WO2022244339A1

WO2022244339A1 - Content-providing terminal and content-providing program used in same

Info

Publication number: WO2022244339A1
Application number: PCT/JP2022/004823
Authority: WO
Inventors: 智幸河添
Original assignee: シャープ株式会社
Priority date: 2021-05-21
Filing date: 2022-02-08
Publication date: 2022-11-24

Abstract

According to the present invention, a user is effectively provided with content that the user is inferred to be interested in while also moderating increases in the user burden of manipulating an input unit for expressing whether the user is interested in the content.　Provided is a content-providing terminal comprising a control unit configured to receive a content vector expressing a distributed representation of content, an interested vector expressing a distributed representation of categories that the user is interested in, and an uninterested vector expressing a distributed representation of categories that the user is not interested in, wherein the control unit infers whether the user is interested in the content on the basis of at least one of a similarity between the content vector and the interested vector or a similarity between the content vector and the uninterested vector.

Description

Content providing terminal and content providing program used therefor

The present disclosure relates to a content providing terminal and a content providing program used therefor. This application claims priority based on Japanese Patent Application No. 2021-085736 filed in Japan on May 21, 2021, the content of which is incorporated herein.

In recent years, for example, as described in Patent Document 1 below, services for distributing content from a distribution server to a user's communication terminal have become widespread. In this service, contents that are presumed to be of interest to users are selected from a large number of contents based on user vectors that represent users in a distributed manner, and the selected contents are distributed to the user's communication terminals.

JP 2018-088051 A

Using the above technology, content such as news containing document data may be distributed to the user's communication terminal. In this case, a large number of news items are distributed from the distribution server to the user's communication terminal every day. Therefore, if the user is asked to perform an operation to select whether each of a large number of distributed news is "interesting news" or "uninteresting news", the user's operation of the communication terminal becomes a heavy burden. Become. Therefore, for example, the following method is being considered.

The communication terminal that receives the content first displays only part of the content information, such as the title of the news, or information that recommends the content to the user, such as an overview of the news, on the display unit. After that, when the user operates the input unit for viewing the entire content, for example, the details of the news, the communication terminal determines that the news is "interesting news". . Also, if the user does not operate the input section for viewing the details of the news on the communication terminal, the communication terminal determines that the news is "uninteresting news". According to this method, it is possible to determine whether distributed news is "interesting news" or "uninteresting news" without increasing the user's operational burden.

However, in this method, either the user is not interested in the news and therefore does not see the details of the news, or the user is interested in the news but already knows the news and therefore does not know the news. I can't tell the difference without looking at the details. Therefore, all news that the user does not look at in detail is determined to be "uninteresting news".

As can be seen from the above, in the above technology, for example, when the user receives partial information or recommended information about the content, but does not receive the entire information about the content, the user receives information about the content. It is not possible to estimate whether or not you are interested in Therefore, it is not possible to effectively provide the user with content that is presumed to be of interest to the user.

The present disclosure has been made in view of the above problems. The purpose of the present disclosure is to suppress an increase in the user's operational burden on the input unit for indicating whether or not the user is interested in the content, while suppressing the content that the user is presumed to be interested in. is effectively provided to the user. Specifically, an object of the present disclosure is to provide a content providing terminal and a content providing program capable of effectively providing the aforementioned content.

One aspect of the content providing terminal of the present disclosure includes a content vector distributed representation of content, an interested vector distributed representation of a field in which the user is interested, and a distributed representation of a field of interest to the user. a control unit capable of receiving a non-interest vector, wherein the control unit receives at least one of a similarity between the content vector and the interested vector and a similarity between the content vector and the non-interest vector Based on this, it is estimated whether the user is interested in the content.

According to one aspect of the content providing program of the present disclosure, a computer is provided with a content vector distributed representation of content, an interested vector distributed representation of a field in which a user is interested, and a distributed representation of a field in which the user is not interested. a computer-readable content providing program for operating as a control unit capable of receiving expressed uninterested vectors, wherein the control unit controls similarity between the content vector and the interested vector, and the content vector and the uninteresting vector, it is estimated whether or not the user is interested in the content.

1 is a diagram showing the overall configuration of a content providing system according to Embodiment 1; FIG. 2 is a diagram showing an internal configuration of a content distribution server according to Embodiment 1; FIG. 4 is a diagram showing an example of news information stored in a content storage unit of the content distribution server of Embodiment 1; FIG. 2 is a diagram showing the internal configuration of the content providing terminal according to Embodiment 1; FIG. 4 is a diagram showing an example of news information stored in a reaction history storage unit and a no-response history storage unit according to Embodiment 1. FIG. FIG. 3 is a diagram showing an example of interesting vectors and uninteresting vectors stored in an interest vector storage unit according to Embodiment 1, and information associated therewith; 4 is a flowchart showing control processing executed by the content distribution server of Embodiment 1; 4 is a first flowchart showing control processing executed by the content providing terminal according to Embodiment 1; Interested vectors and uninterested vectors stored in the interested vector storage unit of the first embodiment, the importance of the interested vectors and the importance of the uninterested vectors, and the similarity of the interested vectors and the similarity of the uninterested vectors is a diagram for explaining the relationship between. 9 is a second flowchart showing control processing executed by the content providing terminal according to Embodiment 1; 4 shows an example of a screen displayed on the display unit of the content providing terminal according to Embodiment 1, which asks the user whether or not to view the details of the news recommended by the content recommendation unit. 9 is a third flowchart showing control processing executed by the content providing terminal according to Embodiment 1; 10 is a fourth flowchart showing control processing executed by the content providing terminal according to Embodiment 1; An example of a screen displayed on the display unit of the content providing terminal according to Embodiment 1, which is a question asking the user whether he wishes to display a word and news related to that word on the display unit in the future. FIG. 10 is a diagram showing an example of a screen displaying information; FIG. 10 is a diagram showing an internal configuration of a content providing terminal according to Embodiment 2; FIG. FIG. 11 is a first flowchart showing control processing executed by the content providing terminal according to Embodiment 2; FIG. FIG. 10 is a second flowchart showing control processing executed by the content providing terminal according to the second embodiment; FIG. 13 is a third flow chart showing control processing executed by the content providing terminal according to the second embodiment;

A content providing terminal according to an embodiment and a content providing program used therein will be described below with reference to the drawings. In the drawings, the same or equivalent elements are denoted by the same reference numerals, and redundant description of the same or equivalent elements will not be repeated.

(Embodiment 1)
A content providing terminal according to the first embodiment and a content providing program used therein will be described with reference to FIGS. 1 to 14. FIG. In the content providing terminal of the present embodiment, the content providing terminal possessed by the user recommends to the user content that is presumed to be of interest to the user from among the content acquired from the content distribution server.

FIG. 1 shows the overall configuration of a content providing system 100 according to this embodiment.

As shown in FIG. 1, the content providing system 100 includes a content distribution server 1 and a content providing terminal 2. The content distribution server 1 and the content providing terminal 2 are connected via a telecommunications information network such as the Internet. In other words, the content distribution server 1 and the content providing terminal 2 are communicably connected by at least one of wireless and wired communication lines.

The content distribution server 1 is a server that distributes content that can be represented in a distributed manner using vectors. The content is assumed to be digitized document data or the like. Also, the content distribution server 1 is, for example, a device for providing content to consumers who subscribe to news articles by businesses that produce news articles.

The content providing terminal 2 is a communication terminal that receives a large amount of content from the content distribution server 1 and selectively provides the user with only content that is presumed to be of interest to the user. In this embodiment, the content providing terminal 2 is a mobile terminal called a smart phone, which has functions of receiving, outputting, and inputting electronic information. It is assumed that the content providing terminal 2 is owned by a consumer who receives content from a business.

However, the content providing terminal 2 may be an installation-type communication terminal such as a personal computer that has the function of transmitting, receiving, outputting, and inputting electronic information. In the case of a personal computer, the parts having the function of inputting and outputting information, such as the keyboard, mouse, display, and speakers, are realized by parts different from the parts having the function of transmitting and receiving information, such as the microcomputer. may In this case, a CPU having a function of transmitting and receiving constitutes the content providing terminal 2 .

The content of this embodiment is a news article expressed as an electronic document. Note that an article is a document in which words composed of letters are combined to describe phenomena, existence, situations, and the like, and is a sentence for conveying a matter.

However, if the content includes electronic document information that can express word data using vectors, electronic books, electronic advertisements of products or services, or electronic music lyrics etc., can be anything.

The content delivery server 1 and the content providing terminal 2 for providing users with news articles (hereinafter simply referred to as "news") as an example of content will be described below.

FIG. 2 shows the internal configuration of the content distribution server 1.

As shown in FIG. 2, the content distribution server 1 includes a content acquisition unit 11, a content vector generation unit 12, a content storage unit 13, and a content distribution unit 14.

The content acquisition unit 11 acquires electronic data of news articles, which are an example of content, from content sources on the Internet. The content vector generation unit 12 generates, from the electronic data of news articles acquired by the content acquisition unit 11, vectors representing the news in a distributed manner. The content storage unit 13 associates and stores the news acquired by the content acquisition unit 11 and the vector representing the news generated by the content vector generation unit 12 in a distributed manner, that is, the content vector. The content distribution unit 14 transmits news and distributed representation vectors corresponding to the news to the content providing terminal 2 .

3 shows an example of news information stored in the content storage unit 13 of the content distribution server 1. FIG.

As shown in FIG. 3, the content storage unit 13 stores IDs (identifiers) of news, dates and times when news is arranged, titles of news, texts of news, and vectors representing distributed news. The title of the news is "partial information of contents". Part of the content information is the title and, for example, the first few pages of an electronic book. Also, hereinafter, the text of the news may be referred to as "details of the news" or "entire information of the content". Data stored in the content storage unit 13 is not limited to that shown in FIG. 3, and may include other data such as image data, audio data, and video data. Note that the distributed representation vector is a representation of a word by a matrix of numbers, such as [0.05, −0.02, −0.09, . . . ].

4 shows the internal configuration of the content providing terminal 2. FIG.

As shown in FIG. 4, the content providing terminal 2 includes a content acquisition unit 21, an interest estimation unit 22, a recommended content storage unit 23, a content output control unit 24, and an output unit 2OUT.

The content acquisition unit 21 acquires news from the content distribution server 1. The content acquisition unit 21 is a radio wave reception unit such as an antenna, or an electric signal reception unit such as a connector to which a cable can be electrically connected. The content acquisition unit 21 may be a part of the content providing terminal 2, or may be an external component. The news acquired by the content acquisition unit 21 is accompanied by a news vector as an example of a content vector in which content is represented in a distributed manner. The interest estimation unit 22 estimates whether or not the user is interested in the news acquired by the content acquisition unit 21 . The interest estimation unit 22 transmits news estimated to be of interest to the user to the recommended content storage unit 23 as recommended content.

In this specification, the term “estimation” refers to the determination result that the user is interested in the content until the user denies the determination result of whether or not the user is interested in the content. shall mean to assume that the result of

The recommended content storage unit 23 stores the news estimated by the interest estimation unit 22 to be of interest to the user and related information. Note that the recommended content storage unit 23 may be an external storage unit provided separately from the content providing terminal 2 . The recommended content storage unit 23 stores news in the same format as the news stored in the content storage unit 13 in the content distribution server 1 (FIG. 3). That is, the recommended content storage unit 23 stores news in a format accompanied by at least content vectors, ie, news vectors.

The content output control unit 24 controls the output unit 2OUT that outputs content. Specifically, the content output control unit 24 outputs the news stored in the recommended content storage unit 23 to the output unit 2OUT. The content output control unit 24 also receives the user's answer information from the input unit 2IN based on the user's operation of the input unit 2IN. The details of this reply information will be described later. In addition, the content output control unit 24 transmits to the vector generation unit 25 the response information received from the input unit 2IN and the content vector representing the news received from the recommended content storage unit 23 in a distributed manner.

The output unit 2 OUT outputs news as content to the user. The output unit 2OUT includes a display unit having a liquid crystal panel or the like and a sound generation unit having a speaker or the like. The output unit 2OUT provides the user with news recommended by the interest estimating unit 22 by displaying the recommended news on the display unit and by emitting a voice reading out the recommended news from a speaker.

As shown in FIG. 4, the content providing terminal 2 includes an input unit 2IN, a vector generation unit 25, an interest vector storage unit 26, a reaction history storage unit 27, a no-response history storage unit 28, and a word vector storage unit 29. ing.

The input section 2IN is for the user to express his/her intention. The input unit 2IN includes an operation unit having a mouse, keyboard, touch panel, etc., and a voice input unit having a voice recognition function. The input unit 2IN recognizes the user's operation on the operation unit and the voice input unit recognizes the user's voice, thereby outputting answer information indicating whether or not the user is interested in news to the content output control unit. 24.

The vector generation unit 25 generates a variance representation vector for estimating whether or not the user is interested in news based on the user's content viewing history. Specifically, the vector generation unit 25 generates a vector representing a field in which the user is interested in a distributed representation (hereinafter referred to as an "interested vector") and a vector representing a field in which the user is not interested in a distributed representation ( hereinafter referred to as a "no interest vector"). The vector-of-interest storage unit 26 stores the vectors of interest and the vectors of no interest stored by the vector generation unit 25 . Interest and non-interest vectors are described in detail later. The vector generation unit 25 uses the vectors stored in the reaction history storage unit 27 and the no-response history storage unit 28 to generate interested vectors and uninterested vectors.

The reaction history storage unit 27 stores the history of the user's viewing of news details. More specifically, the reaction history storage unit 27 stores a news vector distributedly expressing the news for which the user has viewed the details of the news, that is, a content vector distributedly expressing the content for which the overall information is provided (hereinafter referred to as “reaction ) is memorized.

The no-response history storage unit 28 stores a history in which the user did not see the details of the news. Specifically, the no-response history storage unit 28 stores a news vector distributedly representing news for which the user did not see the details of the news, i.e., a content vector distributedly representing the content for which the user did not receive the provision of overall information (hereinafter referred to as a content vector). , say “no response vector”).

The word vector storage unit 29 stores words and vectors representing the words in a distributed manner (hereinafter referred to as "word vectors"). The word vector storage unit 29 stores a large number of words and a large number of word vectors in one-to-one correspondence.

In this embodiment, the interest vector storage unit 26, the reaction history storage unit 27, and the no-response history storage unit 28 all store distributed representation vectors. Assume that you do not remember yourself. The word vector storage unit 29 stores combinations of word vectors and words. In this specification, the interest vector storage unit 26, the reaction history storage unit 27, the no-response history storage unit 28, and the word vector storage unit 29 store at least vectors. It may also be referred to as "vector storage unit 20". However, each storage unit that constitutes the vector storage unit 20 may store the content itself corresponding to the distributed representation vector together with the distributed representation vector that is the content vector.

In the present embodiment, the content providing terminal 2 has a vector storage unit 20. FIG. However, the vector storage unit 20 may be a physically separate and independent storage device that is externally attached to the control unit C of the content providing terminal 2 and communicates with the control unit C. FIG. Also, the vector storage unit 20 may be provided in a server or the like that communicates with the control unit C of the content providing terminal 2 via a telecommunications information network.

In the present embodiment, the interest estimation unit 22, the recommended content storage unit 23, the content output control unit 24, and the vector generation unit 25 are called "control unit C". The control unit C controls the output unit 2OUT using the content vectors acquired by the content acquisition unit 21, the information input from the input unit 2IN, and the vectors stored in the vector storage unit 20. FIG.

The content providing terminal 2 of the present embodiment includes an output section 2OUT and an input section 2IN. However, the content providing terminal 2 may not have the output section 2OUT and the input section 2IN. The output unit 2 OUT may be an output device such as an external display or speaker that can be electrically connected to the content providing terminal 2 . The input unit 2 IN may be an input device such as an external mouse or touch panel that can be electrically connected to the content providing terminal 2 .

FIG. 5 shows a news vector as an example of a content vector stored in the reaction history storage unit 27 and the no-response history storage unit 28, and information accompanying the news vector. The reaction history storage unit 27 stores distributed representation vectors of news that the user has viewed in detail, in other words, content vectors that the user desires to be provided with overall content information, that is, "reaction vector". there is In addition, the no-response history storage unit 28 stores distributed representation vectors of news that the user did not look at in detail, in other words, content vectors that the user did not want to receive the overall information of the content, that is, "no response". vector” is stored.

As shown in FIG. 5, the reaction history storage unit 27 and the no-response history storage unit 28 each store the ID of the news, the date and time when the news was provided to the user, and the vector representing the news in a distributed manner. The data stored in each of the reaction history storage unit 27 and the no-response history storage unit 28 are not limited to those shown in FIG. You can

FIG. 6 shows an example of interested vectors and uninterested vectors stored in the interest vector storage unit 26. FIG.

As shown in FIG. 6, the interest vector storage unit 26 stores the ID of the vector that represents the content in a distributed manner, the date and time of vector generation, the presence or absence of interest, the importance of the vector (or the weight of the vector), and the vector itself. "Interested" in the interest column indicates that the corresponding vector is an interested vector, and "not interested" in the interest column indicates that the corresponding vector is not interested vector. The data stored in the interest vector storage unit 26 are not limited to those shown in FIG. 6, and may include other data such as image data, audio data, and video data.

FIG. 7 is a flowchart showing control processing executed by the content distribution server 1. FIG.

As shown in FIG. 7, in step S101, the content acquisition unit 11 of the content distribution server 1 acquires news as an example of content from content sources on the Internet. Next, in step S102, the content vector generation unit 12 of the content distribution server 1 generates a vector representing the news acquired by the content acquisition unit 11 in a distributed manner. Hereinafter, the news distributed representation vector generated by the content vector generation unit 12 is also referred to as a "content vector".

At this time, in the present embodiment, the content vector generation unit 12 generates all word vectors in which all words included in the news are represented in a distributed manner, and a weighted average vector of all the generated word vectors Calculate Thereby, the content vector generation unit 12 generates the calculated weighted average vector as a content vector representing the content in a distributed manner. The weight of the weighted average is the appearance frequency of words in the document data constituting the news, that is, the number of times the words are used. Therefore, according to this content vector generation method, a word with a higher appearance frequency contributes more to the generated content vector. However, the content vector generation method may be another method.

It should be noted that in the content vector generation method of the present embodiment, words are converted into distributed representation vectors using the word-vector conversion function. Specifically, words included in the received corpus are converted into word vectors as distributed representations. For example, the word-vector conversion function uses a tool (program) called word2vec to convert each of a plurality of words included in the corpus into a word vector. word2vec is a tool that uses a model called a neural network, and converts words included in a corpus into feature vectors (word vectors) that indicate the features of the words and outputs them.

In step S103, the content distribution server 1 stores the news and the distributed representation vector of the news, that is, the content vector, in the content storage unit 13 in a format that allows the mutual correspondence to be specified. The content distribution server 1 periodically repeats the processes of steps S101 to S103. As a result, several pieces of news and several distributed expression vectors representing the pieces of news in a distributed manner, that is, content vectors are stored in the content storage unit 13 in a one-to-one relationship.

The control executed by the content providing terminal 2 will be described with reference to FIGS. 8 to 14. FIG.

FIG. 8 is a first flowchart of control processing executed by the content providing terminal 2. FIG. More specifically, FIG. 8 is a flow chart for explaining control processing by the content providing terminal 2 for selecting news to be recommended to the user.

As shown in FIG. 8, in the content providing terminal 2, in step S201, the content acquisition unit 21 acquires news distributed from the content distribution unit 14 of the content distribution server 1 and distributed representation vectors corresponding to the news. do. The news acquired by the content acquisition unit 21 and its distributed representation vector are transmitted to the interest estimation unit 22 . Hereinafter, the distributed representation vector of news is the content vector.

Next, in step S<b>202 , the interest estimation unit 22 acquires interested vectors and uninterested vectors from the interest vector storage unit 26 . Details of the interested vector and the uninterested vector are described later. In step S203, the interest estimation unit 22 performs calculations to estimate whether the user will be interested in the news. The details of this calculation method will be described later.

In step S204, if it is estimated that the user will be interested in news (YES in S204), in step S205, the interest estimation unit 22 stores the news estimated to be of interest to the user as recommended content. Store in the unit 23 . On the other hand, if it is estimated that the user will not be interested in the news in step S204 (NO in S204), the interest estimation unit 22, in step S206, and delete its distributed representation vector. The details of the method of estimating whether or not the user is interested in news will be described later.

It should be noted that in order to estimate whether the user is interested in the content, as will be described later, it is necessary to obtain several answers from the user regarding whether or not they are interested in the content. However, immediately after the user starts using the content providing terminal 2, the content providing terminal 2 does not have any answer information of the user indicating whether or not the user is interested in the content. Therefore, in the present embodiment, it is presumed that the user will be interested in the content, and the content is recommended to the user until the user's response information indicating whether or not he/she is interested in the content is obtained several times. do.

FIG. 9 shows interested vectors and uninterested vectors stored in the interested vector storage unit of Embodiment 1, the importance of interested vectors and the importance of uninterested vectors, and the similarity of interested vectors and uninterested vectors. shows the similarity relationship between A method of estimating whether or not a user is interested in news will be described with reference to FIG. However, the estimation method described using FIG. 9 is an example, and other estimation methods may be used.

(1) Interested vectors are generated by the vector generator 25 as follows.

In this specification, "whether or not two distributed representation vectors are similar" is determined by "whether or not the degree of similarity of the two distributed representation vectors is equal to or greater than a threshold". Therefore, first, when the user views the details of a piece of news recommended by the user, the vector generation unit 25 calculates the degree of similarity between the particular piece of news and the news the user has viewed in detail in the past. In this embodiment, the degree of similarity between two news hectors is assumed to be the cosine degree of similarity between two content vectors representing distributed representations of two news items. The cosine similarity θ is represented by a value of −1≦θ≦1. In this specification, the term "threshold" is used in several types of similarity determinations, but the thresholds used in the several types of similarity determinations may be the same value or different values. Also, the threshold TH can be arbitrarily set within the range of -1<TH<1, and for example, 0.7 or 0.9 can be used as the threshold.

When the cosine similarity of two variance vectors exceeds a predetermined threshold, it is considered that the user has seen multiple similar news items. In this case, the field to which the similar news items belong is considered to be the field in which the user is interested. Therefore, the average vector of the distributed representation vectors of a plurality of similar news items is considered to be "the vector representing the field in which the user is interested in a distributed representation". In this specification, a vector that represents the field in which the user is interested in a distributed manner is referred to as an "interested vector".

(2) The uninteresting vector is generated by the vector generation unit 25 as follows.

Similar to the interest vector, distributed representation vectors are also generated from news that the user did not see in detail among the recommended news. However, if the user has not seen the details of the recommended news, determine if the user is not interested in the news or if the user has not seen the details of the news for some other reason. I can't. Other causes may be, for example, that the user did not have time to look at the details of the news, or that the user already knew the full content of the news.

For the above reason, in order to determine that the calculated distributed representation vector of news is "a vector representing a distributed representation of a field that the user is not interested in", words similar to the calculated distributed representation vector of news The user is presented with words represented by vectors. This allows the user to select whether or not the field in which the user is interested is similar to the presented word.

By this selection, if the user answers that he/she is not interested in fields similar to the presented word, the distributed representation vector similar to the word vector is "a field in which the user is not interested". is a distributed representation of ”. In this specification, a vector representing a field in which the user is not interested is called a "non-interest vector".

(3) Interested vectors and uninterested vectors are updated by the vector generation unit 25 as follows.

After the interest vector is generated, if the user sees the details of the news distributed in a vector similar to the generated interest vector, this interest vector is the vector that the user is interested in. It is considered that the possibility has increased. Therefore, the importance of this vector of interest is added.

The above-mentioned importance of interest vector is also called "interest importance (or weight of interest vector)". It is assumed that the interest importance is the number of news items distributed and represented by vectors similar to the interest vectors. Therefore, when a vector of interest is generated for the first time, the importance of the vector of interest is incremented by two. That is, when an interested vector is generated for the first time, the initial value of the importance of the interested vector is 2 because the average vector of two similar content vectors becomes the interesting vector. After that, each time the user answers that he or she desires to see news distributedly represented by a content vector similar to the interest vector, the interest importance is incremented by one.

After the disinterest vector is generated, if the user does not see the details of the news distributed representation in the vector similar to the generated disinterest vector, this disinterest vector is the vector that the user is not interested in. It is thought that there is an increased possibility. Therefore, 1 is added to the importance of the no-interest vector.

The importance of the uninterested vector is also called "uninterested importance (or the weight of the uninterested vector)". This uninteresting importance is assumed to be the number of news items distributed and represented by vectors similar to the uninteresting vector. However, when the uninteresting vector is generated for the first time, 2 is added to the importance of the uninteresting vector. That is, when the uninteresting vector is first generated, the initial value of the uninteresting importance is 2 because the average vector of two similar content vectors is the uninteresting vector. After that, each time the user answers that he/she does not want to see the news distributedly represented by the content vectors similar to the uninterest vector, the importance of the uninterest vector is incremented by one.

In addition, along with the addition of interest importance, the interest vector is updated so as to approach the reaction vector, which is a content vector similar to the interest vector. For example, using the interest importance before increasing, ie, the weight, the weighted average vector of the interest vector and the reaction vector similar to the interest vector vector is taken as the updated interest vector. Also, along with the addition of the uninteresting importance, the uninteresting vector is updated to approach the uninteresting vector, which is a content vector similar to the uninteresting vector. For example, using the uninterested importance before increasing, ie, the weight, the weighted average vector of the uninterested vector and the unresponsive vector similar to the uninterested vector is taken as the updated uninterested vector.

However, the method of updating the interested vector and the uninterested vector is not limited to the method of calculating the weighted average vector described above. The update method may be any method as long as it is a method capable of distributed representation of fields in which the user is interested and fields in which the user is not interested.

(4) Whether the user is interested in certain news or not is estimated by the interest estimation unit 22 as follows.

(i) First, the degree of similarity between the news content vector and all vectors of interest is calculated, and the sum of the degrees of importance of all vectors of interest whose degree of similarity is equal to or greater than the threshold is calculated.

(ii) Similarly, calculate the degree of similarity between the news content vector acquired by the content acquisition unit 21 and all the uninterested vectors, and calculate the sum of the importance of all the uninterested vectors whose similarity is equal to or greater than the threshold. calculate.

(iii) Compare the sum of the importance of all interesting vectors calculated in (i) with the sum of the importance of all uninteresting vectors calculated in (ii). In this comparison, if the sum of the importance of all interesting vectors calculated in (i) is equal to or greater than the sum of the importance of all uninterested vectors calculated in (ii), the user obtains presumed to be interested in the news reported. In other words, the acquired news is presumed to be "interesting content". On the other hand, if the sum of the importance of all vectors of interest calculated in (i) is less than the sum of the importance of all uninterested vectors calculated in (ii), the user presumed not to be interested in the news. In other words, the acquired news is presumed to be "uninteresting content".

For example, in the example shown in FIG. 9, the sum of the importance of the interesting vector similar to the acquired news is 8, and the sum of the importance of the uninteresting vector similar to the acquired news is 11. be. Therefore, the acquired news is presumed to be "uninteresting content".

It should be noted that, as described above, "estimation" means indicating the result of the determination of the presence or absence of interest expected based on the information input from the user's input unit 2IN in the past. Therefore, the determination result may be affirmative or negative depending on the subsequent user's actual operation of the input unit 2IN.

The above items can be summarized as follows. Note that, in the present embodiment, the importance of a vector is the weight used when calculating the weighted average vector.

The control unit C includes an interest estimation unit 22 and a vector generation unit 25. The control unit C estimates whether or not the user is interested in the content based on at least one of the similarity between the content vector and the interested vector and the similarity between the content vector and the uninterested vector. . In this case, it can be said that there is a high possibility that the user will be interested in the content distributedly represented by the content vector having a high degree of similarity to the interested vector. Also, it can be said that there is a high possibility that the user will not be interested in the content distributedly represented by the content vector having a high degree of similarity to the no-interest vector. Therefore, the content providing terminal 2 can estimate whether the user is interested in the content and recommend the content that the user is interested in to the user without burdening the user with the operation of the input unit 2IN. . Therefore, while suppressing an increase in the user's operational burden on the input unit for indicating whether or not the user is interested in the content, the content that the user is presumed to be interested in can be effectively displayed. can be provided to the user.

Further, the control unit C determines whether or not the user is interested in the content based on at least one of the importance of the interested vector similar to the content vector and the importance of the uninterested vector similar to the content vector. to estimate Therefore, not only the degree of similarity between the two vectors but also the degree of importance of the vector to be compared is taken into account to estimate whether or not the user is interested in the content, and to recommend the content that the user is interested in to the user. can do. Therefore, it is possible to reliably recommend content that the user is likely to be interested in to the user.

Furthermore, when the user is interested in the content, if the content vector and the interested vector are similar, the control unit C increases the importance of the interested vector that is similar to the content vector. On the other hand, if the user is not interested in the content and the content vector and the uninterest vector are similar, the control unit C increases the importance of the uninterest vector similar to the content vector. Therefore, it is possible to reflect the user's intention of whether or not he/she is interested in individual contents in the importance of the interested vector and the importance of the uninterested vector.

In addition, when the importance of the vector of interest is increased, the control unit C updates the weighted average vector of the vector of interest and the content vector by using the importance of the vector of interest before the increase as a weight. generated as a vector of interest. On the other hand, when the importance of the uninterested vector is increased, the control unit C updates the weighted average vector of the uninterested vector and the content vector by using the importance of the uninterested vector before the increase as a weight. Generate as a no-interest vector. Therefore, it is possible to reflect the intention of the user as to whether or not the user is interested in individual contents in the interested vector itself and the uninterested vector itself.

More specifically, the above is as follows.

When the similarity between the reaction vector, which is the content vector of the content in which the user is interested, and the interest vector is equal to or greater than a predetermined threshold, the control unit C determines the interest importance indicating the weight of the interest vector. Increase by 1. Further, when the similarity between the no-response vector, which is the content vector of the content the user was not interested in, and the no-interest vector is equal to or greater than a specific threshold, the control unit C weights the no-interest vector. Increase the importance by 1.

Every time the weight of the vector of interest is increased by 1, the control unit C uses the vector of interest, the vector of reaction, and the weight of interest before the increment to obtain a weighted average vector of the vector of interest and the vector of reaction. , as the updated vector of interest. At this time, the control unit C calculates the weighted average vector of the no interest vector and the no reaction vector by using the weight of the no interest vector, the no reaction vector, and the weight of the no interest vector before the increase to the updated no interest vector. Generate as

At least one interesting vector and at least one uninteresting vector may already exist. In this case, the control unit C calculates the sum of at least one interest importance of at least one interest vector whose similarity to the last acquired content vector is equal to or greater than a predetermined threshold. In addition, the control unit C calculates the sum of at least one uninteresting importance of at least one uninteresting vector whose similarity to the last-acquired content vector is equal to or greater than a specific threshold. By using the comparison result of at least one sum of interest importance and at least one sum of non-interest importance, whether the user is interested in the content distributedly represented by the content vector finally acquired. Estimate whether or not

Specifically, if the sum of at least one interested importance is greater than or equal to the sum of at least one uninterested importance, the control unit C presumes that the user will be interested in the last acquired content. do. On the other hand, if the sum of at least one interested importance is less than the sum of at least one uninterested importance, the user is not interested in the content distributedly represented by the last acquired vector. presumed to be

In addition, there are cases where the degree of similarity between the new response vector, which is the content vector of the content in which the user is interested, and the interested vector is less than a predetermined threshold. In this case, if the degree of similarity between the existing vector with reaction and the new vector with reaction is equal to or greater than a threshold, the control unit C generates an interested vector different from the existing vector with interest. At this time, another vector of interest is the average vector of the already existing vector with response and the new vector with response. The initial value of the other vector of interest is set to two.

On the other hand, there are cases where the degree of similarity between the new no-response vector, which is the content vector of the content that the user was not interested in, and the no-interest vector is less than a specific threshold. In this case, if the degree of similarity between the existing no-reaction vector and the new no-reaction vector is equal to or greater than a threshold, the control unit C generates a no-interest vector different from the existing no-interest vector. At this time, another no-interest vector is the average vector of the existing no-reaction vector and the new no-reaction vector. The initial value of the other Not Interested vector is set to two.

As can be seen from the above, the importance of the interested vector means the number of content vectors used to generate the interested vector, and the importance of the uninterested vector means the number of content vectors used to generate the uninterested vector. Denotes the number of content vectors.

Above, based on the result of comparing the sum of the importance of at least one interested vector similar to the latest content and the sum of the importance of at least one uninterested vector similar to the latest content, the user may Predicted whether or not they would be interested. However, it is assumed that there may be no no-interest vector similar to the content vector. In this case, if the importance of at least one interested vector similar to the latest acquired content is 2 or higher, it may be assumed that the user is interested in the content, and the content may be recommended to the user. In other words, if there is at least one interest vector whose similarity to the most recently obtained content is greater than or equal to a threshold, it may be inferred that the user is interested in the content, and the content may be recommended to the user. Conversely, it is assumed that there may be no interest vector similar to the content vector at all. In this case, if the importance of at least one uninterest vector similar to the latest acquired content is 2 or more, it may be assumed that the user is not interested in the content, and the content may not be recommended to the user. . In other words, if there is at least one uninteresting vector whose similarity to the most recently acquired content is equal to or greater than a threshold, it may be assumed that the user is not interested in the content, and the content may not be recommended to the user.

Also, the control unit C regards a plurality of content vectors as a plurality of reaction vectors on condition that the user is determined to be interested in each of the plurality of contents. Thereby, if a plurality of vectors with reactions are similar, the control unit C generates an average vector of the vectors with a plurality of reactions as a vector of interest. In this case, the control unit C may determine whether or not the plurality of responsive vectors are similar every time it receives the last responsive vector among the plurality of responsive vectors. Further, the control unit C may determine whether or not a plurality of reaction vectors are similar each time a predetermined period of time elapses.

On the other hand, on the condition that it is determined that the user is not interested in each of the plurality of contents, the control unit C regards the plurality of content vectors as the plurality of non-reaction vectors. Thereby, if a plurality of no-reaction vectors are similar, the control unit C generates an average vector of the plurality of no-reaction vectors as a new no-reaction vector. The control unit C asks whether or not the user desires to recommend words represented by distributed representation using word vectors similar to the new no-response vector and new contents represented by distributed representation using new content vectors similar to the word vector. Execute controls for outputting additional query information and Thereafter, when response information indicating that the user does not wish to recommend new content is input from the input unit, the control unit C generates a new no-reaction vector as a no-interest vector. In this case, the control unit C may determine whether or not the plurality of non-reaction vectors are similar every time it receives the last non-reaction vector among the plurality of non-reaction vectors. Further, the control unit C may determine whether or not a plurality of non-reaction vectors are similar each time a predetermined period of time elapses.

FIG. 10 is a second flowchart showing control processing executed by the content providing terminal 2. FIG. Control processing executed by the content output control unit 24 will be described with reference to FIG.

As shown in FIG. 10, when the recommended content is stored in the recommended content storage unit 23 in the aforementioned step S205 (see FIG. 9), the content output control unit 24 outputs the news title to the output unit in step S301. 2 OUT. Furthermore, it outputs question information asking the user whether or not to see the details of the news recommended by the interest estimation unit 22 . Details of the question information will be described later. Note that the content output control unit 24 may cause the output unit 2OUT to output an advertising comment recommending news instead of or in addition to the title.

That is, in S204, when the interest estimation unit 22 estimates that the user will be interested in the content (YES in step S204), the content output control unit 24 performs the following control. In step S301, the content output control unit 24 outputs at least one of information on a part of the content (for example, the title or the first few lines or pages) and recommendation information for recommending the content (for example, advertising comment). Execute control to output. At this time, the content output control unit 24 controls the output unit 2OUT to output question information asking the user whether or not he or she wishes to receive information on the entire content (whether or not to see the details of the news). (hereinafter referred to as "recommendation control"). In other words, the content output control unit 24 controls the output unit 2OUT to output question information asking the user whether the user is interested in the content.

On the other hand, if the interest estimation unit 22 estimates in S204 that the user is not interested in the content (NO in step S204), the interest estimation unit 22 retrieves the news acquired by the content acquisition unit 21 in step S206. delete. Therefore, the content output control unit 24 does not execute the aforementioned recommendation control. According to such control, it is possible to recommend to the user only content that is presumed to be of interest to the user.

Next, in step S302, the response information input from the input unit 2IN for selecting whether the user wants to see the details of the news, that is, the response information as to whether the user is interested in the content is accepted. After that, in step S303, the content output control unit 24 determines whether or not the user has selected to view the details of the news based on the reply information from the input unit 2IN. According to this step 303, the content output control section 24 can determine whether or not the user is interested in the content. If it is determined in step S303 that the user has selected to view the details of the news, the content output control unit 24 determines that the user is interested in the content in step S304, and views the details of the news. Output to the output unit 2OUT.

That is, in step S302, there is a case where "response response information" indicating that the user desires to receive information on the entire content is input from the input unit 2IN. Reaction response information is "response information indicating that the user is interested in the content". In this case, in step S304, the content output control unit 24 executes control to output the information of the entire content to the output unit 2OUT. According to this control, if the user is interested in the recommended content, it is possible to provide the entire content to the user.

After that, in step S305, the content output control unit 24 converts the information of the news that the user has selected to view in detail, specifically, the distributed representation vector of the news (hereinafter also referred to as "reaction vector") to vector It is transmitted to the generation unit 25 . Accordingly, the vector generation unit 25 starts processing for determining whether to update or newly generate the vector of interest.

On the other hand, in step S303, "no response information" indicating that the user has selected not to see the details of the news may be input from the input unit 2IN (NO in S303). The no-response response information is "response information indicating that the user was not interested in the content". In this case, in step S306, the content output control unit 24 considers that the user was not interested in the content. As a result, the content output control unit 24 outputs the news information for which the user has selected not to see details, specifically, the distributed representation vector of the news (hereinafter also referred to as "no reaction vector") to the vector generation unit. 25. Accordingly, the vector generating unit 25 starts processing for determining whether to update or newly generate the uninteresting vector.

FIG. 11 is an example of a screen displayed on the display unit as the output unit 2 OUT, and shows an example of a screen asking the user whether or not to view the details of the news recommended by the interest estimation unit 22 .

As shown in FIG. 11, the display section as the output section 2 OUT displays the title of the news based on the control processing of the content output control section 24 described using FIG. At this time, the output unit 2OUT also displays an icon for selecting "view details" of the news and an icon for selecting "not viewing" the details of the news. Both the icon for selecting "view details" of the news and the icon for selecting "not viewing" the details of the news function as the input unit 2IN.

Along with the display of the screen of FIG. 11, or instead of the display of the screen of FIG. You may make a sound asking if you don't want to see it.

FIG. 12 is a third flowchart showing control processing executed by the content providing terminal 2. FIG. Control processing executed by the vector generation unit 25 that receives information on news that the user has viewed in detail from the content output control unit 24 will be described with reference to FIG. 12 .

As shown in FIG. 12, in step S401, the vector generation unit 25 receives information on the news that the user has viewed in detail, specifically, vectors representing the news in a distributed manner from the content output control unit 24. The content vector of the news that the user has viewed in detail is also referred to as a "reaction vector".

In step S402, the vector generation unit 25 reads all interested vectors stored in the interested vector storage unit 26. After that, in step S403, the vector generation unit 25 determines whether or not each of all extracted interest vectors is similar to the content vector of the news that the user has seen in detail, that is, the response vector. . Note that “whether or not they are similar” is determined based on whether or not the cosine similarity is equal to or greater than a threshold. If it is determined in step S403 that the interested vector and the reacting vector are similar (YES in S403), the interested vector determined to be similar to the reacting vector is updated in step S404. In addition, the vector generation unit 25 increases by one the importance of the vector of interest that is stored in the vector of interest storage unit 26 and determined to be similar to the vector with reaction.

At this time, the vector generating unit 25 uses the importance of the vector of interest similar to the vector with response to generate the vector of the weighted average of the vector of interest with response and the vector of interest similar to the vector with response as the updated vector of interest. Generate as a vector. The weight of the vector of interest is the importance of the vector of interest. The importance of an interested vector means the number of content vectors used to generate the interested vector, that is, the number of reacted vectors.

In the present embodiment, each time information indicating that the user desires to see the details of news, that is, response information indicating that the user is interested in the content, is input from the input unit 2IN. Calculate the similarity between the vector with response and the vector with interest. If it is determined that the degree of similarity between the reacting vector and the interested vector is greater than or equal to the threshold, the interested vector whose similarity to the reacting vector is determined to be greater than or equal to the threshold is updated. Therefore, the interested vector can be updated in real time according to the user's operation of the input unit 2IN.

If it is determined in step S403 that the interest vector and the reaction vector are not similar (NO in S403), in step S405, the vector generation unit 25 extracts all reaction vectors from the reaction history storage unit 27. read out. After that, in step S406, it is determined whether or not the first vector with reaction received from the content output control unit 24 and each of all the vectors with second reaction stored in the reaction history storage unit 27 are similar. Note that “whether or not they are similar” is determined based on whether or not the cosine similarity is equal to or greater than a threshold.

In step S406, it may be determined that the first reaction vector received from the content output control unit 24 and the second reaction vector stored in the reaction history storage unit 27 are similar (YES in S406). ). In this case, in step S407, the second reaction vector determined to be similar to the first reaction vector is deleted from the reaction history storage unit 27. FIG. After that, in step S408, the vector generation unit 25 calculates the average vector of the first reaction vector received from the content output control unit 24 and the second reaction vector read from the reaction history storage unit 27. FIG. After that, the vector generation unit 25 stores the calculated average vector in the interest vector storage unit 26 as another interesting vector. The vector generation unit 25 sets the importance of the other vector of interest, that is, the initial value of the weight to 2, and stores the initial value of 2 in the vector of interest storage unit 26 in association with the vector of interest.

In the present embodiment, every time response information indicating that the user wishes to see details of news, that is, response information indicating that the user is interested in the content, is input from the input unit 2IN. Then, the similarity between the two vectors with reaction is calculated. If the degree of similarity between the two reaction vectors is greater than or equal to a threshold, it is determined that the two reaction vectors are similar, and a new vector of interest is generated. Therefore, a new interested vector can be generated in real time according to the user's operation of the input unit 2IN. The new vector of interest is the average vector of the two similar responsive vectors.

In step S406, it may be determined that the vector with response received from the content output control unit 24 and the vector with response read from the reaction history storage unit 27 are not similar. In this case, in step S<b>409 , the vector generation unit 25 newly stores the reaction vector received from the content output control unit 24 in the reaction history storage unit 27 separately from the reaction vector already stored in the reaction history storage unit 27 . be memorized.

FIG. 13 is a fourth flowchart showing control processing executed by the content providing terminal 2. FIG. A control process executed by the vector generation unit 25 that receives news information of which the user has not viewed details from the content output control unit 24 will be described with reference to FIG. 13 .

As shown in FIG. 13, in step S501, the vector generation unit 25 receives news information that the user did not see in detail, more specifically, content vectors representing the news in a distributed manner, from the non-response history storage unit 28. do. Hereinafter, the content vector of news that the user did not look at in detail will also be referred to as a "non-reaction vector".

Next, in step S502, the vector generation unit 25 reads out all uninteresting vectors stored in the interest vector storage unit 26. After that, in step S503, the vector generation unit 25 determines that among all the uninterested vectors stored in the interest vector storage unit 26, there is a uninterested vector similar to the unreacted vector received from the content output control unit 24. Determine whether or not Note that “whether or not they are similar” is determined based on whether or not the cosine similarity is equal to or greater than a threshold.

In step S503, it may be determined that there is a no-interest vector similar to the no-reaction vector received from the content output control unit 24 (YES in S503). In this case, the vector generation unit 25 updates the no-interest vector determined to be similar to the no-reaction vector in step S504.

Specifically, the vector generation unit 25 uses the uninterest importance (weight) to generate a weighted average vector of the uninterest vector and the unresponsive vector similar to the uninterest vector as the updated interested vector. do. At this time, the vector generation unit 25 increases the importance of the uninteresting vector stored in the interest vector storage unit 26 by one.

The weight of the uninterested vector in the weighted average is the importance of the uninterested vector. The importance of a no-interest vector is the number of content vectors, ie, no-reaction vectors, used to generate the no-interest vector.

In the present embodiment, each time information indicating that the user does not wish to see the details of the news (hereinafter referred to as "no-response information") is input from the input unit 2IN, the aforementioned no-response vector and the similarity of the no-interest vector. The no-response response information is "response information indicating that the user was not interested in the content". If it is determined that the similarity between the no-reaction vector and the no-interest vector is equal to or greater than the threshold, the no-interest vector determined to have the similarity to the no-reaction vector equal to or greater than the threshold is updated. Therefore, the no-interest vector can be updated in real time according to the user's operation of the input unit 2IN.

On the other hand, in step S503, it may be determined that there is no uninterested vector similar to the no-reaction vector (NO in S503). In this case, in step S505, the vector generation unit 25 reads all no-response vectors from the no-response history storage unit .

Next, in step S506, the vector generation unit 25 determines whether the no-response vector received from the content output control unit 24 is similar to any of all the second no-response vectors stored in the no-response history storage unit 28. determine whether or not Note that “whether or not they are similar” is determined based on whether or not the cosine similarity is equal to or greater than a threshold.

In step S506, it may be determined that the first no-reaction vector received from the content output control unit 24 is not similar to any of the second no-reaction vectors stored in the no-response history storage unit 28 ( NO in S506). In this case, the vector generation unit 25 causes the no-response history storage unit 28 to store the current information, that is, the no-response vector received from the content output control unit 24 in step S507.

On the other hand, in step S506, the first no-reaction vector received from the content output control unit 24 is any one second no-reaction vector of all the second no-reaction vectors stored in the no-response history storage unit 28. It may be determined that they are similar. In this case (YES in S506), the vector generation unit 25 deletes the similar second no-response vector from the no-response history storage unit 28 in step S508. After that, in step S509, the vector generation unit 25 uses the first no-response vector received from the content output control unit 24 and the second no-response vector read from the no-response history storage unit 28 to generate a new response. Generate a null vector.

That is, in step S303, the vector generation unit 25 executes the following process each time the no-response response information indicating that the user does not wish to receive the provision of the entire content information is input from the input unit IN. do. First, in step S508, the vector generation unit 25 calculates the degree of similarity between two no-response vectors obtained by dispersively representing two no-response contents, which are contents corresponding to no-response reply information. After that, in S509, the vector generation unit 25 generates a new non-reaction vector using two non-reaction vectors whose similarity is equal to or higher than the threshold. According to this control, a no-interest vector can be generated in real time based on the no-response reply information from the user.

At this time, the vector generation unit 25 uses the average vector of the first no-response vector received from the content output control unit 24 and the second no-response vector read from the no-response history storage unit 28 as a new no-response vector. Generate.

After that, in step S510, the vector generation unit 25 reads word vectors similar to the new no-reaction vector from the word vector storage unit 29. Note that “whether or not they are similar” is determined based on whether or not the cosine similarity is equal to or greater than a threshold.

Next, in step S511, the vector generation unit 25 performs control to output the word distributedly represented by the word vector to the output unit 2OUT, as shown in FIG. 14, the vector generation unit 25 outputs question information asking the user whether or not he/she wishes to have new news containing or related to the word recommended to the user. 2OUT is controlled to output. New content that includes or is associated with a word is new news distributed representation of the word with a new content vector that is similar to the word vector that distributed representation of the word.

After that, in step S512, the vector generation unit 25 may receive a signal indicating that "do not display much" (see FIG. 14) has been selected (NO in S512). In other words, the vector generation unit 25 may receive the user's answer information from the input unit 2IN indicating that the user does not want new content to be recommended. In this case, in step S513, the vector generation unit 25 stores the generated new no-reaction vector in the interest vector storage unit 26 as a no-interest vector. Accordingly, a new no-response vector can be determined as another no-interest vector based on the answer information from the user. At this time, the vector generation unit 25 sets the importance of the other uninterested vector, that is, the initial value of the weight to 2, and associates the initial value of 2 with the other uninterested vector in the interest vector storage unit 26. Memorize.

On the other hand, in step S512, the vector generation unit 25 may receive a signal indicating that "continue display" (see FIG. 14) has been selected (YES in S512). That is, the vector generation unit 25 may receive, from the input unit 2IN, the user's answer information indicating that the user wishes to have new news recommended. In this case, the new no-reaction vector is deleted in step S514.

FIG. 14 shows an example of a screen displayed on the display section of the output section 2OUT of the content providing terminal 2 according to the embodiment. FIG. 14 shows an example of a screen displaying a word and question information asking the user whether he wishes to continue displaying news related to that word on the display unit.

The vector generation unit 25 causes the output unit 2OUT to output three words, for example, "corona", "PCR", and "self-restraint". "Corona", "PCR", and "self-restraint" are three words distributed and represented by the top three word vectors with the highest similarity among the word vectors similar to the news content vector that the user did not see in detail. is. Note that “whether or not they are similar” is determined based on whether or not the cosine similarity is equal to or greater than a threshold.

In addition, the vector generation unit 25 causes the output unit 2OUT to display another question information that asks the user whether or not he/she wishes to have new content containing or related to the word recommended to the user. Specifically, as shown in FIG. 14, the vector generation unit 25 causes the output unit 2OUT to output question information as to whether such news will be "displayed" in the future or "will not be displayed very often." . The output unit 2OUT may cause the display unit to display the words and question information, or may cause the sound generation unit to generate a voice for reading out the words and question information.

In the case of the present embodiment, the answer information to the question information described above is transmitted to the vector generation unit 25 by clicking either the "display" icon or the "rarely display" icon. In other words, the input section 2IN of the present embodiment is the "display" icon and the "rarely display" icon displayed on the display section. However, the input unit 2IN may be a sound recognizer device. In this case, when the user utters voice, the sound recognition device recognizes the response information indicating either “display” or “do not display”, and the sound recognition device is digitized to the vector generation unit 25. Send response information for

(Embodiment 2)
A content providing terminal according to the second embodiment and a content providing program used therefor will be described with reference to FIGS. 15 to 18. FIG. In the following description, the description of the same points as those of the first embodiment will not be repeated.

In this embodiment, as in the first embodiment, the content providing terminal 2 possessed by the user provides the user with content that is presumed to be of interest to the user from among the content acquired from the content distribution server 1. It is recommended. However, the content providing terminal 2 according to the present embodiment collectively processes reaction vectors and non-reaction vectors corresponding to several content vectors acquired within a predetermined period. It is different from the providing terminal 2 .

Differences between the content providing terminal 2 of the present embodiment and the content providing terminal 2 of the first embodiment will be described below.

FIG. 15 is a diagram showing the internal configuration of the content providing terminal 2 according to the second embodiment.

As can be seen from FIG. 15, in the present embodiment, not only the vector generation unit 25 but also the content output control unit 24 store reaction vectors in the reaction history storage unit 27, and store non-reaction vectors in the non-reaction history. It differs from the first embodiment in that it is stored in the storage unit 28 . Except for this point, the internal configuration of content providing terminal 2 of the present embodiment is the same as the internal configuration of content providing terminal 2 of Embodiment 1, and therefore description of the similar configuration will not be repeated.

FIG. 16 is a first flowchart showing control processing executed by the content providing terminal according to the second embodiment.

The content providing terminal 2 of the present embodiment executes the same processing as the processing from step S201 to step S206 executed by the content providing terminal 2 of the first embodiment. After that, the content providing terminal 2 executes the processes of steps S201 to S206 shown in FIG. 8, and then executes the processes of steps S601 to S606 shown in FIG.

As can be seen from FIG. 16, steps S601 to S604 are the same as steps S301 to S304 in Embodiment 1, and therefore description thereof will not be repeated. The present embodiment differs from the first embodiment in that the content providing terminal 2 executes the processes of steps S605 and S606 instead of steps S305 and S306 of the embodiment.

In step S605, the content output control unit 24 executes processing for storing, in the reaction history storage unit 27, the response vector, which is the content vector of the news for which the user has selected to view details in step S603 (YES in S603). do. That is, when response information indicating that the user is interested in the content is input from the input unit 2IN, the content output control unit 24 determines that the user is interested in the content. Thereby, the content output control unit 24 stores the distributed representation vector of the content in the reaction history storage unit 27 as a reaction vector. On the other hand, in step S606, the content output control unit 24 stores, in the non-response history storage unit 28, the no-response vector, which is the content vector of the news for which the user selected not to view the details in step S603 (NO in S603). Execute the process that causes the That is, when response information indicating that the user is not interested in the content is input from the input unit 2IN, the content output control unit 24 determines that the user is not interested in the content. Thereby, the content output control unit 24 causes the reaction history storage unit 27 to store the distributed representation vector of the content as a non-reaction vector.

In the present embodiment, the processing of steps S201 to S206 shown in FIG. 8 and the processing of steps S601 to S606 shown in FIG. Repeated until time elapses.

FIG. 17 is a second flowchart showing control processing executed by the content providing terminal according to the second embodiment. FIG. 17 shows a process executed using several responsive vectors representing distributed representations of several responsive contents estimated by the interest estimation unit 22 within a predetermined period of time that the user is likely to be interested.

In step S701, the vector generation unit 25 reads all reaction vectors (news information) stored in the reaction history storage unit 27 within a predetermined period of time each time a predetermined period of time has elapsed from the reference time. This response vector is a content vector of news that the user is interested in among the news acquired within a predetermined period. That is, the vector with response is a content vector corresponding to response information with response input from the input unit 2IN within a predetermined period. Next, in step S<b>702 , the vector generation unit 25 reads out all interested vectors stored in the interested vector storage unit 26 . After that, in step S703, it is determined whether or not there is an interested vector similar to any of all the reacted vectors among all the interested vectors. Determination of whether or not each of the reacting vectors acquired within a predetermined time is similar to the interesting vector is sequentially executed according to the order of acquiring the reacting vectors. Note that “whether or not they are similar” is determined based on whether or not the cosine similarity is equal to or greater than a threshold. If it is determined in step S703 that there is no interested vector similar to the reacting vector (NO in S703), the vector generation unit 25 executes the process of step S706.

On the other hand, if it is determined that there is an interested vector similar to the reactive vector, the vector generation unit 25 updates the interested vector similar to the reactive vector in step S704. At this time, the vector generating unit 25 uses the importance of the vector of interest similar to the vector with response to generate the vector of the weighted average of the vector of interest with response and the vector of interest similar to the vector with response as the updated vector of interest. Generate as a vector. The weight of the vector of interest is the importance of the vector of interest. The importance of an interested vector means the number of content vectors, ie, reacted vectors used to generate the interested vector. At this time, the vector generating unit 25 increases the importance of the vector of interest stored in the vector of interest storage unit 26, that is, the weight by one.

Next, in step S705, the vector generation unit 25 causes the reaction history storage unit 27 to store the reaction vector determined to be similar to the interested vector in step S703 as a processed reaction vector. After that, in step S<b>706 , the vector generation unit 25 reads all unprocessed reaction vectors from the reaction history storage unit 27 .

Next, in step S707, the vector generation unit 25 determines whether or not there is a set of similar unprocessed vectors with reaction in the reaction history storage unit 27. Specifically, the vector generation unit 25 calculates the degree of similarity of each pair of two similar reaction vectors among the unprocessed reaction vectors stored in the reaction history storage unit 27 as Calculations are performed sequentially according to the order in which the vectors are stored. After that, the vector generation unit 25 determines whether or not the cosine similarity of each pair of two vectors with reaction similar to the unprocessed vector with reaction is equal to or greater than a threshold. In step S707, it may be determined that there are two sets of unprocessed vectors with reactions that are similar to each other in the reaction history storage unit 27 (YES in S707). In this case, in step S708, the vector generation unit 25 generates an average vector of two similar unprocessed vectors with a reaction as a new different vector of interest, and converts the generated another vector of interest to an interest vector. Store in the storage unit 26 . At this time, the vector generation unit 25 sets the importance of another vector of interest, that is, the initial value of the weight to 2, associates the initial value of 2 with the other vector of interest, and stores it in the vector-of-interest storage unit 26. Memorize.

For example, first, among the unprocessed vectors with reaction, the degree of similarity between the vector with reaction that is stored first and the vector with reaction that is stored second is calculated. If the degree of similarity is less than the threshold, the degree of similarity between the first stored vector with reaction and the third stored vector with reaction is calculated. In this way, the degree of similarity between the vector with response stored first and the vector with response stored after the first vector with response is sequentially calculated according to the order in which the vectors with response were stored. Thereby, if there are other reactive vectors similar to the first stored reactive vector, the first stored reactive vector and other similar reactive vectors are processed. At this time, the average vector of the two similar responsive vectors is generated as a new responsive vector. In this case, if the first stored vector with response and, for example, the tenth stored vector with response are similar, then the second stored vector with response and the third stored vector with response are similar. The degree of similarity with the reacted vector is calculated.

On the other hand, if the degree of similarity between the first stored vector with reaction and the second stored vector with reaction is equal to or greater than the threshold, the first stored vector with reaction and the second stored vector with reaction Let the vector and be the processed vector with reaction. Also, the average vector of the first reacted vector and the second reacted vector is generated as a new reacted vector. After that, the degree of similarity between the third stored vector with reaction and the fourth stored vector with reaction is calculated. Such processing is sequentially executed according to the order in which the response vector is stored.

On the other hand, in step S707, it may be determined that there is no set of two unprocessed vectors with reactions similar to the reaction history storage unit 27 (NO in S707). In this case, the vector generator 25 terminates the processing.

As described above, in the present embodiment, unlike the first embodiment, the vector generation unit 25 generates a plurality of reaction vectors (process ) are calculated for each similarity. Thereby, the vector generation unit 25 determines whether or not there is a set of similar unprocessed vectors with reaction in the reaction history storage unit 27 . Therefore, for example, the processing for grasping the user's interest can be performed during a period when the user does not use the content providing terminal 2, for example, at night. Therefore, it is possible to reduce the processing load on the content providing terminal 2 while the user is using the content providing terminal 2 .

FIG. 18 is a third flowchart showing control processing executed by the content providing terminal according to the second embodiment. FIG. 18 shows processing executed by using several no-response vectors representing distributed representations of non-response contents estimated by the interest estimating unit 22 within a predetermined period of time that the user would not be interested.

In step S801, the vector generation unit 25 reads all no-response vectors (news information) stored in the no-response history storage unit 28 within a predetermined period of time every time a predetermined period of time elapses from the reference time. This no-response vector is a content vector of news that the user is not interested in among the news acquired within a predetermined period of time. That is, the no-response vector is a content vector corresponding to the no-response response information input from the input unit 2IN within a predetermined period. Next, in step S802, the vector generation unit 25 reads all uninteresting vectors stored in the interest vector storage unit 26. FIG. After that, in step S803, the vector generation unit 25 determines whether or not there is a non-interest vector similar to any of all the non-reaction vectors among all the non-interest vectors. Note that “whether or not they are similar” is determined based on whether or not the cosine similarity is equal to or greater than a threshold.

If it is determined in step S803 that there is no uninterested vector similar to the no-reaction vector (NO in S803), the vector generation unit 25 executes the process of step S806. On the other hand, if it is determined that there is a no-interest vector similar to the no-reaction vector (YES in S803), the vector generation unit 25 updates the no-interest vector similar to the no-reaction vector in step S804. At this time, the vector generating unit 25 uses the importance of the uninterested vector similar to the unreacted vector to generate a vector of the weighted average of the uninterested vector similar to the unreacted vector and the updated uninterested vector. Generate as a none vector. The weight of the uninterested vector is the importance of the uninterested vector. The importance of a no-interest vector means the number of content vectors, ie, no-reaction vectors, used to generate the no-interest vector. At this time, the vector generation unit 25 increases the importance, that is, the weight of the uninteresting vector stored in the interest vector storage unit 26 by one.

Next, in step S805, the vector generation unit 25 causes the no-reaction history storage unit 28 to store the no-reaction vector determined to be similar to the no-interest vector in step S803 as a processed no-reaction vector. After that, in step S806, the vector generation unit 25 extracts all unprocessed no-reaction vectors from the no-reaction history storage unit .

Next, in step S807, the vector generation unit 25 determines whether or not there is a set of similar unprocessed no-response vectors in the no-response history storage unit 28 or not. Specifically, the vector generation unit 25 calculates the cosine similarity of each pair of two non-reaction vectors out of the unprocessed no-reaction vectors stored in the non-reaction history storage unit 28 as the unprocessed reaction None vectors are calculated sequentially according to the order in which they were acquired. After that, the vector generation unit 25 determines whether or not the cosine similarity of each pair of two unreacted vectors among the unprocessed unreacted vectors is equal to or greater than a threshold. If it is determined in step S807 that there is no set of two similar unprocessed no-response vectors in the no-response history storage unit 28, the vector generation unit 25 terminates the process.

On the other hand, in step S807, it may be determined that there are two sets of similar unprocessed no-reaction vectors in the no-reaction history storage unit . In this case, in step S808, the vector generation unit 25 generates an average vector of two similar unprocessed vectors without reaction as a new vector with no reaction.

For example, first, the degree of similarity between the first stored non-reaction vector and the second stored non-reaction vector among the unprocessed non-reaction vectors is calculated. If the degree of similarity is less than the threshold, the degree of similarity between the first stored no-reaction vector and the third stored no-reaction vector is calculated. In this way, the degree of similarity between the first stored no-reaction vector and the second and subsequent no-reaction vectors is sequentially calculated according to the order in which the other no-reaction vectors are stored. Thereby, if there are other non-reaction vectors similar to the first stored non-reaction vector, the first stored non-reaction vector and other similar non-reaction vectors are processed. At this time, the average vector of the two similar non-reaction vectors is generated as a new non-reaction vector. In this case, if the first stored no reaction vector and, for example, the tenth stored no reaction vector are similar, then the second stored no reaction vector and the third stored no reaction vector are similar. The degree of similarity with the no-response vector is calculated.

On the other hand, if the degree of similarity between the first stored no-response vector and the second stored no-response vector is equal to or greater than the threshold, the first stored no-response vector and the second stored no-response vector Let vectors be the processed non-reaction vectors. Also, the average vector of the first no-reaction vector and the second no-reaction vector is generated as a new no-reaction vector. After that, the degree of similarity between the third stored no-reaction vector and the fourth stored no-reaction vector is calculated. Such processing is sequentially executed according to the order in which the no-reaction vectors are stored.

After that, in step S809, the vector generation unit 25 reads word vectors similar to the new no-reaction vector from the word vector storage unit 29. Note that “whether or not they are similar” is determined based on whether or not the cosine similarity is equal to or greater than a threshold. Next, in step S810, the vector generation unit 25 performs control to output the word distributedly represented by the word vector to the output unit 2OUT, as shown in FIG. Also, in step S811, the vector generation unit 25 asks the user whether he wants new news containing or related to the word to be recommended to the user, as shown in FIG. It executes control to output information to the output unit 2OUT. New content that includes or is associated with a word is new news distributed representation of the word with a new content vector similar to the word vector that distributed representation of the word.

After that, in step S811, the vector generation unit 25 may receive a signal indicating that "do not display much" (see FIG. 14) has been selected (NO in S811). In other words, the vector generation unit 25 may receive the user's answer information from the input unit 2IN indicating that the user does not want new content to be recommended. In this case, in step S812, the vector generation unit 25 causes the interest vector storage unit 26 to store the generated new no-reaction vector as a no-interest vector. Accordingly, a new no-response vector can be determined as another no-interest vector based on the answer information from the user. At this time, the vector generation unit 25 sets the importance of the other uninterested vector, that is, the initial value of the weight to 2, and associates the initial value of 2 with the other uninterested vector in the interest vector storage unit 26. Memorize.

On the other hand, in step S811, the vector generation unit 25 may receive a signal indicating that "continue displaying" (see FIG. 14) has been selected (YES in S811). That is, the vector generation unit 25 may receive, from the input unit 2IN, the user's answer information indicating that the user wishes to have new news recommended. In this case, the new no-reaction vector is deleted in step S813.

The content providing terminal 2 in the present disclosure includes a computer as a control unit C. The main function of the content providing terminal 2 in the present disclosure is realized by the computer executing the content providing program.

The computer as the control unit C has a processor, for example, a CPU (Central Processing Unit) that operates according to the content providing program, as its main hardware configuration. The processor can be of any type as long as it can implement the functions by executing the content providing program. A processor is composed of one or more electronic circuits including a semiconductor integrated circuit, for example, an IC (Integration Circuit) or an LSI (Large Scale Integration). A plurality of electronic circuits may be integrated on one chip or may be provided on a plurality of chips. A plurality of chips may be integrated into one device, or may be provided in a plurality of devices.

The content providing program is recorded in a non-temporary recording medium such as a computer-readable ROM (Read Only Memory), optical disk, hard disk drive, or the like. The content providing program may be pre-stored in the recording medium, or may be supplied to the recording medium via a wide area network including the Internet.

Claims

A control unit capable of receiving a content vector distributed representation of content, an interested vector distributed representation of a field in which the user is interested, and a no-interest vector distributed representation of a field in which the user was not interested. ,
The control unit determines whether the user is interested in the content based on at least one of a similarity between the content vector and the interested vector and a similarity between the content vector and the uninterested vector. A content providing terminal that estimates whether or not.
When the user is estimated to be interested in the content, the control unit checks at least one of information of a part of the content and recommendation information and whether or not the user is interested in the content. 2. The content providing terminal according to claim 1, which executes control for outputting question information to ask said user.
3. The content providing terminal according to claim 2, wherein when response information indicating that the user is interested in the content is input, the control unit performs control to provide information on the entire content. .
The control unit determines whether the user is interested in the content based on at least one of the importance of the interested vector similar to the content vector and the importance of the uninterested vector similar to the content vector. 4. The content providing terminal according to any one of claims 1 to 3, which estimates whether or not it has
The control unit
if the vector of interest is at least one vector of interest and the vector of no interest is at least one vector of no interest,
Whether the user is interested in the content based on a comparison result between the sum of the importance of each of the at least one interested vector and the sum of the importance of each of the at least one uninterested vector. 5. The content providing terminal according to claim 4, which estimates whether or not.
The control unit
When it is determined that the user is interested in the content, if the content vector and the interest vector are similar, increasing the importance of the interest vector,
6. When it is determined that the user is not interested in the content, if the content vector and the uninterest vector are similar, the importance of the uninterest vector is increased. The content providing terminal described in .
The control unit
When the importance of the vector of interest is increased, the importance of the vector of interest before the increase is regarded as a weight, and a weighted average vector of the vector of interest and the content vector is updated to the While generating as an interest vector,
When the importance of the uninterested vector is increased, the importance of the uninterested vector before the increase is regarded as a weight, and a weighted average vector of the uninterested vector and the content vector is used as the updated 7. The content providing terminal according to claim 6, which is generated as a no-interest vector.
On the condition that the user is determined to be interested in each of the plurality of contents, the control unit regards the plurality of content vectors as a plurality of reaction vectors, and determines that the plurality of reaction vectors are similar. 8. The content providing terminal according to any one of claims 1 to 7, wherein if so, an average vector of said plurality of responsive vectors is generated as said interested vector.
9. The content providing terminal according to claim 8, wherein the control unit determines whether or not the plurality of reaction vectors are similar every time it receives the last vector with reaction among the plurality of vectors with reaction. .
The content providing terminal according to claim 8, wherein the control unit determines whether or not the plurality of reaction vectors are similar each time a predetermined period elapses.
The control unit
If it is determined that the user is not interested in each of the plurality of contents, the plurality of content vectors are regarded as a plurality of non-reaction vectors, and if the plurality of non-reaction vectors are similar , generating an average vector of the plurality of no-reaction vectors as a new no-reaction vector;
Another question that asks whether or not the user wishes to recommend words represented in a distributed representation with a word vector similar to the new no-response vector and new content represented in a distributed representation with a new content vector similar to the word vector. 8. The content providing terminal according to any one of claims 1 to 7, which executes control for outputting information.
12. The control unit according to claim 11, wherein said control unit generates said new no-reaction vector as said no-interest vector when reply information indicating that said new content is not recommended is input from said input unit. Content provider terminal.
The content according to claim 11 or 12, wherein the control unit determines whether or not the plurality of non-reaction vectors are similar every time it receives the last non-reaction vector among the plurality of non-reaction vectors. Provided terminal.
The content providing terminal according to claim 11 or 12, wherein the control unit determines whether the plurality of non-reaction vectors are similar each time a predetermined period elapses.
Controlling a computer to receive a content vector distributed representation of content, an interested vector distributed representation of a field in which a user is interested, and a no-interest vector distributed representation of a field in which the user was not interested a computer-readable content-providing program for operating as a unit,
The control unit determines whether the user is interested in the content based on at least one of a similarity between the content vector and the interested vector and a similarity between the content vector and the uninterested vector. A content providing program that estimates whether or not.