JP6257076B2 - Information processing apparatus, information processing method, and information processing program - Google Patents

Description

The present invention relates to an information processing apparatus, an information processing method, and an information processing program.

Extraction of n-grams (n-gram (n is an integer)) is made possible by modeling target document data in units of n words and as a result of this modeling. For example, Non-Patent Document 1 describes that n-gram extraction is performed based on a hierarchical probability model based on Bayesian theory.
In the model shown in Non-Patent Document 1, a latent variable of a topic is acquired from a document and a random variable of a word is acquired from a topic using a hierarchical class. In the model shown in Non-Patent Document 1, the state between the next words is acquired from the previous word and the previous topic, and the random variable of the next word is acquired from the state between the next words. . In this way, in Non-Patent Document 1, n-grams consisting of consecutive n words are extracted.

Xuerul Wang, Mc Callum A. et al. Xing Wei, Topical N-grams: Phrase and Topic Discovery, with an Application to Information Retrieval, ICDM 2007, 2007 Oct. pp. 679-702

  However, in the model shown in Non-Patent Document 1, the number of latent variables of a topic needs to be determined in advance. Further, in the model shown in Non-Patent Document 1, it is difficult to extract an n-gram having an optimal number of words. Thus, there was a problem that convenience when extracting n-grams was not sufficient.

This invention is made in view of the above-mentioned subject, and this invention makes it a subject to provide the information processing apparatus, the information processing method, and information processing program which can improve the convenience at the time of extracting n-gram. .

(1) The present invention has been made in order to solve the above-described problems, and one aspect of the present invention is a classification apparatus, which determines the number K of latent variables (K is an integer) from a probability distribution, a latent variable estimator for estimating a latent variable for extracting words of n (n is an integer) gram, and for each latent variable estimated by the latent variable estimator, n gram words or n gram words And a word estimation unit for estimating an appearance probability.

(2) One aspect of the present invention is a classification device, a latent variable estimation unit for estimating a latent variable for extracting n (n is an integer) gram words, and the number of n gram words from a probability distribution And a word estimation unit for estimating the appearance probability of an n-gram word or an n-gram word for each latent variable estimated by the latent variable estimation unit. .

(3) One aspect of the present invention is the classification device according to any one of the above, wherein the latent variable estimation unit determines the number K of latent variables by a Chinese restaurant process. It is.

(4) One aspect of the present invention is the classification device according to any one of the above-described classification devices, wherein the word estimation unit estimates n-gram words by a Pitman-Yor process.

(5) One aspect of the present invention is any one of the above-described classification apparatuses, wherein a topic is used as the latent variable.

(6) One aspect of the present invention is a classification method, wherein the number of latent variables K (K is an integer) is determined from a probability distribution, and latent variables for extracting n (n is an integer) gram words are determined. A latent variable estimation process to be estimated; and a word estimation process for estimating an n-gram word or an appearance probability of the n-gram word for each latent variable estimated by the latent variable estimation process. This is a classification method.

(7) One aspect of the present invention is a classification method, a latent variable estimation process for estimating latent variables for extracting n (n is an integer) gram words, and the number of n gram words from a probability distribution For each latent variable estimated by the latent variable estimation process, and a word estimation process for estimating an appearance probability of the n-gram word or the word of the n-gram. .

(8) One aspect of the present invention is a classification program, in which a computer of a classification apparatus determines the number of latent variables K (K is an integer) from a probability distribution, and extracts n (n is an integer) gram words A latent variable estimating step for estimating a latent variable for performing, and for each latent variable estimated by the latent variable estimating step, a word estimation step for estimating an n-gram word or an appearance probability of the n-gram word. This is a classification program for execution.

(9) One aspect of the present invention is a classification program, a latent variable estimation step for estimating a latent variable for extracting words of n (n is an integer) gram in a computer of a classification device, and a probability distribution a word estimation step for determining the number of n-gram words and estimating an occurrence probability of the n-gram word or the n-gram word for each latent variable estimated by the latent variable estimation step. Classification program.

  According to the present invention, convenience when extracting n-grams can be improved.

It is the schematic which shows an example of the graphical model which concerns on embodiment of this invention. It is explanatory drawing which shows an example of the symbol which concerns on this embodiment. It is a system configuration figure showing an example of the composition of the classification system concerning this embodiment. It is a schematic block diagram which shows an example of a structure of the classification device which concerns on this embodiment. It is a flowchart which shows an example of the description of the calculation process in the classification device concerning this embodiment. It is explanatory drawing explaining an example of the processing content of the calculation process in the classification device concerning this embodiment. It is explanatory drawing explaining an example of the processing content of the calculation process in the classification device concerning this embodiment. It is the schematic which shows an example of the graphical model which concerns on the comparative example of this invention. It is explanatory drawing explaining an example of the effect of the classification device concerning this embodiment.

<Embodiment>
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a graphic model according to an embodiment of the present invention. The symbols and their definitions in this embodiment are as shown in FIG.

Node 11 is a node of parameter α. The parameter α is a hyper parameter (Dirichlet parameter) that generates a probability distribution for obtaining the number of topics and topics.
The node 12 (latent variable estimation unit) is a node for _obtaining the latent variable Z _{ij of the} topic. Here, i is a token (word) number, and j is a document number. The topic latent variable Z _ij represents the i th token in the j th review document. In the present embodiment, a Chinese restaurant process is introduced as a topic probability distribution generation process. By introducing the Chinese restaurant process, the number of topics (K (K is an integer)) is appropriately determined without presetting.

Node 13 is a node of parameter λ. The parameter λ is a concentration parameter that generates a probability distribution for each topic for obtaining the rating.
The node 14 is a node for obtaining a rating observation variable v _j . The review document includes metadata that is accompanying information accompanying the review document. The metadata includes information indicating evaluation on the review document, information indicating the creation date and time of the review document, information indicating the browsing date and time of the review data, information on the number of reviews of the review document, and the like. Rating is performed based on metadata information for each review document. The same word that exists in each review document may have different meanings depending on the topic of the review document. For example, “small” or “light” has a positive meaning when the topic is a mobile device, but has a negative meaning when the topic is fruit. The node 14 obtains a rating observation variable v _j for each review document j using a beta distribution of metadata for each topic.

  In this embodiment, the rating is treated as a continuous value and the beta distribution is used as the rating probability distribution. However, the rating may be treated as a discrete value and the multinomial distribution may be used as the rating probability distribution.

Nodes 15 ₁ , 15 ₂ ,..., 15 _n−1 are nodes of the parameter γ (λ ₀ , λ ₀ ,..., Λ _n−1 ). The parameter γ (λ ₀ , λ ₀ ,..., Λ _n−1 ) is a concentration parameter.
Nodes 16 ₁ , 16 ₂ ,..., 16 _n−1 are nodes of the parameter d (d ₀ , d ₀ ,..., D _n−1 ). The parameter d (d ₀ , d ₀ ,..., D _n−1 ) is a discount parameter.

Nodes 17 ₁ , 17 ₂ ,..., 17 _n−1 are nodes for obtaining a probability distribution G (G _b , G _k ,..., G _k ^u ) of each word in the n-gram. In this embodiment, the nodes 17 ₁ , 17 ₂ ,..., 17 _n-1 use a Pitman-Yor process hierarchically from the parameter γ, the parameter d, and the previous parameter G. Thus, the probability distribution of each word of the n-gram is obtained. That is, the node 17 ₁ generates a probability distribution G _b underlying used throughout general document set. Node _{17 2, ..., 17 n-} 1 is, Pitman-Yor process is hierarchically used things, the probability distribution of each probability distribution, for each word n-gram topic k probability distribution Gk, _{..., G} ^{k u} Ask for.

Node 18 (word estimator) includes a latent variable _{Z ji} topics determined by the node 12, the probability distribution of each word by topic _{k G b, G k, ...} , and _G ^{k u,} from the review document j Find the observation variable w _ji of the n-gram word of token i. Here, the number of n-gram words is appropriately set by hierarchically using the Pitman-Yor process and introducing a technique similar to the Chinese restaurant process.

  The Chinese restaurant process is used in a non-parametric directory process to generate a number separator. In the present embodiment, this Chinese restaurant process is used as a topic prior probability distribution.

The Chinese restaurant process is described using a restaurant metaphor with an infinite number of circular tables, each table having an infinite number of seating capacities. Assume that the tables are numbered sequentially, and Z _i indicates the number of the table (topic) where the i-th customer sits. When a customer enters a store, the probability that the customer will choose a table where someone is sitting is proportional to the number of people who are already sitting, and the possibility that a customer will choose a table where no one is sitting is a certain number. Proportional to parameter. That is, the first customer sits at the first table (Z _i = 1), and then the probability distribution of the i-th customer sitting at the table is as shown in equation (1).

  The Chinese restaurant process can be described as the probability of the Dirichlet process category in the sense of obtaining a random break from the order of sitting at the table in the restaurant. The customer selects the table and sits down to get a number break. It is similar to the cluster structure obtained from the Dirichlet process.

  As described above, by introducing the Chinese restaurant process as a process for generating the probability distribution of topics, the number K of topics is appropriately determined without being set in advance.

Next, rating estimation in the present embodiment will be described. As described above, in this embodiment, the beta distribution is used to obtain the observation variable v _j of the rating in each review document. This processing procedure can be described in a manner similar to the Chinese restaurant process. It is described as a probability similar to the Dirichlet process in the sense of obtaining a random break from the order of customers sitting at a restaurant table.

That is, if the topic Z _ij is an index of a table selected by the i-th customer in the review document (restaurant) j, the distribution is as follows.

The hyperparameter α in this equation is estimated by auxiliary variable sampling. Words and given ratings are combined through a topic using a beta distribution to simultaneously achieve dimensionality reduction in word presentation and corresponding rating decisions. Therefore, the rating probability v _j is given as follows.

  Next, n-gram word estimation in this embodiment will be described. As described above, in this embodiment, the probability distribution of each word of the n-gram is obtained using the Pitman-Yor process hierarchically. The power law can be controlled by using the Pitman-Yor process hierarchically.

  In other words, the power law is a mathematical relationship between two quantities, and is known in linguistics as Zip's law. For example, the probability P (nw) of a word w that occurs at a frequency nw in a sufficiently large corpus is given as follows:

  The Pitman-Yor process performs non-parametric estimation in a Bayesian framework by defining a Dirichlet process and defining a probability distribution of a probability distribution on a dimensionless parameter space. The Pitman-Yor process is shown as follows.

  Here, the discount parameter d is a parameter for controlling the nature of the power law.

In the Pitman-Yor process of the above equation, the probability distribution G _b is set as the parameter G ₀ , and the probability distribution obtained thereby is set as G _0. A probability distribution of words can be obtained.

The procedure for generating G to PYP (γ, d, G ₀ ) can be described using the Chinese restaurant process metaphor according to equation (1). The parameters d and γ produce the effect of smoothing the power law.

That is, the first customer sits at the first table (z ₁ = 1), and the table selected by the i-th customer follows the following distribution.

  Here, when the number of tables increases as well as the number of customers entering the restaurant, this discount parameter produces a power law. That is, the Pitman-Yor process yields results that follow a power law distribution.

  Moreover, the number of words of the n-gram connected model can be set appropriately by introducing the Chinese restaurant process method. That is, in this embodiment, n-gram words are generated by recursively using the Pitman-yor process. In order to form n-gram words, the present embodiment samples words hierarchically.

Probability distribution G _b of the basic word is the probability vector of the unigram (1 grams) word underlying assigned to a topic k = 0. The probability distribution _Gb is a probability distribution obtained for the entire general document set without using previous information. Probability distribution G _b of the basic words, can be obtained as follows.

Probability distribution of topic-specific unigram is acquired by the probability distribution G _b and the current topic k of the base of a given word. The probability distribution G _{k is} obtained by placing the obtained probability distribution G _b as the next parameter G ₀ .

  Here, each topic is regarded as a restaurant owner in order to generate a probability distribution of n-gram words for each topic. This extends the Chinese restaurant process to the Chinese restaurant franchise process.

For example, each word probability distribution G _k ^u is indexed by a restaurant (word number) u and has an associated restaurant managed by the owner (topic) k. Here, probability distribution G _k ^u is the probability of conditional w following u, and u is the number of the previous n−1 words in topic k. The restaurant customers are taken from the probability distribution G _k ^u , the table is obtained from the basic probability distribution before the probability distribution, and the dish is the value obtained from the word.

  By placing the Pitman-yor process recursively as a prior probability distribution of word probabilities, each n-gram derived from the same consecutive topic assignment in each document is defined.

The probability distribution of the next word can be acquired by recursively placing the previous probability distribution G _k ^u using equation (9). The concentration parameter γ and the discount parameter d have a function related to the length of the phrase. This calculation is repeated until a background word probability distribution _Gb is obtained.

For each restaurant (number of words) u and dishes (words) w, c ^k _uwl and t ^k _uwl are respectively _stored in the table u at the restaurant u managed by k and the customers who are eating dishes (words) w And the number of tables of dishes w to be provided by k at the same restaurant u. By indicating the limit number, c ^k _ul is the number of customers sitting in the table l managed by ^k , c ^k _uw is the eaten dish w in u managed by ^k , and c ^k _u is , K is the number of customers managed by ^k , and t ^k _u is the number of u tables managed by k. As a result, the next word w extracted from G _k ^u , G _k and G _b is recursively calculated as follows.

FIG. 2 is an explanatory diagram illustrating an example of symbols according to the present embodiment.
As shown in the figure, D (K, W) represents the number of review documents (number of topics, number of words). N _j represents the number of word tokens in the review document j. v _j represents the association number rating for the review document j. z _ji represents the topic associated with the rank token in the review document j. w _ji represents the i-th token in the review document j. u represents the order of words (phrase). u (/) represents the order of words. G _b represents the probability distribution of words in a given corpus. G _k represents a probability distribution of words specific to topic k. G _k ^u represents the probability distribution of a unique word u in topic k. α represents a hyper parameter. d _n represents the discount parameters. γ _n represents a concentration parameter.

Next, the process shown by the graphic model as shown in FIG. 1 will be specifically described. As described above, in the present embodiment, a Chinese restaurant process is introduced, the number of latent variables Z _ji is determined, and the Pitman-yor process is used hierarchically to estimate an n-gram connected model. ing. These processes can be realized by Gibbs sampling.

  FIG. 3 is a schematic diagram illustrating an example of the configuration of the classification system according to the present embodiment, and FIG. 4 is a schematic block diagram illustrating an example of the configuration of the classification device. As shown in FIG. 3, the system according to the embodiment of the present invention includes a file server 51, a calculation server 52, a database 53, and a service server 54.

  As illustrated in FIG. 4, the file server 51 includes a data file storage unit 61 that stores review data to be calculated. The review data stored in the data file storage unit 61 is a blog document on the Internet, a web page document, or the like. The review data is document data in which metadata and a review are associated with each other. Note that the review data to be calculated is not limited to documents on the Internet.

  The calculation server 52 extracts the review data to be calculated from the file server 51, performs a calculation process as shown by the graphic model shown in FIG. 1, and outputs the calculation result. The calculation server 52 includes a preprocessing unit 71 and a calculation processing unit 72, as shown in FIG. The pre-processing unit 71 extracts words from a file including review data to be calculated. Then, the preprocessing unit 71 adds the text ID and the word ID, and stores the correspondence table in the file system. The calculation processing unit 72 performs calculation processing corresponding to the graphic model shown in FIG.

  As shown in FIG. 4, the database 53 has a calculation result storage unit 81. The calculation result of the calculation processing unit 72 of the calculation server 52 is sent to the database 53 and stored in the calculation result storage unit 81.

The service server 54 is a server for providing calculation results for use of the service. As shown in FIG. 4, the service server 54 includes a calling unit 91. In response to the call from the user terminal 55, the calling unit 91 sends the calculation result to the user terminal 55. This calculation result is used for various services such as product search, review search, marketing, etc. that handle a document set including observation values (for example, continuous value observation variable v _j in FIG. 2) such as feedback from the user and subjective evaluation. it can.

FIG. 5 is a flowchart illustrating an example of description of calculation processing in the calculation server 52.
In FIG. 5, first, the pre-processing unit 71 of the calculation server 52 extracts words and metadata (creation date, evaluation, etc.) for each text to be calculated. And the pre-processing part 71 performs processing which allocates text ID to each text, and allocates word ID to each word (step S1).
That is, in the table T1 in FIG. 6, “text AAAB” is described in the text of the first record, and “text XDCFR” is described in the text of the last record. In the token (word) of the first record, “apple”, “operability”, and “heritage” are described. In the token of the last record, “music”, “art”, and “apple” are described.

  A table T2 in FIG. 6 is an example in the case where a process of assigning a text ID and a word ID is performed on such data in step S1. In the table T2 of FIG. 6, “000” is assigned as the text ID to the text of the first record, and “086” is assigned as the text ID to the text of the last record. In addition, “0000”, “0003”, and “0120” are assigned as word IDs to tokens (words) of the first record. Then, “12111234”, “03042”, and “0000” are assigned as word IDs to the token of the last record.

  Next, the calculation processing unit 72 of the calculation server 52 generates a random number and sets the value as a random variable (Z) (step S2). The table T3 of FIG. 7 is a table in which random numbers are generated for the table T2 to which the text ID and the word ID are assigned in step S2, and are used as a topic random variable (Z). That is, in the first record of the table T3, random numbers “11”, “8”, and “3” are entered as topic random variables. In the last record, random numbers “2”, “1”, and “11” are entered as topic random variables.

  Next, the calculation processing unit 72 of the calculation server 52 estimates a latent variable by Gibbs sampling (step S3). And if the number of sampling is preset, a process will be complete | finished (step S4). Accordingly, it is possible to estimate the topic latent variables by optimally determining the number of topic latent variables, and to estimate the n-gram using the estimated topics.

<Comparative example>
FIG. 8 is a graphic model of a comparative example. In FIG. 8, a node 101 is a hyper parameter α node. The hyper parameter α is used to obtain a topic probability distribution θ _d .
The node 102 is a node for obtaining a probability distribution θ _d of a unique topic for each document data. Here, D represents the number of documents, and the topic probability distribution θ _d is D types existing for each document.

Nodes 103 ₁ ,..., 103 _i , 103 _{i + 1} ,... Are nodes that acquire latent variables z (z ₁ ,..., Z _i , z _{i + 1} ,...) Of topics. That is, the node _{103 1, ..., 103 i,} 103 i + 1, ... , from the document data, based on the probability distribution theta _d specific topics for each document obtained by the node 102, the topic of the latent variable _z 1, ..., Get z _i , z _{i + 1} ,... Here, z _i is a topic related to the i-th word (i is an arbitrary integer) (token: the smallest unit of words).

The node 104 is a hyper parameter β node. The hyperparameter β is used to obtain a word probability distribution φ.
The node 105 is a node for obtaining a unique word probability distribution φ for each topic. The number of topics is Z, and the word probability distribution φ of the node 105 is of Z types.

The node 106 is a node of the hyper parameter ε. The hyperparameter ε is used to obtain the probability distribution σ of the next inter-word state.
The node 107 is a node for obtaining the probability distribution σ of the next inter-word state peculiar to the previous word and the previous topic. Note that the probability distribution σ of the next inter-word state is of (Z × W) types.

The node 108 is a hyper parameter γ node. The hyperparameter γ is used to obtain the probability distribution ψ of the next word. For the hyperparameter γ, a random value is used as an initial value.
The node 109 is a node for obtaining a probability distribution ψ of the next word that is unique for each of the previous word and the current topic. Here, Z represents the number of topics, W represents the number of words, and the probability distribution ψ of the next word is of (Z × W) types.

Nodes 110 ₁ ,..., 110 _i , 110 _{i + 1} ,... Are nodes that acquire word observation variables. That is, the node _{110 1, ..., 110 i,} 110 i + 1, ... , the node _{103 1, ..., 103 i,} 103 i + 1, topic _z 1 obtained in _{..., ..., z i, z} i + 1, ... from the node Based on the word probability distribution φ obtained in 105, word observation variables w ₁ ,..., W _i , w _{i + 1} ,. Further, the nodes 110 _i , 110 _{i + 1} ,... Represent the next word observation variables w _i , w _{i + 1} ,... Based on the probability distribution ψ of the next word obtained at the node 109 from the previous word and the current topic. get. Here, w _i represents the i-th word (i is an arbitrary integer).

Nodes 111 _i , 111 _{i + 1} ,... Are nodes that acquire the state between the next word from the previous word and the previous topic. That is, the node ₁₁₁ i, _{111 i + 1,} ..., the node _{103 1, ..., 103 i,} 103 i + 1 ... previous topic _z 1 obtained by, _..., z _{i, z i + 1,} ... and the nodes 110 _1, ... , 110 _i , 110 _{i + 1} ,... Based on the probability distribution σ of the next inter-word state obtained at the node 107 from the previous words w ₁ ,..., W _i , w _{i + 1} ,. The latent variables x _i , x _{i + 1} ,... Of the state between words are acquired.

  As shown in FIG. 8, in the comparative example, when topics are used as latent variables, the number of topics needs to be determined in advance. In the example of FIG. 8, the number of topics is Z. In this case, Z types of probability distributions are necessary as the probability distribution φ of unique words for each topic. Further, (Z × W) types of probability distributions are required as the probability distribution ψ of the next word. Increasing the number of topics increases the number of processes. If the number of topics is reduced, words for each topic cannot be accurately estimated. On the other hand, in this embodiment, the number of topics can be appropriately set by introducing a Chinese restaurant process.

  In the comparative example shown in FIG. 8, the state between the next words is acquired from the previous word and the previous topic, and the next word is estimated. This configuration is basically a 2-gram connected model. Even if n (n> 3) gram extraction is performed in the comparative example shown in FIG. 8, it is based on the 2-gram model, so that meaningful n-gram extraction cannot be performed. On the other hand, in this embodiment, an n-gram connection model can be realized by hierarchically introducing the Pyman-Yor process. In addition, by expanding and introducing the Chinese restaurant process, the number of words for n-gram extraction can be set appropriately.

FIG. 9 compares the phrases (2 grams, 3 grams) extracted according to the present embodiment with the phrases (2 grams, 3 grams) extracted according to the comparative example.
Here, for the DVD title, the book title, and the music title, the phrase extracted by the present embodiment and the phrase extracted by the comparative example shown in FIG. Rate).

  Here, Precision indicates (number of correct answers in calculation results / number of calculation results), and Recall indicates (number of correct answers in calculation results / number of all correct answers). Precision is an index of how many wrong results are included in the extraction result, and Recall is an index of how much extraction is missed. Basically, Precision and Recall are in a trade-off relationship.

  From the results shown in FIG. 9, it is confirmed that in this embodiment, a better extraction result can be obtained for both Precision and Recall than the comparative example. For example, in the case of 2 grams, in the book title, the Precision is “0.67” and the Recall is “0.60” in the comparative example, but in the present embodiment, the Precision is “0.87” and the Recall. Becomes “0.90”, and good extraction results are obtained for both Precision and Recall. In the case of 3 grams, in the title of the book, the Precision is “0.46” and the Recall is “0.42” in the comparative example, but in this embodiment, the Precision is “0.82” and the Recall. Becomes “0.86”, and good extraction results are obtained for both Precision and Recall.

As described above, according to the present embodiment, the classification device automatically determines the number of latent variables K (K is an integer) from the probability distribution and extracts a word of n (n is an integer) gram. A latent variable estimation unit that estimates a variable, and a word estimation unit that estimates an n-gram word or an appearance probability of the n-gram word for each latent variable estimated by the latent variable estimation unit.
As a result, it is not necessary to determine the number of latent variables (number of topics) in advance when extracting n-grams, and the number of latent variables can be determined automatically. It can be set, and convenience when extracting n-grams can be improved.

Further, according to the present embodiment, the classification device determines a latent variable estimation unit for estimating a latent variable for extracting n (n is an integer) gram word, and determines the number of n gram words from the probability distribution. A word estimation unit that estimates n-gram words or the appearance probability of the n-gram words for each latent variable estimated by the latent variable estimation unit.
Thereby, the number of words representing n in the extraction of n-gram words can be set appropriately, and convenience when extracting n-grams can be improved.

  In addition, by introducing a Chinese restaurant process, when n-grams are estimated using topics as latent variables, the number of topics need not be determined in advance, and the number of topics can be set appropriately. . In addition, the n-gram connection model can be realized by hierarchically introducing the Pyman-Yor process, and the number of n-gram extraction words can be appropriately increased by introducing the Chinese restaurant process. Can be set to

  In the above-described embodiment, extraction of n-grams is explained mainly with respect to English words. However, the present invention is not limited to extraction of n-grams with English words, but also in Japanese and other multilingual languages. can do.

  In the above-described embodiment, the case where n-gram extraction is performed using a topic as a latent variable has been described, but the latent variable is not limited to a topic. Also, when using latent variables other than topics, the number of latent variables can be determined optimally in the same manner by the Chinese restaurant process.

  Further, a program for executing each process of the calculation server 52 of the present embodiment is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read by a computer system and executed. The above-described various processes related to the calculation server 52 may be performed.

  Here, the “computer system” may include an OS and hardware such as peripheral devices. Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. The “computer-readable recording medium” means a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a CD-ROM, a hard disk built in a computer system, etc. This is a storage device.

  Further, the “computer-readable recording medium” refers to a volatile memory (for example, DRAM (Dynamic) in a computer system serving as a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. Random Access Memory)) that holds a program for a certain period of time is also included. The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium.

  Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, what is called a difference file (difference program) may be sufficient.

  As mentioned above, although embodiment of this invention was explained in full detail with reference to drawings, the concrete structure is not restricted to this embodiment, The design of the range which does not deviate from the summary of this invention, etc. are included.

51 File Server 52 Calculation Server 53 Database 54 Service Server 55 User Terminal 61 Data File Storage Unit 71 Preprocessing Unit 72 Calculation Processing Unit 81 Calculation Result Storage Unit 91 Calling Unit

Claims (9)

The number of latent variables K (K is an integer) is determined from the probability distribution of unique words for each topic of the sentence in the sentence, and the number of latent variables used to extract n (n is an integer) gram words. A latent variable estimator for estimating K latent variables;
For each latent variable estimated by the latent variable estimation unit, a word estimation unit that estimates an n-gram word or an appearance probability of the n-gram word;
An information processing apparatus comprising: The information processing apparatus according to claim 1, wherein the latent variable estimation unit determines a number K of the latent variables by a Chinese restaurant process. a latent variable estimator for estimating a latent variable for extracting words of n (n is an integer) gram;
The number of n-gram words is determined from the probability distribution of unique words for each topic of the sentence in the sentence, and for each latent variable estimated by the latent variable estimation unit, n-gram words or n-gram words are determined. A word estimator for estimating the appearance probability;
An information processing apparatus comprising: The information processing apparatus according to claim 1, wherein the word estimation unit estimates n-gram words by a Pitman-Yor process. The information processing apparatus according to any one of claims 1 to 4, wherein a topic is used as the latent variable. Information processing device
The number of latent variables K (K is an integer) is determined from the probability distribution of unique words for each topic of the sentence in the sentence, and the number of latent variables used to extract n (n is an integer) gram words. A latent variable estimation process for estimating K latent variables;
For each latent variable estimated by the latent variable estimation process, a word estimation process for estimating an n-gram word or an appearance probability of the n-gram word;
And an information processing method. Information processing device
a latent variable estimation process for estimating a latent variable for extracting words of n (n is an integer) gram;
The number of n-gram words is determined from the probability distribution of unique words for each topic of the sentence in the sentence, and for each latent variable estimated by the latent variable estimation process, n-gram words or n-gram words are determined. A word estimation process for estimating the appearance probability;
An information processing method characterized by comprising: In the computer of the information processing device,
The number of latent variables K (K is an integer) is determined from the probability distribution of unique words for each topic of the sentence in the sentence, and the number of latent variables used to extract n (n is an integer) gram words. A latent variable estimating step for estimating K latent variables;
For each latent variable estimated by the latent variable estimation step, a word estimation step for estimating the occurrence probability of an n-gram word or the n-gram word;
Information processing program to execute. In the computer of the information processing device,
a latent variable estimation step for estimating a latent variable for extracting n (n is an integer) gram words;
The number of n-gram words is determined from the probability distribution of unique words for each topic of the sentence in the sentence, and for each latent variable estimated by the latent variable estimation step, n-gram words or n-gram words are determined. A word estimation step for estimating an appearance probability;
Information processing program to execute.

Images