latent variable model learning unit 41, a signal / latent variable relationship model learning unit 42, an initial additional information / latent variable relationship model learning unit 43, an additional information / latent variable relationship model updating unit 44, The additional information / latent variable relationship model correcting unit 45 and the convergence determining unit 46 are included.

  The latent variable model learning unit 41 receives the accumulated latent variable set Z, learns (extracts) a latent variable model that is a model describing the stochastic structure of the accumulated latent variable itself, and outputs the latent variable model. The learning method of the latent variable model is not particularly limited, but in the present embodiment, the following three examples will be described. In either case, the occurrence probability p (z) of the accumulated latent variable z is adopted as the latent variable model.

[First embodiment]
The latent variable model learning unit 41 calculates a latent variable model p (z) by multipoint approximation of the following equation (15) using the accumulated latent variable set Z. Here, Kz (z) is a probability density function with z as a parameter. For example, Dirac delta function or normal distribution can be used as Kz. That is, the latent variable model learning unit 41 extracts a latent variable model p (z) by superimposing probability density functions using the accumulated latent variable z as a parameter.

[Second Embodiment]
The latent variable model learning unit 41 considers the probabilistic structure of the canonical correlation analysis and calculates the latent variable model p (z) by the following equation (16). However, the function N is a normal distribution in which the first argument is an average and the second argument is a covariance matrix.

[Third embodiment]
The latent variable model learning unit 41 calculates a latent variable model p (z) by a mixed normal distribution represented by the following equation (17).

However, the number of Mz is mixed distribution, omega _m is the weight of the m-th distribution ^{_{(Σ m = 1 Mz ω m}} = 1), ¯z m average of the m-th distribution, sigma _zm is the m-th distribution Represents the variance. If Mz is determined in advance, ω _m , ￣z _m , and Σ _zm can be obtained by the EM algorithm.

  The signal / latent variable relationship model learning unit 42 receives a stored image feature set X and a stored latent variable set Z, and uses these sets as a signal that describes a relationship between the stored image and the stored latent variable.・ Learning (extracting) the latent variable relationship model and outputting the signal / latent variable relationship model. The learning method of the signal / latent variable relationship model is not particularly limited, but in the present embodiment, the following two examples will be described. In either case, the conditional occurrence probability p (x | z) of the stored image feature x when the stored latent variable z is given is adopted as the signal / latent variable relationship model.

[First embodiment]
The signal / latent variable relationship model learning unit 42 calculates a signal / latent variable relationship model p (x | z _n ) for each element z _n of the accumulated latent variable set by the following equation (18). That is, the signal / latent variable relationship model learning unit 42 extracts a signal / latent variable relationship model individually for each accumulated latent variable z _n and based on a probability density function using the accumulated image feature x as a parameter.

  Here, ~ z is an accumulated latent variable obtained by converting the accumulated image feature x by the accumulated latent variable extracting unit 3, and Kx (z) is a probability density function having z as a parameter. As Kx, for example, Dirac delta function or normal distribution can be used.

[Second Embodiment]
The signal / latent variable relationship model learning unit 42 calculates the signal / latent variable relationship model p (x | z) by the following equation (19) in consideration of the probabilistic structure of the canonical correlation analysis.

  The initial additional information / latent variable relationship model learning unit 43 is a model that uses the stored additional information feature set Y and the stored latent variable set Z as inputs and describes the relationship between the additional information and the stored latent variables using these sets. A certain initial additional information / latent variable relationship model is learned (extracted), and this initial additional information / latent variable relationship model is output. The learning method of the initial additional information / latent variable relationship model is not particularly limited, but in the present embodiment, the following three embodiments will be described. In either case, the conditional occurrence probability p (y | z) of the additional information feature y when the accumulated latent variable z is given is adopted as the initial additional information / latent variable relationship model.

[First embodiment]
A method similar to that of the signal / latent variable relationship model learning unit 42 can be used.

[Second Embodiment]
The initial additional information / latent variable relationship model learning unit 43 calculates the initial additional information / latent variable relationship model p (y | z _n ) from the following equations (20) to (23) for each element z _n of the accumulated latent variable set. calculate. That is, the initial additional information / latent variable relationship model learning unit 43 extracts an additional information / latent variable relationship model individually for each accumulated latent variable z _n and based on a probability density function using additional information as a parameter. .

Here, mu is a constant satisfying 0 ≦ μ ≦ 1, y n , d is the d component of the accumulated additional information feature y _n. Ky (y) is a probability density function with y as a parameter. For example, a Kronecker delta function or a multinomial distribution can be used. That is, the above relational expression first assumes that each text label occurs independently (formula (20)), and the occurrence probability of each text label is expressed as an empirical distribution (formula ( 21) Ky (equivalent to y _d −y _{n, d} )) and the empirical distribution of text labels in all samples (corresponding to M _d / M in equation (21)) are generated by mixing at a mixing ratio μ. Means that.

[Third embodiment]
The initial additional information / latent variable relationship model learning unit 43 calculates the initial additional information / latent variable relationship model p (y | z) by the following equation (24) in consideration of the probabilistic structure of the canonical correlation analysis.

  The additional information / latent variable relationship model updating unit 44 includes an initial additional information / latent variable relationship model (or additional information / latent variable relationship model corrected by the additional information / latent variable relationship model correcting unit 45) and a stored additional information feature. From the set Y and the stored latent variable set Z, the (initial) additional information / latent variable relation model is updated in consideration of the distribution of the stored latent variables corresponding to the incompletely stored image feature set, and new additional information / latency Output as a variable relation model.

  Hereinafter, in order to simplify the notation, a matrix notation represented by the following equation (25) is introduced.

P _{m, n} is an (m, n) component of the matrix P and is expressed by the following equation (26).

F _{n, d} is an (n, d) component of the matrix F and is expressed by the following equation (27).

Here, the matrix F is the additional information / latent variable relationship model. Further, the initial additional information / latent variable relationship model obtained by the initial additional information / latent variable relationship model learning unit 43 is F ^(0), and the additional information / latent variable relationship model updated for the kth time is F ^(k) . . The additional information / latent variable relationship model updating unit 44 updates the additional information / latent variable relationship model by the following equation (28), and outputs F ^(k) as a new additional information / latent variable relationship model.

  When there are a large number of stored images, there is a problem that the matrix P becomes too large. In order to avoid this, the following method can be taken.

  First, among the elements of the matrix P calculated by the equation (26), the element having a small value is forcibly rewritten to 0. As a criterion for selecting an element to be rewritten, for example, a method of selecting a predetermined number of elements for each row of the matrix P or selecting an element that falls below a predetermined threshold can be considered. In many embodiments, each element of the matrix P is determined by the distance between two vectors in the latent variable space, so in this case, determining which elements to leave unmodified is in the latent variable space. Equivalent to performing a neighborhood search or range search at.

  Next, a new matrix P hat that retains only the remaining elements without being replaced with 0 is considered, and this is used instead of the matrix P to calculate equation (28).

  The original matrix P requires an area about the square of the number of stored images to hold it, while the new matrix P hat reduces the area required to hold it to the same order as the number of stored images. be able to.

  The additional information / latent variable relationship model correcting unit 45 stores the additional information / latent variable relationship model updated by the additional information / latent variable relationship model updating unit 44 and the stored additional information feature set Y. The additional information / latent variable relationship model is corrected so as to hold the feature information to a certain degree (predetermined value), and is output as a new additional information / latent variable relationship model. The additional information / latent variable relationship model F can be described by the following equation (29).

Here, F ^(k) _L is the additional information / latent variable relationship model of the N × Dy matrix corresponding to the completely accumulated image, and F ^(k) _U is the Nx × Dy matrix of the incompletely accumulated image. Represents additional information / latent variable relationship model. The additional information / latent variable relationship model correcting unit 45 corrects the additional information / latent variable relationship model for the completely stored image in the additional information / latent variable relationship model F by the following equation (30).

Here, T is a diagonal matrix with the (n, n) component being t _n (0 <t _n ≦ 1, n = 1, 2,..., N), It is a regularization parameter for deciding whether to correct the additional information / latent variable relationship model with emphasis on the degree. The value of the matrix T can be freely determined in advance.

The convergence determination unit 46 includes the additional information / latent variable relationship model corrected by the additional information / latent variable relationship model update unit 44 and the additional information / latent variable relationship model correction unit 45, and the additional information / latent variable relationship before correction. The model (the output immediately before the additional information / latent variable relationship model correction unit 45) is compared. If the convergence determination unit 46 determines that the additional information / latent variable relationship model has converged, the convergence determination unit 46 outputs the additional information / latent variable relationship model and ends the process, while determining that the additional information / latent variable relationship model has not converged. For example, the processing shifts to the additional information / latent variable relationship model update unit 44, and the updating / correction processing of the additional information / latent variable relationship model is repeated again. At this time, the convergence determination unit 46 calculates a matrix norm of the corrected additional information / latent variable relationship model F ^{(k + 1)} and the corrected additional information / latent variable relationship model F ^(k−1) , and this matrix norm If the difference is within a predetermined threshold value, it is determined that it has converged, and if it is greater than the threshold value, it is determined that it has not converged.

  The topic model learning unit 4 integrates the latent variable model, the signal / latent variable relationship model, and the additional information / latent variable relationship model into a topic model, and outputs the topic model.

Next, the semi-teacher topic model learning process in the semi-teacher topic model learning apparatus 1 according to the present embodiment will be described with reference to FIGS. FIG. 4 is a flowchart showing the procedure of the semi-teacher topic model learning process.
First, in step S1, the stored signal feature extraction unit 1 extracts the complete stored image set and the image feature of each image from each is an element of the incomplete storage image set, full storage image feature set X _C and incomplete storage An image feature set _XI is generated. Next, in step S2, the accumulated additional information feature extraction unit 2 extracts accumulated additional information features from each accumulated additional information that is an element of the accumulated additional information set, and generates an accumulated additional information feature set Y. Next, in step S3, the accumulation latent variable extractor 3, and a full storage image feature set X _C and incomplete accumulated image feature set X _I and accumulating additional information feature set Y, extracts the accumulated latent variable set Z. Finally, based on in step S4, the topic model learning unit 4, a full storage image feature set X _C, and an incomplete storage image feature set X _I, the accumulated additional information feature set Y, and a storage latent variable set Z Thus, a topic model that is an image / additional information relation model is extracted.

FIG. 5 is a flowchart showing a procedure of accumulated latent variable extraction processing in the accumulated latent variable extracting unit 3 according to the present embodiment.
First, in step S31, the latent variable space generation unit 31, and a full storage image feature set X _C and incomplete accumulated image feature set X _I and accumulating additional information feature set Y, to generate a latent variable space w. Next, in step S32, fully accumulated latent variable set extraction unit 32, full storage image feature set using the X _C and accumulating additional information feature set Y and latent variable space w, a complete accumulation latent variable set Z _C extraction To do. Finally, in step S33, accumulated latent variable set extraction unit 33, by using the latent variable space w completely stored image feature set X _C and incomplete accumulated image feature set X _I, incomplete storage latent variable set Z _I Are extracted to form the entire accumulated latent variable set Z.

FIG. 6 is a flowchart showing the procedure of the topic model learning process in the topic model learning unit 4 according to the present embodiment.
First, in step S41, the latent variable model learning unit 41 learns a latent variable model p (z) from the accumulated latent variable set Z. Next, in step S42, the signal / latent variable relationship model learning unit 42 learns the signal / latent variable relationship model p (x | z) from the accumulated image feature set X and the accumulated latent variable set Z. Next, in step S43, the initial additional information / latent variable relationship model learning unit 43 learns the initial additional information / latent variable relationship model p (y | z) from the accumulated additional information feature set Y and the accumulated latent variable set Z. To do.

  Next, in step S 44, the additional information / latent variable relationship model updating unit 44 adds the initial additional information / latent variable relationship model p (y | z) or the additional information / latent variable relationship model correction unit 45 corrected. Update information / latent variable relationship model. Next, in step S45, the additional information / latent variable relationship model correcting unit 45 corrects the additional information / latent variable relationship model. In step S46, the convergence determination unit 46 compares the corrected additional information / latent variable relationship model with the uncorrected additional information / latent variable relationship model to determine whether or not the convergence has occurred. If it is determined that it has converged, the convergence determination unit 46 outputs the corrected additional information / latent variable relationship model and ends the process. On the other hand, when it determines with not having converged, it transfers a process to step S44.

Thus, according to the present embodiment, the accumulated latent variable extraction unit 3 describes the relationship between the image and the additional information from both the image associated with the additional information and the image not associated with the additional information. By extracting the storage latent variables for this purpose, even when only a small amount of the image associated with the additional information can be used, the relationship between the image and the additional information can be learned with high accuracy. That is, when learning the relationship between an image and additional information, an image that is not associated with additional information is used at the same time, which is higher than when only a small number of images associated with additional information are used. The relationship between the image and the additional information can be learned with accuracy.
Further, the additional information / latent variable relationship model update unit 44 and the additional information / latent variable relationship model correction unit 45 update and correct the additional information / latent variable relationship model until the additional information / latent variable relationship model converges, thereby providing more accurate additional information.・ Latent variable relationship model can be output.

Also, a program for realizing each step shown in FIGS. 4 to 6 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed, thereby executing a semi-teacher. You may perform a topic model learning process. Here, the “computer system” may include an OS and hardware such as peripheral devices.
Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” means a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a CD-ROM, a hard disk built in a computer system, etc. This is a storage device.

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

As described above, the embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to the above, and various design changes and the like can be made without departing from the scope of the present invention. It is possible to
For example, in the present embodiment, an accumulated image set and an accumulated additional information set are input. However, the semi-teacher topic model learning device 10 includes a storage unit, and the accumulated image set and the accumulated additional information set are stored in advance in the storage unit. You may keep it.

  DESCRIPTION OF SYMBOLS 1 ... Accumulated signal feature extraction part 2 ... Accumulated additional information feature extraction part 3 ... Accumulation latent variable extraction part 4 ... Topic model learning part 10 ... Semi-teacher topic model learning device 31 ... Latent variable space generation part 32 ... Complete accumulation latent variable set Extraction unit 33 ... accumulated latent variable set extraction unit 41 ... latent variable model learning unit 42 ... signal / latent variable relationship model learning unit 43 ... initial additional information / latent variable relationship model learning unit 44 ... additional information / latent variable relationship model update unit 45 ... Additional information / latent variable model correction unit 46 ... Convergence determination unit

Claims (9)

A semi-teacher topic model learning device that learns the relationship between a given signal and additional information that is information describing the signal,
A complete accumulation signal that is an element of a complete accumulation signal set that is a set of signals to which additional information is given in advance, and an incomplete accumulation signal that is an element of an incomplete accumulation signal set that is a set of signals to which no additional information is given An accumulated signal feature extraction unit that extracts an accumulated signal feature that is a vector expressing the characteristics of each signal from each of the
Accumulated additional information that extracts a stored additional information feature that is a vector expressing the characteristics of the accumulated additional information from each of the accumulated additional information that is an element of the accumulated additional information set that is a set of additional information assigned to the complete accumulated signal A feature extraction unit;
A complete accumulated signal feature set that is a set of accumulated signal features of the complete accumulated signal, an incomplete accumulated signal feature set that is a set of accumulated signal features of the incomplete accumulated signal, and an accumulation that is a set of the accumulated additional information features. An accumulated latent variable extracting unit that extracts an accumulated latent variable that is a variable for describing the relationship between the signal and the additional information from the additional information feature set;
Describe the relationship between the signal and additional information from the complete stored signal feature set, the incomplete stored signal feature set, the stored additional information feature set, and the stored latent variable set that is a set of the stored latent variables. and a topic model learning unit for extracting topic model is a model,
The accumulated latent variable extraction unit includes:
A latent variable space generation unit that generates a latent variable space that is a multidimensional vector space in which a stored latent variable exists, from the complete stored signal feature set, the incomplete stored signal feature set, and the stored additional information feature set;
A fully stored latent variable that extracts a fully stored latent variable set that is a stored latent variable set corresponding to the complete stored signal feature set using the complete stored signal feature set, the stored additional information feature set, and the latent variable space. A set extraction unit;
Using the complete stored signal feature set, the incomplete stored signal feature set, and the latent variable space, an incomplete stored latent variable set that is a stored latent variable set corresponding to the incomplete stored signal feature set is extracted, An accumulated latent variable set extraction unit that forms the entire accumulated latent variable set by combining the incompletely accumulated latent variable set and the complete accumulated latent variable set;
Semi-supervised topic model learning device which is characterized in that said Rukoto equipped with. The topic model learning unit
A latent variable model learning unit that extracts a latent variable model that is a model describing a stochastic structure of the accumulated latent variable itself from the accumulated latent variable set;
A signal / latent variable for extracting a signal / latent variable relationship model, which is a model describing the relationship between a signal and a latent variable, from the complete accumulated signal feature set, the incompletely accumulated signal feature set, and the accumulated latent variable set A relationship model learning unit;
An additional information / latent variable relationship model learning unit that extracts an additional information / latent variable relationship model, which is a model describing the relationship between additional information and latent variables, from the stored additional information feature set and the stored latent variable set; With
The semi-teacher topic model learning device according to claim 1, wherein the latent variable model, the signal / latent variable relationship model, and the additional information / latent variable relationship model are integrated into a topic model. The topic model learning unit
From the additional information / latent variable relationship model, the accumulated additional information feature set, and the accumulated latent variable set, the additional information / latent variable relationship model is considered in consideration of the distribution of latent variables corresponding to the incompletely accumulated signal feature set. Additional information / latent variable relationship model update unit for updating
From the additional information / latent variable relationship model updated by the additional information / latent variable relationship model and the stored additional information feature set, the additional information / latent variable relationship model so that the information of the stored additional information feature set is retained to a certain extent. The semi-teacher topic model learning apparatus according to claim 2 , further comprising: an additional information / latent variable relationship model correcting unit that corrects The latent variable model learning unit extracts a latent variable model by superimposing a probability density function using each accumulated latent variable as a parameter,
The signal / latent variable relationship model learning unit extracts a signal / latent variable relationship model individually for each accumulated latent variable and based on a probability density function having a stored signal feature as a parameter,
The additional information / latent variable relationship model learning unit extracts the additional information / latent variable relationship model individually for each accumulated latent variable and based on a probability density function using the additional information feature as a parameter. The semi-teacher topic model learning device according to claim 2 or 3 . A semi-teacher topic model learning method used in a semi-teacher topic model learning device that learns the relationship between a given signal and additional information that is information describing the signal,
The accumulated signal feature extraction unit includes a complete accumulated signal that is an element of a complete accumulated signal set that is a set of signals to which additional information is given in advance, and an incomplete accumulated signal set that is a set of signals to which no additional information is given. Extracting a stored signal feature, which is a vector representing the characteristics of each signal, from each of the incompletely stored signals as elements;
The accumulated additional information feature extraction unit is a vector that expresses the characteristics of accumulated additional information from each of the accumulated additional information that is an element of the accumulated additional information set that is a set of additional information added to the complete accumulated signal. Extracting information features;
An accumulated latent variable extraction unit includes a complete accumulated signal feature set that is a set of accumulated signal features of the completely accumulated signal, an incomplete accumulated signal feature set that is a set of accumulated signal features of the incomplete accumulated signal, and the accumulation addition. Extracting a stored latent variable that is a variable for describing a relationship between a signal and additional information from a stored additional information feature set that is a set of information features;
A topic model learning unit includes a signal and additional information from the complete stored signal feature set, the incomplete stored signal feature set, the stored additional information feature set, and a stored latent variable set that is a set of the stored latent variables. look including a step of extracting the topic model is a model that describes the relationship between,
Extracting the accumulated latent variable comprises:
A latent variable space generation unit generates a latent variable space, which is a multidimensional vector space in which accumulated latent variables exist, from the complete accumulated signal feature set, the incomplete accumulated signal feature set, and the accumulated additional information feature set. When,
A complete storage latent variable set extraction unit uses the complete storage signal feature set, the stored additional information feature set, and the latent variable space to store a complete storage latency that is a storage latent variable set corresponding to the complete storage signal feature set. Extracting a set of variables;
The accumulated latent variable set extraction unit uses the complete accumulated signal feature set, the incomplete accumulated signal feature set, and the latent variable space, and is an incomplete accumulation that is an accumulated latent variable set corresponding to the incomplete accumulated signal feature set. Extracting a latent variable set and combining the incompletely stored latent variable set and the complete stored latent variable set to form an entire stored latent variable set;
Semi-supervised topic model learning method which is characterized in containing Mukoto a. Extracting the topic model comprises:
A latent variable model learning unit extracting a latent variable model, which is a model describing a stochastic structure of the accumulated latent variable itself, from the accumulated latent variable set;
The signal / latent variable relationship model learning unit is a signal / latency model that describes a relationship between a signal and a latent variable from the complete accumulated signal feature set, the incompletely accumulated signal feature set, and the accumulated latent variable set. Extracting a variable relationship model;
The additional information / latent variable relationship model learning unit extracts an additional information / latent variable relationship model, which is a model describing the relationship between additional information and latent variables, from the accumulated additional information feature set and the accumulated latent variable set. Including steps and
The semi-teacher topic model learning method according to claim 5, wherein the latent variable model, the signal / latent variable relationship model, and the additional information / latent variable relationship model are integrated into a topic model. Extracting the topic model comprises:
The additional information / latent variable relationship model update unit calculates a latent variable distribution corresponding to the incompletely stored signal feature set from the additional information / latent variable relationship model, the stored additional information feature set, and the stored latent variable set. Updating the additional information / latent variable relationship model in consideration of;
The additional information / latent variable relationship model correction unit obtains a certain amount of information of the accumulated additional information feature set from the additional information / latent variable relationship model updated by the additional information / latent variable relationship model and the accumulated additional information feature set. The method of learning a semi-teacher topic model according to claim 6 , further comprising: correcting the additional information / latent variable relationship model so as to be retained. The step of extracting the latent variable model extracts a latent variable model by superimposing a probability density function having each accumulated latent variable as a parameter,
The step of extracting the signal / latent variable relationship model extracts a signal / latent variable relationship model individually for each accumulated latent variable and based on a probability density function having the accumulated signal feature as a parameter,
The step of extracting the additional information / latent variable relationship model includes extracting the additional information / latent variable relationship model individually for each accumulated latent variable and based on a probability density function having the additional information feature as a parameter. The semi-teacher topic model learning method according to claim 6 or 7 . To a computer as a semi-teacher topic model learning device that learns the relationship between a given signal and additional information that is information describing the signal,
The accumulated signal feature extraction unit includes a complete accumulated signal that is an element of a complete accumulated signal set that is a set of signals to which additional information is given in advance, and an incomplete accumulated signal set that is a set of signals to which no additional information is given. Extracting a stored signal feature, which is a vector representing the characteristics of each signal, from each of the incompletely stored signals as elements;
The accumulated additional information feature extraction unit is a vector that expresses the characteristics of accumulated additional information from each of the accumulated additional information that is an element of the accumulated additional information set that is a set of additional information added to the complete accumulated signal. Extracting information features;
An accumulated latent variable extraction unit includes a complete accumulated signal feature set that is a set of accumulated signal features of the completely accumulated signal, an incomplete accumulated signal feature set that is a set of accumulated signal features of the incomplete accumulated signal, and the accumulation addition. Extracting a stored latent variable that is a variable for describing a relationship between a signal and additional information from a stored additional information feature set that is a set of information features;
A topic model learning unit includes a signal and additional information from the complete stored signal feature set, the incomplete stored signal feature set, the stored additional information feature set, and a stored latent variable set that is a set of the stored latent variables. to execute the steps of extracting the topic model is a model that describes the relationship between,
Extracting the accumulated latent variable comprises:
A latent variable space generation unit generates a latent variable space, which is a multidimensional vector space in which accumulated latent variables exist, from the complete accumulated signal feature set, the incomplete accumulated signal feature set, and the accumulated additional information feature set. When,
A complete storage latent variable set extraction unit uses the complete storage signal feature set, the stored additional information feature set, and the latent variable space to store a complete storage latency that is a storage latent variable set corresponding to the complete storage signal feature set. Extracting a set of variables;
The accumulated latent variable set extraction unit uses the complete accumulated signal feature set, the incomplete accumulated signal feature set, and the latent variable space, and is an incomplete accumulation that is an accumulated latent variable set corresponding to the incomplete accumulated signal feature set. Extracting a latent variable set and combining the incompletely stored latent variable set and the complete stored latent variable set to form an entire stored latent variable set;
Program characterized Rukoto allowed to run.

Images