JP7043243B2 - Classification device, classification method, and program - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act Date: September 4, 2017 (Symposium 2nd day) Meeting name: NLP Young People's Association (YANS) 12th Symposium Venue: Okinawa Kariyushi Urban Resort Naha (Okinawa) 3-25-1 Maejima, Naha City, Okinawa Prefecture)

The present invention relates to a classification device, a data structure, a classification method, and a program.

Conventionally, the technology of extracting topics (topic analysis) from large-scale text streams accumulated over time such as blogs and WEB news, and extracting correspondence with past and new topics, and the evolution process of topics has been known. ing.

Japanese Unexamined Patent Publication No. 2009-187395

However, in the conventional technology, the number of contents such as news articles to be retained is limited, and what kind of contents should be retained has not been sufficiently examined.

The present invention has been made in view of the above problems, and an object of the present invention is to provide a classification device, a data structure, a classification method, and a program capable of more effectively sorting contents.

In one aspect of the present invention, an acquisition unit for acquiring new content and a category to which the new content is classified by reinforcement learning when the new content is acquired by the acquisition unit are included in a plurality of existing categories. It is a classification device including a learning processing unit that determines whether to select from the above or to generate a new category and classify it into the generated category.

According to one aspect of the present invention, it is possible to provide a classification device, a data structure, a classification method, and a program capable of sorting contents more effectively.

It is a figure which shows an example of the content classification system 1 including the classification apparatus 100 of an embodiment. It is a figure which shows an example of the structure of the classification apparatus 100 of an embodiment. It is a figure for demonstrating the method of classifying a new content CT into a category. It is a flowchart which shows the flow of a series of processing by a learning processing unit 114. It is a figure which shows typically the process of updating a topic vector. It is a figure which shows typically the process of exchanging a new article and an existing topic. It is a figure which shows typically the process which does not classify a new article into an existing topic. It is a figure which shows an example of the result of the reward obtained by the reinforcement learning of this embodiment. It is a figure which shows an example of the hardware composition of the classification apparatus 100 of an embodiment.

Hereinafter, the classification device, the data structure, the classification method, and the program to which the present invention is applied will be described with reference to the drawings.

[Overview]
The classifier is implemented by one or more processors. When the classification device acquires new content, it selects a new category from a plurality of existing categories based on the value function whose parameters are determined by reinforcement learning. Decide if you want to generate it and put it in that category. This allows the classification device to effectively sort the content.

More specifically, the classification device deletes one category from the existing plurality of categories and sets it as the classification destination of the new content. As a result, the content can be classified into suitable categories while keeping the number of categories to be retained constant.

The content in the present embodiment is, for example, an article posted on a blog, a website, or the like, and is content including text. In addition, the content may be content (including text and images) that posts a product or service to be put up for auction, or content such as video, image, or audio posted on a posting site or the like. There may be.

[overall structure]
FIG. 1 is a diagram showing an example of a content classification system 1 including the classification device 100 of the embodiment. The content classification system 1 in the embodiment includes, for example, one or more information providing devices 10 (10-1 to 10-n; n is an arbitrary natural number) and a classification device 100. These devices are connected to each other, for example, via a network NW.

Each device shown in FIG. 1 transmits and receives various information via the network NW. The network NW includes, for example, a radio base station, a Wi-Fi access point, a communication line, a provider, the Internet, and the like. It should be noted that not all combinations of the devices shown in FIG. 1 need to be able to communicate with each other, and the network NW may include a local network in part.

The information providing device 10 is, for example, a server device that distributes the various contents described above.

The classification device 100 collects content from one or more information providing devices 10. For example, the classification device 100 collects contents (for example, news articles) distributed from the information providing device 10. Then, the classification device 100 classifies the collected contents into categories. For example, when the content is an article to be posted on a news site, the classification device 100 classifies the article into a topic (an example of a category).

FIG. 2 is a diagram showing an example of the configuration of the classification device 100 of the embodiment. For example, the classification device 100 includes a communication unit 102, a control unit 110, and a storage unit 130.

The communication unit 102 includes, for example, a communication interface such as a NIC (Network Interface Card) and a DMA (Direct Memory Access) controller. The communication unit 102 communicates with the information providing device 10 and the like via the network NW.

The control unit 110 includes, for example, an acquisition unit 112 and a learning processing unit 114. These components are realized by, for example, a processor such as a CPU (Central Processing Unit) executing a program stored in the storage unit 130. Further, some or all of the components of the control unit 110 are hardware such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or GPU (Graphics Processing Unit). It may be realized by (circuit unit; circuitry), or it may be realized by the cooperation of software and hardware.

The storage unit 130 is realized by, for example, an HDD (Hard Disc Drive), a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), a ROM (Read Only Memory), a RAM (Random Access Memory), or the like. In addition to various programs such as firmware and application programs, the storage unit 130 stores the collected content CT, value function information FX, and the like. Value function information FX is information that defines a value function described later.

The acquisition unit 112 uses the communication unit 102 to communicate with the information providing device 10 and the like to acquire the content CT. Then, the acquisition unit 112 stores the acquired content CT in the storage unit 130.

When the content CT is newly acquired by the acquisition unit 112, the learning processing unit 114 uses a value function whose parameters are determined by a machine learning method called reinforcement learning for the new content (hereinafter referred to as new content). Classify into categories based on. The value function is, for example, a function for evaluating how good the current state of the category in which the content CT is classified is. Good condition is determined, for example, by the amount of rewards that will be obtained in the future if the content CT included in that category is retained.

For example, when the content CT classified into a category is an article, the reward obtained by holding the article may be the number of PV (Page View), CTR (Click Through Rate), or the like. Further, for example, when the content CTs classified into categories are moving images, images, and sounds, the rewards obtained by holding those content CTs include the number of views, views, views, and favorites. If the content CTs classified into categories are items for auction, the rewards obtained by holding those content CTs are the number of successful bids, the amount of successful bids, the number of views, and the number of bids. , Bid amount, etc.

FIG. 3 is a diagram for explaining a method of classifying the new content CT into categories. In the figure, each of A to D represents a category type different from each other. For example, the learning processing unit 114 converts the new content CT to be classified into categories into a vector, and converts this vector (hereinafter referred to as a new content vector) into a category vector obtained by vectorizing each category, or already into each category. By comparing with one or more classified existing content vectors (hereinafter referred to as existing content vectors), it is determined which category the new content vector falls into or does not fall into any category.

For example, when the new content CT is a text, the learning processing unit 114 divides the text into words (morphemes) by using a morphological analysis method such as TF (Term Frequency) -IDF (Inverse Document Frequency) or Okapi BM25b. , A multidimensional vector whose elements are statistics such as the frequency of appearance of the divided words may be derived as a new content vector. Further, when a plurality of new content CTs exist, the learning processing unit 114 may have searched a matrix having each of the plurality of new content CTs as an element and one or more contents among the plurality of new content CTs. A multidimensional vector obtained by multiplying a matrix having each of a plurality of users as an element may be derived as a new content vector. At this time, the learning processing unit 114 may use a statistic such as the number of views of the new content CT by the user as the weight of each user's element included in the multidimensional vector.

Further, the learning processing unit 114 may reduce the number of dimensions of the multidimensional vector derived as a new content vector by using a dimensional compression method such as Principal Component Analysis (PCA) or AutoEncoder.

Further, the learning processing unit 114 uses a non-negative matrix factor as a multidimensional vector (matrix) obtained by multiplying a word matrix and a content matrix, or a multidimensional vector (matrix) obtained by multiplying a content matrix and a user matrix. It may be changed to a low rank (less number of dimensions) matrix by matrix decomposition such as decomposition (Non-negative Matrix Factorization: NMF).

Further, when the new content CT is a text, the learning processing unit 114 uses algorithms such as word2vec and doc2vec to learn a task of predicting a word appearing before and after a word of interest in the text, thereby using the new content. A feature vector as each element of the vector may be obtained as a secondary.

Further, the learning processing unit 114 learns (determines) the parameters of the value function referred to when classifying the new content into categories by reinforcement learning. For example, the learning processing unit 114 redetermines the parameters of the value function based on the reward obtained when the new content is classified into categories by using a reinforcement learning method called Q-learning. The details of reinforcement learning will be described later.

[Category classification processing]
Hereinafter, the flow of the category classification processing by the learning processing unit 114 will be described according to the flowchart. FIG. 4 is a flowchart showing a flow of a series of processes by the learning process unit 114. The process of this flowchart is executed, for example, when a new content CT is acquired by the acquisition unit 112. In the following description of the flowchart, as an example, the content CT is assumed to be an "article", and the category of the article is assumed to be a "topic".

First, when the learning processing unit 114 acquires a new article (an example of a new content CT) at a certain time t by the acquisition unit 112, the new article vector x ₀ is a vector of the new article based on the equation (1). And, the similarity with the topic vector x _k obtained by vectorizing each topic k of a plurality of existing topics is derived, and among the similarity derived for each topic vector x _k , the largest similarity is the threshold value s or more. It is determined whether or not there is (S100). In the formula, T represents transposition.

The topic vector x _k is, for example, a vector obtained by averaging a set of existing article vectors in which one or more articles (hereinafter referred to as existing articles) already classified into the topic k are vectorized. At this time, a weight is given to each existing article vector, and the topic vector x _k may be a weighted average of a plurality of existing article vectors. In this case, the weight may be set to increase according to the freshness of the article (short time from the time of submission), for example.

For example, the topic vector x _k , which is the set average of existing article vectors, is expressed by Eq. (2). K represents the total number of existing topics, and k represents the topic of interest among the K topics. For example, in the example of FIG. 3, since the topics A to D (an example of the category) exist, K is 4.

As shown in the equation (1), for example, the learning processing unit 114 selects one topic vector x _k of interest from the set {x _k } of K topic vectors, and transposes the new article vector x ₀ . , Derives the cosine similarity with the selected topic vector x _k . At this time, it is assumed that the new article vector x ₀ and the topic vector x _k are represented by, for example, a linear sum of the same basis (base vector).

The learning processing unit 114 selects one topic vector x _{k + 1} , which is different from the previously selected topic vector x _k , from the set of K topic vectors {x _k }, and again with the transposed new article vector x ₀ . Derivation of the cosine similarity of. By repeating this process K times, the learning processing unit 114 derives the cosine similarity with the new article vector x ₀ for each topic vector x _k . Then, the learning processing unit 114 determines whether or not the largest cosine similarity among the cosine similarity derived for each topic vector x _k is the threshold value s or more.

When the learning processing unit 114 determines that the maximum similarity (for example, cosine similarity) is equal to or higher than the threshold value s, the learning processing unit 114 determines that the new article is a follow-up report of an existing topic, and based on the equation (3), 1 to 1 to From a plurality of topics up to K, one topic corresponding to the existing topic vector x _k having the highest similarity to the new article vector x ₀ is selected (S102).

The k (-) in the equation (3) represents an existing topic corresponding to the existing topic vector x _k having the highest degree of similarity to the new article vector x ₀ . In addition, (-) represents a bar of a symbol.

Next, the learning processing unit 114 classifies the new article determined as a follow-up report of the existing topic into the existing topic, and based on the value function indicated by the value function information FX, the topic vector x of the selected existing topic k (-). Update _{k (−)} (S104).

For example, the learning processing unit 114 derives a weighted sum of the new article vector x ₀ and the topic vector x _{k (-)} obtained by vectorizing the existing topic k (-) based on the equation (4). Change the topic vector x _{k (-)} to a vector represented by the weighted sum of each vector.

Α (−) in the equation (4) represents the weight of the new article vector x ₀ , and (1-α (−)) represents the weight of the topic vector x _{k (−)} . The weight α (−) may be derived, for example, by the equation (5).

V ({x _k }) in equation (5) represents the value function of topic k. The value function is defined as, for example, V (x _k ; {x.}, T). For example, in the equation (5), the learning processing unit 114 sets the value derived by the value function of the topic k, that is, the weight α of the new article vector x ₀ that maximizes the value of the topic k, to the weight α of the equation (4). Determined as α (-).

FIG. 5 is a diagram schematically showing a process of updating a topic vector. For example, if the vector space is a two-dimensional space, the existing topics A to D are represented as circles. In the space inside the circle corresponding to each topic, it indicates that the similarity with the topic is equal to or higher than the threshold value s. For example, when a new article is included in the existing topic B among the existing topics A to D, the topic having the highest degree of similarity to the new article is the topic B. In this case, the topic B # in which the weighted sum of the new article and the topic B is represented as a topic vector becomes the new topic B. That is, the topic in which the new article is classified has the article vector of the article included in the deleted topic as a subspace containing the article vector of the article contained in the topic in the vector space represented by the basis vector. Shift from the contained subspace.

In this way, when the new article is similar to any of the existing topics, the learning processing unit 114 determines that the new article is an article of a topic related to the existing topic, and classifies this new article into the existing topic. .. Then, the learning processing unit 114 changes the topic vector of the existing topic to which the new article is classified, and updates the property of the existing topic. As a result, for example, when multiple articles that are common to a topic are acquired in quick succession, the articles acquired later are more likely to resemble the topics in which the articles acquired earlier are classified. The nature of the topic can be changed according to the trend). As a result, the accuracy of article classification can be improved.

On the other hand, when the learning processing unit 114 determines that the maximum similarity (for example, cosine similarity) is less than the threshold value s in the processing of S100, the new article is not a follow-up report of the existing topic but an article of the new topic. Is determined, and any one topic k'is selected from the set of all topics (K + 1) when the new article is set as a new topic (S106).

Next, the learning processing unit 114 excludes the selected topic k'from the set of all topics (K + 1), and combines the remaining plurality of topics excluding the topic k'(S108).

Next, the learning processing unit 114 derives the value of each of the combined plurality of topics based on the value function indicated by the value function information FX, and derives the sum of these values (S110). For example, the sum of the values of a topic may be expressed by equation (6).

Next, the learning processing unit 114 determines whether or not the sum of the values of the topics has been derived for all the combinations in the set of topics (K + 1) (S112), and the topic is still for all the combinations. If it is determined that the sum of values has not been derived, the process returns to S106 and a topic different from the previous one is selected as the topic k'.

On the other hand, when the learning processing unit 114 determines that the sum of the values of the topics has been derived for all the combinations, the learning processing unit 114 selects the topic k'excluded at the time of the combination of the topics having the maximum sum of the values of the topics (S114). ).

For example, the learning processing unit 114 may perform the above-mentioned processing from S106 to S114 based on the equation (7).

K (-) in the equation (7) is the sum of the values of a set of a plurality of topic vectors {x _k } in which the remaining topic vectors x _k are combined when the topic vector x _k'of a certain topic k'is excluded. When V {x _k } (see equation (6)) becomes maximum, it corresponds to the topic vector x _k'excluded when a plurality of topic vectors x _k are combined as a set {x _k } of the topic vectors. Represents the topic k'. Further, k'= 0 in the equation (7) represents the new article vector x ₀ .

Next, the learning processing unit 114 determines whether or not the topic k (−) is an existing topic (S116). For example, the learning processing unit 114 selects an existing topic as the topic k'from the set of all topics (K + 1) in the processing of S106, and further, this topic k'is the total value V {x _k }. If the topic k (-) was excluded when was maximized, it is determined that the topic k (-) is an existing topic.

On the other hand, the learning processing unit 114 selects a new article treated as one topic from the set of all topics (K + 1) as the topic k'in the processing of S106, and further, this topic k'is of value. If the topic k (-) was excluded when the sum V {x _k } was maximized, it is determined that the topic k (-) is not an existing topic but a new article.

When the learning processing unit 114 determines that the topic k (-) is an existing topic, that is, when any of the existing topics is excluded, at least the sum of the values of the set of topics including the new article becomes the maximum. If so, the new article is replaced with the existing topic selected as the topic k (−) based on the formula (8) (S118).

For example, the learning processing unit 114 redefines the topic vector x _{k (-) of the existing topic selected as the topic k (-)} to the new article vector x ₀ , as shown in the equation (8). Swap articles with existing topics.

FIG. 6 is a diagram schematically showing a process of replacing a new article with an existing topic. For example, when the similarity between the new article and each existing topic is less than the threshold value s, the new article is not included in any of the topics as shown in the illustrated example. At this time, for example, when the total value of the set of topics including new articles excluding the existing topic C becomes the maximum, the learning processing unit 114 discards (deletes) the existing topic C as shown in the illustrated example. ) And add the new article as a new topic C #.

On the other hand, when the learning processing unit 114 determines that the topic k (-) is not an existing topic but a new article, that is, when the new article is excluded, the sum of the values of the set of topics including only the existing topic. When is the maximum, the new article is discarded without being classified into any of the existing topics (S120). This ends the processing of this flowchart.

FIG. 7 is a diagram schematically showing a process of not classifying a new article into an existing topic. For example, when the similarity between the new article and each existing topic is less than the threshold value s, the new article is not included in any of the existing topics as shown in the illustrated example. At this time, for example, when the sum of the values of the sets of the existing topics A to D excluding the new articles is maximized, the learning processing unit 114 puts the new articles into any existing topic as shown in the illustrated example. Discard this new article without classifying it.

[Reinforcement learning of value function]
Hereinafter, reinforcement learning of the value function by the learning processing unit 114 will be described. For example, the learning processing unit 114 learns the parameters of the value function V described above by using the Q-learning method. It is assumed that the value function V described below is realized by, for example, a neural network having an input layer, one or more intermediate layers (hidden layers), and an output layer. The value function information FX includes various information such as connection information of neurons (units) constituting the neural network and connection coefficients (model parameters of the neural network described later). For example, the number of units in the input layer is 6 or more, the number of units in the output layer is 1, the number of intermediate layers is 3 or more, the number of units in each intermediate layer is 128, and all the units in each layer of the intermediate layer. They shall be connected to each other. Further, the activation function of the unit of each intermediate layer may be a rectified linear function (ReLU function), a sigmoid function, a step function, or other functions.

For example, the input layer of a neural network is input with some or all of the following multiple elements, and the output layer outputs a value indicating the value of an existing topic according to the input (these). It is composed. That is, the value function may be a function in which some or all of the following plurality of elements are variables.

Identity 1: Weighted average value of the vector of articles contained in the existing topic.
Identity 2: Weighted average of all article vectors.
Identity 3: The degree of similarity with the upper predetermined number of high values among the degree of similarity with existing topics.
Identity 4: Time elapsed since the topic was created.
Identity 5: The time elapsed since the last new article was classified into an existing topic.
Identity 6: The number of articles included in the existing topic.

For example, element 1 represents a vector (topic vector) that serves as a reference for an existing topic, and is a weighted average of each article vector of one or more articles included in the existing topic as shown in equation (4). expressed.

For example, identity 2 is represented by a weighted average of all article vectors, regardless of topic. At this time, the earlier the acquisition time, the larger the weight may be given to each article vector. For example, when the article B is acquired after the article A is acquired, the weight of the article vector of the article B may be larger than the weight of the article vector of the article A. This allows you to compare trends in articles that are currently being submitted with trends in existing topics.

For example, the identity 3 is represented by a predetermined number of similarities, such as the top five, among the similarities of the plurality of existing topics to the new article. The learning processing unit 114 trains the neural network so that the output value increases as the degree of similarity increases. By inputting only the upper predetermined number of similarities without inputting all the similarities as the element 3, the same neural network can be continuously used even if the number K of the existing topic fluctuates.

For example, the elements 4 and 5 are represented by elapsed time, and the learning processing unit 114 updates the parameters of the neural network by Q-learning, which will be described later, so that the longer these elapsed times are, the smaller the output value becomes. Train a neural network.

For example, the element 6 is represented by the number of articles, and the learning processing unit 114 updates the parameters of the neural network by Q-learning so that the larger the number of articles, the greater the output value. ..

The learning processing unit 114 determines the parameter θ of the neural network based on the equation (9) to which the Q-learning method is applied. Equation (9) is a model of the problem targeted for reinforcement learning. The parameter θ represents a model parameter of the neural network at a certain time _ti . The model parameters include, for example, in the middle layer of the neural network, the weight given to the output data when the data is output from the unit of one layer to the unit of the deeper layer. Further, the model parameters of the neural network may include a bias component peculiar to each layer.

The first term r (k, ti _{+ 1} ) in the equation (9) is added when a new article is added to the existing topic _k at a future time ti _{+ 1} that is ahead of a certain time ti + 1. It represents a reward. The time _ti and the time ti _{+ 1} represent the time when the new article is submitted to the posting site or the like, or the time when the new article submitted to the posting site or the like is acquired by the acquisition unit 112. If no new article is added (classified) to the existing topic k at time ti _{+ 1} , the term of r (k, ti _{+ 1} ) becomes 0. The reward r may be a predetermined constant such as 1 or 0, or may be some actually obtained measured value such as the number of PVs.

Further, V _θ '(x _k ; {x.}, Ti _{+ 1} ) of the second term in the equation (9) represents the value (predicted value) of the existing topic k at the time ti _{+ 1} . θ'represents the model parameters of the neural network at time ti _{+ 1} . The second term V _θ'is an example of the “first value”.

Further, the second term e ^ γ (ti _{+ 1 −} ti) represents a coefficient term (discount rate) that reduces the value of the existing topic _k at the time ti _{+ 1} . γ is a negative constant, and the value of _e decreases as the acquisition interval (ti _{+ 1 −} ti) of new articles increases. That is, the coefficient term e of the second term further reduces the value of the existing topic _k at the future time ti _{+ 1} as the acquisition interval (ti _{+ 1 −} ti) of the new article increases.

The V _θ (x _k ; {x. _} , Ti) of the third term in the equation (9) represents the value (predicted value) of the existing topic _k at the time ti. θ represents the model parameter of the neural network at time _ti . Since the model parameters of the neural network are updated by error backpropagation over time, θ'and θ described above represent parameters that are different from each other. That is, the third term V _θ represents the value of the existing topic k output by the neural network adjusted by the parameter θ before being updated to the parameter θ ′ used at time ti _{+ 1} . V _θ in the third term is an example of “second value”.

For example, assuming that the current time is time _ti + 1, it is predicted that the learning processing unit 114 classifies new articles into existing topics according to the processing of the above-mentioned flowchart at the future time ti _{+ 1} . In this case, the learning processing unit 114 refers to the equation (9), and determines the reward at the predicted future time ti _{+ 1} (first term) and the value of the existing topic k at the predicted future time ti _{+ 1} (second term). Is derived. At this time, the learning processing unit 114 reduces the value of the existing topic _k at the predicted future time ti _{+ 1 according} to the time difference from the time ti + 1. Then, the learning processing unit 114 searches for the parameter θ that minimizes the square of the value obtained by subtracting the value (third term) of the existing topic _k at the current time ti from the derived sum.

The learning processing unit 114 adjusts the weight of the neural network by error back propagation so that the parameter θ obtained as a result of the search is obtained. As a result, the value of the existing topic _k predicted at the current time point ti is modified so that the difference in the value of the existing topic k at different time points becomes small. For example, the learning processing unit 114 updates the parameter θ of the value function V by reinforcement learning while feeding back the classification result of the content each time the processing of the above-mentioned flowchart is executed. By such processing, the value function V whose parameter θ is determined by reinforcement learning is referred to at the time of category classification, so that the content classification accuracy can be improved.

FIG. 8 is a diagram showing an example of the result of the reward obtained by the reinforcement learning of the present embodiment. The vertical axis in the figure represents the value (score) of the reward r in the equation (9), and the horizontal axis represents the time during which reinforcement learning is continued. Further, LN1 in the figure represents a change in the total sum of rewards r of each topic obtained by reinforcement learning of the present embodiment, and LN2 is a reward r of each topic when new articles are preferentially left. It represents a change in sum. In addition, LN3 in the figure represents the change in the sum of rewards r for each topic when the topic with the longest time elapsed since the last article was classified (the oldest topic) and the new article are replaced. LN4 represents the change in the sum of the rewards r of each topic when the topic with the smallest number of articles is replaced with the new article. Further, LN5 in the figure represents a change in the sum of rewards r of each topic when the topic that does not cover the space area with other topics is preferentially left in the vector space. The sum of each of these rewards r is normalized by a predetermined value. The predetermined value is the sum of the rewards r of each topic when a new article is acquired and one of the randomly selected topics out of the K topics is replaced with the new article.

As shown in the illustrated example, when a certain amount of time has passed, LN1 has a higher score than all of LN2 to 5, and the reinforcement learning in this embodiment is compared with the heuristic method of LN2 to 5. Can actually get more rewards.

According to the embodiment described above, when the acquisition unit 112 for acquiring new content (for example, a new article) and the acquisition unit 112 acquire new content, the parameters are learned by reinforcement learning based on the value function. It is most valuable by having a learning processing unit 114 that determines whether to select the category to which the new content is classified from among a plurality of existing categories or to generate a new category and classify it into that category. It can be preferentially deleted from categories (for example, topics) that are expected to become smaller. As a result, for example, it is possible to more efficiently sort out the contents that should be kept and the contents that should not be kept, rather than adopting a heuristic method in which the oldest category is deleted in order.

Further, according to the above-described embodiment, when a new article is classified into an existing topic, the topic vector representing the existing topic is changed, so that the nature of the topic can be changed according to the trend of the submitted article. can. As a result, the accuracy of article classification can be improved. In addition, when the content is a product put up for auction instead of an article, the nature of the category can be changed according to the selling trend.

Further, according to the above-described embodiment, a new category can be created in order to replace the topic whose value is predicted to be the smallest with the new content. For example, the general category classification method determines the content classification destination from the categories whose properties are defined in advance, and unknown content that does not correspond to any of the currently prepared categories was obtained. If so, the content does not fall into any category. On the other hand, in the present embodiment, since the category with low value is replaced with the new content, when unknown content whose property is not defined is obtained, the content does not correspond to the existing category. Based on the degree of similarity with the new content, it is possible to collect contents with similar properties and generate a new category.

Further, according to the above-described embodiment, as the acquisition interval (ti _{+ 1} - _ti ) of new articles increases, the value of the existing topic at the future time ti _{+ 1} decreases, so that the learning accuracy can be further improved. ..

In the case of general reinforcement learning, the discount rate γ is often constant regardless of time. Using a constant discount rate γ, for example, if an article is frequently trafficked, the future value of an existing topic, even though little time has passed since the last article was trafficked. It tends to drop significantly. On the other hand, in the present embodiment, since the discount rate γ is changed according to the acquisition interval of new articles, the future value of the existing topic can be predicted accurately.

<Hardware configuration>
The classification device 100 of the above-described embodiment is realized by, for example, a hardware configuration as shown in FIG. FIG. 9 is a diagram showing an example of the hardware configuration of the classification device 100 of the embodiment.

The classification device 100 includes NIC100-1, CPU100-2, RAM100-3, ROM100-4, a secondary storage device 100-5 such as a flash memory or an HDD, and a drive device 100-6 by an internal bus or a dedicated communication line. It has a structure that is connected to each other. A portable storage medium such as an optical disk is mounted on the drive device 100-6. A program stored in a portable storage medium mounted on the secondary storage device 100-5 or the drive device 100-6 is expanded in the RAM 100-3 by a DMA controller (not shown) or the like, and executed by the CPU 100-2. As a result, the control unit 110 is realized. The program referred to by the control unit 110 may be downloaded from another device via the network NW.

Although the embodiments for carrying out the present invention have been described above using the embodiments, the present invention is not limited to these embodiments, and various modifications and substitutions are made without departing from the gist of the present invention. Can be added.

1 ... Content classification system, 10 ... Information providing device, 100 ... Classification device, 102 ... Communication unit, 110 ... Control unit, 112 ... Acquisition unit, 114 ... Learning processing unit, 130 ... Storage unit, CT ... Content, FX ... Value Function information

Claims (17)

The acquisition department that acquires new content, and
When the new content is acquired by the acquisition unit, the category to be classified into the new content is selected from a plurality of existing categories by reinforcement learning, or a new category is generated and generated. The learning processing unit that decides whether to classify into categories,
Equipped with
The learning processing unit
It is determined whether or not the largest similarity between the feature amount of each of the existing plurality of categories and the feature amount of the new content is equal to or higher than the threshold value.
When it is determined that the similarity is equal to or higher than the threshold value, the existing category having the highest similarity between the features with the new content is selected from the existing plurality of categories as the category to which the new content is classified. Select as
When classifying the new content into the selected existing category, the value when the existing category to which the new content is classified is left unerased based on the value function for evaluating the state of the category is Update the features of the existing category to the maximum,
Sorting device. The acquisition department that acquires new content, and
When the new content is acquired by the acquisition unit, the category to be classified into the new content is selected from a plurality of existing categories by reinforcement learning, or a new category is generated and generated. The learning processing unit that decides whether to classify into categories,
Equipped with
The learning processing unit
It is determined whether or not the largest similarity between the feature amount of each of the existing plurality of categories and the feature amount of the new content is equal to or higher than the threshold value.
When it is determined that the similarity is less than the threshold value, a plurality of categories are combined except for one category from the existing plurality of categories.
Based on the value function that evaluates the state of the categories, the sum of the values when each of the multiple categories combined is left unerased is derived.
Of the sum of values derived for each combination of the plurality of categories, one category excluded when the category with the largest sum of values is combined is determined as the category to be eliminated.
A new category is generated to classify the new content into the generated category, and one category determined to be deleted is deleted from the existing plurality of categories .
Sorting device. The learning processing unit further
Include the new content as one category in multiple categories to be combined.
Combining a plurality of categories excluding one category from the set of the existing plurality of categories and the new content as one category,
If one of the categories excluded from the combination of the categories having the largest sum of values is the new content, it is decided to delete the new content.
The classification device according to claim 2 . The learning processing unit learns the parameters of the value function by the reinforcement learning based on the classification result of the new content.
The classification device according to any one of claims 1 to 3 . The learning processing unit receives a reward given when the new content is classified into an existing category at a certain first time, and from a second time before the first time to the first time. In the process, when the new content is classified into an existing category, the sum with the first value obtained by the value function is derived.
The value function is based on the difference between the sum of the reward and the first value and the second value obtained by the value function when the new content is classified into an existing category at the second time. Learn the parameters,
The classification device according to claim 4 . The learning processing unit learns the parameters of the value function so that the value obtained by squaring the difference is minimized.
The classification device according to claim 5 . The first time and the second time are times when the new content is acquired by the acquisition unit.
The learning processing unit reduces the first value based on the time difference from the second time to the first time.
The classification device according to claim 5 or 6 . The value function is a category when at least a weighted average of the features of one or more existing contents included in the category selected as the classification destination of the new contents and the features of the new contents is input. Deriving value,
The classification device according to any one of claims 4 to 7 . The value function derives the value of a category when at least a weighted average of the features of one or more existing contents and the features of the new content is entered.
The classification device according to any one of claims 4 to 8 . The value function derives the value of a category when at least the similarity between each of the existing categories and the new content is entered.
The classification device according to any one of claims 4 to 9 . The value function derives the value of a category when at least the time elapsed since each of the existing categories was generated is entered.
The classification device according to any one of claims 4 to 10 . The value function derives the value of a category when at least the time elapsed since the content was classified into each of the existing categories is entered.
The classification device according to any one of claims 4 to 11 . The value function derives the value of a category when at least the number of contents contained in each of the existing plurality of categories is input.
The classification device according to any one of claims 4 to 12 . The computer
Get new content,
When the new content is acquired, whether the category to be classified into the new content is selected from a plurality of existing categories or a new category is generated and classified into the generated category by reinforcement learning. Decide and
In the process of determining, the computer
It is determined whether or not the largest similarity between the feature amount of each of the existing plurality of categories and the feature amount of the new content is equal to or higher than the threshold value.
When it is determined that the similarity is equal to or higher than the threshold value, the existing category having the highest similarity between the features with the new content is selected from the existing plurality of categories as the category to which the new content is classified. Select as
When classifying the new content into the selected existing category, the value when the existing category to which the new content is classified is left unerased based on the value function for evaluating the state of the category is Update the features of the existing category to the maximum,
Classification method. On the computer
Get new content,
When the new content is acquired, whether the category to be classified into the new content is selected from a plurality of existing categories by reinforcement learning, or whether a new category is generated and classified into the generated category. Let me decide
In the process of making the determination, the computer
Among the similarities between the features of each of the existing plurality of categories and the features of the new content, it is determined whether or not the largest similarity is equal to or higher than the threshold value.
When it is determined that the similarity is equal to or higher than the threshold value, the existing category having the highest similarity between the features with the new content is selected from the existing plurality of categories as the category to which the new content is classified. Let me choose as
When classifying the new content into the selected existing category, the value when the existing category to which the new content is classified is left unerased based on the value function that evaluates the state of the category. Update the feature quantity of the existing category so as to be the maximum.
program. The computer
Get new content,
When the new content is acquired, whether the category to be classified into the new content is selected from a plurality of existing categories or a new category is generated and classified into the generated category by reinforcement learning. Decide and
In the process of determining, the computer
It is determined whether or not the largest similarity between the feature amount of each of the existing plurality of categories and the feature amount of the new content is equal to or higher than the threshold value.
When it is determined that the similarity is less than the threshold value, a plurality of categories are combined except for one category from the existing plurality of categories.
Based on the value function that evaluates the state of the categories, the sum of the values when each of the multiple categories combined is left unerased is derived.
Of the sum of values derived for each combination of the plurality of categories, one category excluded when the category with the largest sum of values is combined is determined as the category to be eliminated.
A new category is generated to classify the new content into the generated category, and one category determined to be deleted is deleted from the existing plurality of categories .
Classification method. On the computer
Get new content,
When the new content is acquired, whether the category to be classified into the new content is selected from a plurality of existing categories by reinforcement learning, or whether a new category is generated and classified into the generated category. Let me decide
In the process of making the determination, the computer
Among the similarities between the feature amounts of each of the existing plurality of categories and the feature amounts of the new content, it is determined whether or not the largest similarity is equal to or higher than the threshold value.
When it is determined that the similarity is less than the threshold value, a plurality of categories are combined by excluding one category from the existing plurality of categories.
Based on the value function that evaluates the state of the category, the sum of the values when each of the plurality of combined categories is left unerased is derived.
Of the sum of values derived for each combination of the plurality of categories, one category excluded when the category with the largest sum of values is combined is determined as the category to be eliminated.
A new category is generated to classify the new content into the generated category, and one category determined to be deleted is deleted from the existing plurality of categories .
program.

Images