WO2019167748A1 - Program, device and method for estimating empathetic influence of users with respect to content - Google Patents

Program, device and method for estimating empathetic influence of users with respect to content Download PDF

Info

Publication number
WO2019167748A1
WO2019167748A1 PCT/JP2019/006283 JP2019006283W WO2019167748A1 WO 2019167748 A1 WO2019167748 A1 WO 2019167748A1 JP 2019006283 W JP2019006283 W JP 2019006283W WO 2019167748 A1 WO2019167748 A1 WO 2019167748A1
Authority
WO
WIPO (PCT)
Prior art keywords
user
matrix
content
ability
feature
Prior art date
Application number
PCT/JP2019/006283
Other languages
French (fr)
Japanese (ja)
Inventor
一則 松本
啓一郎 帆足
池田　和史
Original Assignee
Kddi株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Kddi株式会社 filed Critical Kddi株式会社
Publication of WO2019167748A1 publication Critical patent/WO2019167748A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Definitions

  • the present invention relates to a technique for estimating a user's sympathy ability for content.
  • Non-Patent Document 1 In recent years, dialogue systems that interact with people have also been installed in smartphone applications and robots. In the old days, there is a technique of responding to user utterances in a question-and-answer format, such as ELIZA and SHRDLU (see Non-Patent Document 1, for example). According to this technique, a question example close to a user utterance is detected by pattern matching from a set of question examples accumulated in advance. On the other hand, the current dialogue system can be roughly divided into “task-oriented dialogue function” and “chat conversation dialogue function”, and these functions are combined. There is also a dialogue system technology that focuses on “social dialogue” for the purpose of establishing and maintaining interpersonal relationships (see, for example, Non-Patent Document 2).
  • Non-Patent Document 3 Furthermore, there is a technique for increasing the user's familiarity with a dialogue system that applies empathy to social dialogue (see, for example, Non-Patent Document 3). Furthermore, there is a technique in which an information recommendation function in a user's field of interest is applied to a dialog system (see, for example, Non-Patent Document 4).
  • a user who browses an article understands the characters (including anthropomorphic objects) of the article and the thoughts and feelings of the author, and has some feelings. This is based on the user's ability to sympathize with the content.
  • the user's sympathy ability varies from person to person. For example, even if the content is the same, each user has different sympathy.
  • an object of the present invention is to provide a program, an apparatus, and a program for estimating a user's empathy ability with respect to content.
  • a program for causing a computer mounted on a device for estimating the user's empathy ability ⁇ E of the estimation target user sE to the content c to function In order to realize the learning function, the computer is associated with the sensitivity (t11 to t1L,..., TN1 to tNL) of each content cl (c1 to cL) for each user sn (s1 to sN).
  • a co-sensitivity matrix storage means for storing Tnl;
  • User feature value matrix storage means for storing a user feature value matrix Fnj in which feature values (f11 to f1J,..., FN1 to fNJ) of predetermined user feature elements (v1 to vJ) are associated with each user sn.
  • Item response theory calculation means for inputting the co-sensitivity matrix Tnl of each content cl for each user sn and calculating sympathy ability ⁇ n based on item response theory IRT (Item Response Theory);
  • IRT Items Response Theory
  • the computer calculates a user feature matrix FE in which the feature quantities (fE1 to fEJ) of the predetermined user feature elements (v1 to vJ) are associated with the estimation target user sE.
  • the machine learning engine functions to input the user feature matrix FE of the estimation target user sE and output the sympathy ability ⁇ E.
  • the co-sensitivity is Recognition tR to understand the emotions of the characters or authors of the content
  • the emotional level tE that produces the same emotion as the content character or author
  • the item response theory calculation means preferably causes the computer to function so as to calculate the sympathetic ability ⁇ n as a two-parameter logistic model.
  • the user characteristic element is based on a plurality of questions for the user sn; It is also preferable to cause the computer to function so that the feature amount is based on an answer of the user sn to the question.
  • the content is posted to a web server from an unspecified number of third parties, It is also preferable that the content influence M is caused to cause the computer to function based on an emotion received when the user browses the content on the Web server.
  • the machine learning engine also preferably causes the computer to function based on linear regression, support vector machine, or deep learning regression learning.
  • an apparatus for estimating the sympathy ability ⁇ E of the estimation target user sE for the content c As a learning function, A sensitivity matrix storage means for storing a sensitivity matrix Tnl in which the sensitivity (t11 to t1L,..., TN1 to tNL) of each content cl (c1 to cL) is associated for each user sn (s1 to sN); User feature value matrix storage means for storing a user feature value matrix Fnj in which feature values (f11 to f1J,..., FN1 to fNJ) of predetermined user feature elements (v1 to vJ) are associated with each user sn.
  • Item response theory calculation means for inputting the empathy matrix Tnl of each content cl for each user sn and calculating the sympathy ability ⁇ n based on the item response theory;
  • a machine learning engine that learns teacher data associating the user feature matrix Fnj and the sympathetic ability ⁇ n for each user sn;
  • User feature value matrix calculating means for calculating a user feature value matrix FE in which the feature values (fE1 to fEJ) of the predetermined user feature elements (v1 to vJ) are associated with the estimation target user sE;
  • the machine learning engine inputs the user feature matrix FE of the estimation target user sE and outputs a sympathetic ability ⁇ E.
  • an apparatus estimation method for estimating a user's empathy ability ⁇ E of an estimation target user sE for content c For each user sn (s1 to sN), a cosensitivity matrix Tnl in which the cosensitivities (t11 to t1L,..., TN1 to tNL) of the contents cl (c1 to cL) are associated is stored in the cosensitivity matrix storage unit.
  • a user feature quantity matrix Fnj in which feature quantities (f11 to f1J,..., FN1 to fNJ) of predetermined user feature elements (v1 to vJ) are associated with each user sn is stored in a user feature quantity matrix storage unit.
  • FIG. 1 is a system configuration diagram according to the present invention.
  • the Web server 2 connected to the Internet publishes various contents.
  • the Web server 2 is, for example, a news site, a blog (Web log) site, a mini blog site (for example, twitter (registered trademark), an SNS (Social Networking Service) site (for example, facebook (registered trademark) or LINE (registered trademark)).
  • the content may be posted on a Web server from an unspecified number of third parties.
  • the terminal 3 accesses the Web server 2 via the access network and the Internet, and causes the user to browse these contents.
  • the estimation apparatus 1 in the present invention estimates the user's sympathy ability for the content. That is, the user group can be classified according to psychological empathy.
  • FIG. 2 is a functional configuration diagram of the estimation apparatus according to the present invention.
  • the estimation apparatus 1 performs estimation by estimating the user's empathy ability ⁇ E of the estimation target user sE with respect to the content c.
  • the estimation apparatus 1 is roughly divided into ⁇ learning function> and ⁇ operation function>, and has a machine learning engine 12 in common for both functions.
  • the estimation apparatus 1 includes a content storage unit 101, a co-sensitivity matrix storage unit 102, a user feature amount matrix storage unit 103, and an item response theory calculation unit 11 as learning functions.
  • a user feature matrix calculation unit 13 is provided as an operation function.
  • These functional components are realized by executing a program that causes a computer installed in the apparatus to function. The processing flow of these functional components can be understood as an apparatus estimation method.
  • the learning function is a machine learning engine based on item response theory results (sympathetic ability ⁇ n) and user feature values of a plurality of users sn (s1 to sN) for each content cl (c1 to cL) as teacher data.
  • a learning model is built in 12.
  • the content storage unit 101 stores a plurality of contents (c1 to cL) as teacher data.
  • the content may be in a format that can be browsed by the user, such as text (articles, blogs, etc.), images, videos, and the like.
  • the co-sensitivity matrix storage unit 102 stores a co-sensitivity matrix Tnl in which the co-sensitivities (t11 to t1L,..., TN1 to tNL) of the contents cl (c1 to cL) are associated with each user (s1 to sN). .
  • FIG. 3 is an explanatory diagram showing the co-sensitivity matrix.
  • each content cl as teacher data for example, a questionnaire is used to obtain a co-sensitivity answer as a subjective evaluation.
  • the co-sensitivity matrix Tnl (t11 to t1L,..., TN1 to tNL) created thereby is stored.
  • the user feature amount matrix storage unit 103 associates, for each user sn, a feature amount (f11 to f1J,..., FN1 to fNJ) of each predetermined user feature element (v1 to vJ) with a user feature amount matrix Fnj (N Row J).
  • the user characteristic element may generally be user profile information (gender, age, occupation, etc.). In particular, it is preferably based on the personality of the user (recognition ability, physical ability, intellectual ability, personality ability, patience, etc.).
  • the user characteristic element may be a parameter element that identifies the characteristics of the user according to the application.
  • the user characteristic element may be based on a plurality of questions for the user sn. In this case, the user feature amount is based on the answer of the user sn to the question.
  • FIG. 4 is an explanatory diagram showing a user feature matrix.
  • each user who becomes a test subject answers a plurality of questions.
  • the contents of the question are preferably those related to user empathy.
  • an answer matrix (N rows and M columns) for each question qm is obtained for each user sn.
  • each user's answer is compressed to J dimension by Singular Value Decomposition.
  • a matrix U in which compressed feature vectors are collected by all users sn is defined as a user feature amount matrix.
  • Singular value decomposition is a matrix decomposition method for a matrix having components of complex numbers or real numbers in linear algebra (see, for example, Non-Patent Document 8).
  • the item response theory calculation unit 11 inputs the cosensitivity matrix Tnl of each content cl for each user sn, and calculates the empathy ability ⁇ n based on the item response theory IRT (Item Response Theory).
  • the item response theory refers to a test theory for measuring the user ability ⁇ (and the identification degree ⁇ / the difficulty level ⁇ ) based on the user's response to the evaluation item group (see, for example, Non-Patent Document 7). This calculates the user ability (and the degree of identification / difficulty) stochastically from the user's discrete answers to the evaluation items. It is suitable for computerized adaptive testing, which maintains a stock of test items for users and regards the difficulty of multiple tests as equivalent.
  • User ability ⁇ l a real value representing the magnitude of the ability of user n for content l (identification degree al: a real value representing the ability of content l to identify the user's empathy ability) (Difficulty level bl: real value representing difficulty of content l)
  • identity degree al a real value representing the ability of content l to identify the user's empathy ability
  • difficulty level bl real value representing difficulty of content l
  • the machine learning engine 12 of the learning function learns as teacher data in which the user feature matrix Fnj and the sympathy ability ⁇ n are associated with each user sn.
  • the machine learning engine 12 may be based on linear regression, support vector machine, or deep learning regression learning.
  • the user feature quantity matrix calculation unit 13 calculates a user feature quantity matrix FE in which feature quantities (fE1 to fEJ) of predetermined user feature elements (v1 to vJ) are associated with the estimation target user sE.
  • the user feature amount matrix calculation unit 13 calculates the user feature amount matrix FE in exactly the same manner as the user feature amount matrix storage unit 103 described above.
  • the machine learning engine 12 of the operational function inputs the user feature matrix FE of the estimation target user (unknown) sE, and outputs the empathy ability ⁇ E.
  • FIG. 5 is an explanatory diagram showing the degree of recognition and emotion of the user with respect to the content.
  • FIG. 5 shows a subjective evaluation obtained by asking many users (subjects) to read the same article and asking each user “a topic that attracts interest” or “a topic that resonates”. The answer result is shown.
  • FIG. 6 is an explanatory diagram in which recognition and emotion are applied as co-sensitivity.
  • FIG. 7 is an explanatory diagram in which recognition and emotion are applied as the co-sensitivity matrix.
  • the co-sensitivity matrix storage unit 102 stores the recognition degree T1 and the emotional degree T2 separately as the co-sensitivity matrix T as compared with FIG.
  • the item response theory calculation unit 11 calculates the empathy ability ⁇ as follows for the user sn (s1 to sN) as a two-parameter logistic model. Awareness of awareness: ⁇ 11 ⁇ ⁇ 1N Emotional empathy ability: ⁇ 21 ⁇ ⁇ 2N In addition, when a part of users can answer at random, the item reaction theory calculation part 11 may be calculated as a 3 parameter logistic model.
  • the machine learning engine 12 learns, for each content cl, teacher data in which the feature quantity matrix Fl, the identification degree ⁇ 11 and the difficulty level ⁇ 11 for the recognition degree, and the identification degree ⁇ 12 and the difficulty level ⁇ 12 for the emotion degree are associated with each other. Then, a learning model is constructed as follows. Learning model for awareness of sympathetic ability ⁇ 11: M1 Learning model for emotional empathy ability ⁇ 12: M2
  • the machine learning engine 12 outputs the sympathy ability ⁇ 1E for the discrimination degree and the sympathy ability ⁇ 2E for the emotion degree.
  • 8A, 8B, 8C, and 8D are explanatory diagrams showing item response category characteristic curves for each content.
  • An item response category characteristic curve (IRCCC) is created as follows.
  • Horizontal axis empathy ability value (left side: low sensitivity, right side: high sensitivity)
  • Vertical axis probability p Curve: each choice (“do not sympathize”"do not know”"sympathize") ("I'm not interested”, “I don't know”, “I'm interested”) ("I don't resonate”"Idon'tknow”"Idon'tresonate”)
  • the item response category characteristic curves of FIGS. 8A to 8D show the co-sensitivity matrices T1 and T2 of the recognition degree and emotion degree for 300 users (s1 to s300) and eight dimensions for 100 contents (c1 to c100). Obtained from the user feature matrix F.
  • the difficulty level b is on the same scale as the empathy ability ⁇ .
  • the degree of discrimination a determines the slope of the item response category characteristic curve. The greater the slope of the curve, the clearer the answer is, the greater the difference between the difficulty level b and the empathy ability ⁇ .
  • the item response category characteristic curves of the co-sensitivity matrices of the recognition degree and the emotion degree are represented for the four contents (ID36, ID72, ID82, ID94).
  • the content ID 82 shown in FIG. 8A and the content ID 94 shown in FIG. 8B have higher recognition than the content ID 72 shown in FIG. 8D.
  • the content ID 82 shown in FIG. 8A has higher emotion than the content ID 94 shown in FIG. 8B and the content ID 72 shown in FIG. 8D.
  • the content ID 82 shown in FIG. 8A has high recognition and high emotion.
  • the content ID 72 shown in which the co-sensitivity of a certain content is estimated to be low can also be obtained when the user's empathy capability is high.
  • the estimation apparatus As described above in detail, according to the program, the estimation apparatus, and the method of the present invention, it is possible to estimate the user's sympathy ability with respect to the content.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

As a learning function, this device which estimates empathetic capacity θE has a machine learning engine learn, for each user (s1-sN), teacher data associating a user characteristic quantity matrix Fnj associating user characteristic quantities (f11-f1J to fN1-fNK) of predetermined user characteristic elements (v1-vJ) with empathetic capacity θn on the basis of item response theory (IRT) in a degree of resonance matrix Tnl associating the degree of resonance (t11-t1L to tN1 to tNL) for content c1 (c1-cL). As an operation function, the device calculates a user characteristic quantity matrix FE associating the characteristic quantities (fE1-fEJ) of each predetermined user characteristic element (v1-vJ) for users SE which are the subject of estimation, and the machine learning engine inputs the user characteristic quantity matrix FE for users SE which are the subject of estimation and outputs the empathetic capacity θE.

Description

コンテンツに対するユーザの共感能力を推定するプログラム、装置及び方法Program, apparatus and method for estimating user's sympathy ability for content
　本発明は、コンテンツに対するユーザの共感能力を推定する技術に関する。
　本願は、２０１８年３月１日に、日本に出願された特願２０１８－３７００１号に基づき優先権を主張し、その内容をここに援用する。 The present invention relates to a technique for estimating a user's sympathy ability for content.
This application claims priority on March 1, 2018 based on Japanese Patent Application No. 2018-37001 filed in Japan, the contents of which are incorporated herein by reference.
　近年、人と対話する対話システムが、スマートフォンアプリケーションやロボットにも搭載されてきている。
　古くはELIZAやSHRDLUのように、ユーザ発話に対して一問一答形式で返答する技術がある（例えば非特許文献１参照）。この技術によれば、事前に蓄積された質問例の集合の中から、パターンマッチによってユーザ発話に近い質問例を検出する。
　これに対し、現在の対話システムは、「タスク指向対話機能」及び「雑談対話機能」に大別でき、これらの機能を組み合わせて構成されている。
　また、対人関係の確立と維持とを目的として、「社会的対話」に注目した対話システムの技術もある（例えば非特許文献２参照）。
　更に、社会的対話に共感性を適用した対話システムによって、ユーザの親近感を高めようとする技術もある（例えば非特許文献３参照）。
　更に、ユーザの興味分野の情報レコメンド機能を、対話システムに適用した技術もある（例えば非特許文献４参照）。 In recent years, dialogue systems that interact with people have also been installed in smartphone applications and robots.
In the old days, there is a technique of responding to user utterances in a question-and-answer format, such as ELIZA and SHRDLU (see Non-Patent Document 1, for example). According to this technique, a question example close to a user utterance is detected by pattern matching from a set of question examples accumulated in advance.
On the other hand, the current dialogue system can be roughly divided into “task-oriented dialogue function” and “chat conversation dialogue function”, and these functions are combined.
There is also a dialogue system technology that focuses on “social dialogue” for the purpose of establishing and maintaining interpersonal relationships (see, for example, Non-Patent Document 2).
Furthermore, there is a technique for increasing the user's familiarity with a dialogue system that applies empathy to social dialogue (see, for example, Non-Patent Document 3).
Furthermore, there is a technique in which an information recommendation function in a user's field of interest is applied to a dialog system (see, for example, Non-Patent Document 4).
　前述した従来技術によれば、対話システムにおけるユーザの親近感を高めようとしているが、心理学の知見に基づいた社会的対話については何ら検討されていない。
　心理学の分野によれば、他者の気持ちを認知し、他者と同じ情動を持つ体験をする性質として、「共感性(Empathy)」がある（例えば非特許文献５参照）。但し、共感性は、単一の観点のみから判断することはできない。具体的には、共感性は、「他者の心理状態を正確に認知する能力（認知度）」と「他者の心理状態に対する代理的な反応の強さ（情動度）」との多元的な観点から判断する必要がある（例えば非特許文献６参照）。 According to the above-described prior art, an attempt is made to enhance the familiarity of the user in the dialogue system, but no social dialogue based on knowledge of psychology has been studied.
According to the field of psychology, there is “Empathy” as a property of recognizing the feelings of others and experiencing the same emotions as others (see, for example, Non-Patent Document 5). However, empathy cannot be judged from a single point of view. Specifically, empathy is a multi-factor of “the ability to accurately recognize another person's psychological state (recognition level)” and “the strength of a proxy response to another person's psychological state (emotion level)”. It is necessary to make a judgment from various viewpoints (see, for example, Non-Patent Document 6).
　例えば、ある記事を閲覧したユーザは、その記事の登場人物（擬人化された物も含む）や著作者の思考及び感情を理解し、何らかの感情を抱く。これは、コンテンツに対するユーザの共感能力に基づく。ユーザの共感能力は、人によって異なり、例えば同じコンテンツであっても、各ユーザが持つ共感性は違ってくる。 For example, a user who browses an article understands the characters (including anthropomorphic objects) of the article and the thoughts and feelings of the author, and has some feelings. This is based on the user's ability to sympathize with the content. The user's sympathy ability varies from person to person. For example, even if the content is the same, each user has different sympathy.
　そこで、本発明は、コンテンツに対するユーザの共感能力を推定するプログラム、装置及びプログラムを提供することを目的とする。 Therefore, an object of the present invention is to provide a program, an apparatus, and a program for estimating a user's empathy ability with respect to content.
　本発明によれば、コンテンツｃに対する推定対象ユーザｓEのユーザの共感能力θEを推定する装置に搭載されたコンピュータを機能させるプログラムであって、前記プログラムは、
　学習機能を実現するために、前記コンピュータを
　ユーザｓn(ｓ1～ｓN)毎に、各コンテンツｃl(ｃ1～ｃL)の共感度(ｔ11～ｔ1L,～,ｔN1～ｔNL)を対応付けた共感度行列Ｔnlを記憶する共感度行列記憶手段と、
　前記ユーザｓn毎に、所定の各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆ11～ｆ1J,～,ｆN1～ｆNJ)を対応付けたユーザ特徴量行列Ｆnjを記憶するユーザ特徴量行列記憶手段と、
　前記ユーザｓn毎に、前記各コンテンツｃlの前記共感度行列Ｔnlを入力し、項目反応理論ＩＲＴ(Item Response Theory)に基づく共感能力θnを算出する項目反応理論算出手段と、
　前記ユーザｓn毎に、前記ユーザ特徴量行列Ｆnjと前記共感能力θnとを対応付けた教師データとして学習する機械学習エンジンと
して機能させ、
　運用機能を実現するために、前記コンピュータを
　前記推定対象ユーザｓEについて、所定の前記各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆE1～ｆEJ)を対応付けたユーザ特徴量行列ＦEを算出するユーザ特徴量行列算出手段と
して機能させると共に、
　前記機械学習エンジンは、前記推定対象ユーザｓEの前記ユーザ特徴量行列ＦEを入力し、共感能力θEを出力するように機能させる。 According to the present invention, there is provided a program for causing a computer mounted on a device for estimating the user's empathy ability θE of the estimation target user sE to the content c to function,
In order to realize the learning function, the computer is associated with the sensitivity (t11 to t1L,..., TN1 to tNL) of each content cl (c1 to cL) for each user sn (s1 to sN). A co-sensitivity matrix storage means for storing Tnl;
User feature value matrix storage means for storing a user feature value matrix Fnj in which feature values (f11 to f1J,..., FN1 to fNJ) of predetermined user feature elements (v1 to vJ) are associated with each user sn. ,
Item response theory calculation means for inputting the co-sensitivity matrix Tnl of each content cl for each user sn and calculating sympathy ability θn based on item response theory IRT (Item Response Theory);
For each user sn, function as a machine learning engine that learns as teacher data in which the user feature matrix Fnj and the sympathetic ability θn are associated,
In order to realize the operation function, the computer calculates a user feature matrix FE in which the feature quantities (fE1 to fEJ) of the predetermined user feature elements (v1 to vJ) are associated with the estimation target user sE. While functioning as a user feature matrix calculation means,
The machine learning engine functions to input the user feature matrix FE of the estimation target user sE and output the sympathy ability θE.
　本発明のプログラムにおける他の実施形態によれば、
　前記共感度は、
　　コンテンツの登場人物又は著作者の感情を理解する認知度ｔRと、
　　コンテンツの登場人物又は著作者と同じ感情を生じる情動度ｔEと
からなり、
　項目反応理論算出手段は、２母数ロジスティックモデルとして前記共感能力θnを算出するようにコンピュータを機能させることも好ましい。 According to another embodiment of the program of the present invention,
The co-sensitivity is
Recognition tR to understand the emotions of the characters or authors of the content,
The emotional level tE that produces the same emotion as the content character or author,
The item response theory calculation means preferably causes the computer to function so as to calculate the sympathetic ability θn as a two-parameter logistic model.
　本発明のプログラムにおける他の実施形態によれば、
　前記ユーザ特徴要素は、前記ユーザｓnに対する複数の質問に基づくものであり、
　前記特徴量は、前記質問に対する前記ユーザｓnの回答に基づくものである
ようにコンピュータを機能させることも好ましい。 According to another embodiment of the program of the present invention,
The user characteristic element is based on a plurality of questions for the user sn;
It is also preferable to cause the computer to function so that the feature amount is based on an answer of the user sn to the question.
　本発明のプログラムにおける他の実施形態によれば、
　前記コンテンツは、不特定多数の第三者からＷｅｂサーバに投稿されたものであり、
　前記コンテンツ影響力Ｍは、ユーザがＷｅｂサーバの当該コンテンツを閲覧した際に受ける感情に基づくものである
ようにコンピュータを機能させることも好ましい。 According to another embodiment of the program of the present invention,
The content is posted to a web server from an unspecified number of third parties,
It is also preferable that the content influence M is caused to cause the computer to function based on an emotion received when the user browses the content on the Web server.
　本発明のプログラムにおける他の実施形態によれば、
　前記機械学習エンジンは、線形回帰、サポートベクターマシン(Support Vector Machine)、又は、深層学習の回帰学習に基づくものであるようにコンピュータを機能させることも好ましい。 According to another embodiment of the program of the present invention,
The machine learning engine also preferably causes the computer to function based on linear regression, support vector machine, or deep learning regression learning.
　本発明によれば、コンテンツｃに対する推定対象ユーザｓEの共感能力θEを推定する装置であって、
　学習機能として、
　ユーザｓn(ｓ1～ｓN)毎に、各コンテンツｃl(ｃ1～ｃL)の共感度(ｔ11～ｔ1L,～,ｔN1～ｔNL)を対応付けた共感度行列Ｔnlを記憶する共感度行列記憶手段と、
　前記ユーザｓn毎に、所定の各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆ11～ｆ1J,～,ｆN1～ｆNJ)を対応付けたユーザ特徴量行列Ｆnjを記憶するユーザ特徴量行列記憶手段と、
　前記ユーザｓn毎に、前記各コンテンツｃlの前記共感度行列Ｔnlを入力し、項目反応理論に基づく共感能力θnを算出する項目反応理論算出手段と、
　前記ユーザｓn毎に、前記ユーザ特徴量行列Ｆnjと前記共感能力θnとを対応付けた教師データとして学習する機械学習エンジンと
を有し、
　運用機能として、
　前記推定対象ユーザｓEについて、所定の前記各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆE1～ｆEJ)を対応付けたユーザ特徴量行列ＦEを算出するユーザ特徴量行列算出手段と
を有し、
　前記機械学習エンジンは、前記推定対象ユーザｓEの前記ユーザ特徴量行列ＦEを入力し、共感能力θEを出力する。 According to the present invention, an apparatus for estimating the sympathy ability θE of the estimation target user sE for the content c,
As a learning function,
A sensitivity matrix storage means for storing a sensitivity matrix Tnl in which the sensitivity (t11 to t1L,..., TN1 to tNL) of each content cl (c1 to cL) is associated for each user sn (s1 to sN);
User feature value matrix storage means for storing a user feature value matrix Fnj in which feature values (f11 to f1J,..., FN1 to fNJ) of predetermined user feature elements (v1 to vJ) are associated with each user sn. ,
Item response theory calculation means for inputting the empathy matrix Tnl of each content cl for each user sn and calculating the sympathy ability θn based on the item response theory;
A machine learning engine that learns teacher data associating the user feature matrix Fnj and the sympathetic ability θn for each user sn;
As an operation function,
User feature value matrix calculating means for calculating a user feature value matrix FE in which the feature values (fE1 to fEJ) of the predetermined user feature elements (v1 to vJ) are associated with the estimation target user sE;
The machine learning engine inputs the user feature matrix FE of the estimation target user sE and outputs a sympathetic ability θE.
　本発明によれば、コンテンツｃに対する推定対象ユーザｓEのユーザの共感能力θEを推定する装置の推定方法であって、
　学習機能として、
　ユーザｓn(ｓ1～ｓN)毎に、各コンテンツｃl(ｃ1～ｃL)の共感度(ｔ11～ｔ1L,～,ｔN1～ｔNL)を対応付けた共感度行列Ｔnlを共感度行列記憶部に記憶し、
　前記ユーザｓn毎に、所定の各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆ11～ｆ1J,～,ｆN1～ｆNJ)を対応付けたユーザ特徴量行列Ｆnjをユーザ特徴量行列記憶部に記憶し、
　前記ユーザｓn毎に、前記各コンテンツｃlの共感度行列Ｔnlを入力し、項目反応理論ＩＲＴ(Item Response Theory)に基づく共感能力θnを算出し、
　機械学習エンジンを用いて、前記ユーザｓn毎に、前記ユーザ特徴量行列Ｆnjと前記共感能力θnとを対応付けた教師データとして学習し
　運用機能として、
　前記推定対象ユーザｓEについて、所定の前記各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆE1～ｆEJ)を対応付けたユーザ特徴量行列ＦEを算出し、
　前記機械学習エンジンが、前記推定対象ユーザｓEの前記ユーザ特徴量行列ＦEを入力し、共感能力θEを出力する第２２のステップと
を実行することを特徴とする。 According to the present invention, there is provided an apparatus estimation method for estimating a user's empathy ability θE of an estimation target user sE for content c,
As a learning function,
For each user sn (s1 to sN), a cosensitivity matrix Tnl in which the cosensitivities (t11 to t1L,..., TN1 to tNL) of the contents cl (c1 to cL) are associated is stored in the cosensitivity matrix storage unit.
A user feature quantity matrix Fnj in which feature quantities (f11 to f1J,..., FN1 to fNJ) of predetermined user feature elements (v1 to vJ) are associated with each user sn is stored in a user feature quantity matrix storage unit. ,
For each user sn, input a co-sensitivity matrix Tnl of each content cl, and calculate a sympathy ability θn based on an item response theory IRT (Item Response Theory),
Using a machine learning engine, for each user sn, learning as teacher data in which the user feature matrix Fnj and the empathy ability θn are associated with each other,
For the estimation target user sE, a user feature value matrix FE that associates feature values (fE1 to fEJ) of the predetermined user feature elements (v1 to vJ) is calculated,
The machine learning engine executes the twenty-second step of inputting the user feature matrix FE of the estimation target user sE and outputting the empathy ability θE.
　本発明のプログラム、装置及び方法によれば、コンテンツに対するユーザの共感能力を推定することができる。 According to the program, apparatus, and method of the present invention, it is possible to estimate the user's empathy ability with respect to the content.
本発明の一実施形態を示すシステム構成図である。It is a system configuration figure showing one embodiment of the present invention. 本発明の一実施形態における推定装置を示す機能構成図である。It is a functional block diagram which shows the estimation apparatus in one Embodiment of this invention. 一実施形態における共感度行列を示す図である。It is a figure which shows the cosensitivity matrix in one Embodiment. 一実施形態におけるユーザ特徴量行列を表す図である。It is a figure showing the user feature-value matrix in one Embodiment. 一実施形態において、コンテンツに対してユーザが持つ認知度及び情動度を示す図である。In one Embodiment, it is a figure which shows the recognition degree and emotion degree which a user has with respect to a content. 一実施形態において、共感度として認知度及び情動度を適用した場合の説明図である。In one embodiment, it is explanatory drawing at the time of applying a recognition degree and emotion degree as co-sensitivity. 一実施形態において、共感度行列として認知度及び情動度を適用した場合の説明図である。In one embodiment, it is explanatory drawing at the time of applying a recognition degree and emotional degree as a co-sensitivity matrix. 一実施形態において、コンテンツ毎の項目反応カテゴリ特性曲線を示すグラフである。In one embodiment, it is a graph which shows the item response category characteristic curve for every content. 一実施形態において、コンテンツ毎の項目反応カテゴリ特性曲線を示すグラフである。In one embodiment, it is a graph which shows the item response category characteristic curve for every content. 一実施形態において、コンテンツ毎の項目反応カテゴリ特性曲線を示すグラフである。In one embodiment, it is a graph which shows the item response category characteristic curve for every content. 一実施形態において、コンテンツ毎の項目反応カテゴリ特性曲線を示すグラフである。In one embodiment, it is a graph which shows the item response category characteristic curve for every content.
　以下、本発明の実施の形態について、図面を用いて詳細に説明する。 Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
　図１は、本発明におけるシステム構成図である。 FIG. 1 is a system configuration diagram according to the present invention.
　インターネットに接続されるＷｅｂサーバ２は、様々なコンテンツを公開する。Ｗｅｂサーバ２は、例えば、ニュースサイトや、ブログ(Web log)サイト、ミニブログサイト（例えばtwitter（登録商標）、ＳＮＳ(Social Networking Service)サイト（例えばfacebook（登録商標）やLINE（登録商標））、掲示板サイトのようなものであってもよい。コンテンツは、不特定多数の第三者からＷｅｂサーバに投稿されたものであってもよい。
　端末３は、アクセスネットワーク及びインターネットを介して、Ｗｅｂサーバ２へアクセスし、それらコンテンツをユーザに閲覧させる。
　本発明における推定装置１は、コンテンツに対するユーザの共感能力を推定する。即ち、ユーザ群を、心理学的な共感性に応じて分類することができる。 The Web server 2 connected to the Internet publishes various contents. The Web server 2 is, for example, a news site, a blog (Web log) site, a mini blog site (for example, twitter (registered trademark), an SNS (Social Networking Service) site (for example, facebook (registered trademark) or LINE (registered trademark)). The content may be posted on a Web server from an unspecified number of third parties.
The terminal 3 accesses the Web server 2 via the access network and the Internet, and causes the user to browse these contents.
The estimation apparatus 1 in the present invention estimates the user's sympathy ability for the content. That is, the user group can be classified according to psychological empathy.
　図２は、本発明における推定装置の機能構成図である。 FIG. 2 is a functional configuration diagram of the estimation apparatus according to the present invention.
　本発明における推定装置１は、コンテンツｃに対する推定対象ユーザｓEのユーザの共感能力θEを推定する推定する。
　図２によれば、推定装置１は、＜学習機能＞及び＜運用機能＞に大別され、両機能に共通して、機械学習エンジン１２を有する。
　推定装置１は、学習機能として、コンテンツ蓄積部１０１と、共感度行列記憶部１０２と、ユーザ特徴量行列記憶部１０３と、項目反応理論算出部１１とを有する。また、運用機能として、ユーザ特徴量行列算出部１３を有する。
　これら機能構成部は、装置に搭載されたコンピュータを機能させるプログラムを実行することによって実現される。また、これら機能構成部の処理の流れは、装置の推定方法として理解できる。 The estimation apparatus 1 according to the present invention performs estimation by estimating the user's empathy ability θE of the estimation target user sE with respect to the content c.
According to FIG. 2, the estimation apparatus 1 is roughly divided into <learning function> and <operation function>, and has a machine learning engine 12 in common for both functions.
The estimation apparatus 1 includes a content storage unit 101, a co-sensitivity matrix storage unit 102, a user feature amount matrix storage unit 103, and an item response theory calculation unit 11 as learning functions. In addition, a user feature matrix calculation unit 13 is provided as an operation function.
These functional components are realized by executing a program that causes a computer installed in the apparatus to function. The processing flow of these functional components can be understood as an apparatus estimation method.
＜学習機能＞
　学習機能は、教師データとしてのコンテンツｃl(ｃ1～ｃL)毎に、複数のユーザｓn(ｓ1～ｓN)の項目反応理論結果（共感能力θn）と、ユーザ特徴量とに基づいて、機械学習エンジン１２内に学習モデルを構築する。 <Learning function>
The learning function is a machine learning engine based on item response theory results (sympathetic ability θn) and user feature values of a plurality of users sn (s1 to sN) for each content cl (c1 to cL) as teacher data. A learning model is built in 12.
［コンテンツ蓄積部１０１］
　コンテンツ蓄積部１０１は、教師データとして、複数のコンテンツ(ｃ1～ｃL)を蓄積したものである。コンテンツとしては、例えばテキスト（記事やブログなど）、画像、映像など、ユーザに閲覧される形式のものであればよい。 [Content accumulation unit 101]
The content storage unit 101 stores a plurality of contents (c1 to cL) as teacher data. The content may be in a format that can be browsed by the user, such as text (articles, blogs, etc.), images, videos, and the like.
［共感度行列記憶部１０２］
　共感度行列記憶部１０２は、ユーザ(ｓ1～ｓN)毎に、各コンテンツｃl(ｃ1～ｃL)の共感度(ｔ11～ｔ1L,～,ｔN1～ｔNL)を対応付けた共感度行列Ｔnlを記憶する。 [Co-sensitivity matrix storage unit 102]
The co-sensitivity matrix storage unit 102 stores a co-sensitivity matrix Tnl in which the co-sensitivities (t11 to t1L,..., TN1 to tNL) of the contents cl (c1 to cL) are associated with each user (s1 to sN). .
　図３は、共感度行列を表す説明図である。 FIG. 3 is an explanatory diagram showing the co-sensitivity matrix.
　図３によれば、教師データとしてのコンテンツｃl毎に、例えばアンケートによって主観評価としての共感度の回答を得ている。例えば、被験者(ｓ1～ｓN)それぞれに、各コンテンツ(ｃ1～ｃL)を明示し、その共感度について例えば３段階(w＝3)の回答を得る。これによって作成された共感度行列Ｔnl(ｔ11～ｔ1L,～,ｔN1～ｔNL)が記憶されている。 According to FIG. 3, for each content cl as teacher data, for example, a questionnaire is used to obtain a co-sensitivity answer as a subjective evaluation. For example, each content (c1 to cL) is clearly indicated for each of the subjects (s1 to sN), and answers of, for example, three levels (w = 3) are obtained for the co-sensitivity. The co-sensitivity matrix Tnl (t11 to t1L,..., TN1 to tNL) created thereby is stored.
［ユーザ特徴量行列記憶部１０３］
　ユーザ特徴量行列記憶部１０３は、ユーザｓn毎に、所定の各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆ11～ｆ1J,～,ｆN1～ｆNJ)を対応付けたユーザ特徴量行列Ｆnj（N行J列）を記憶する。 [User feature matrix storage unit 103]
The user feature amount matrix storage unit 103 associates, for each user sn, a feature amount (f11 to f1J,..., FN1 to fNJ) of each predetermined user feature element (v1 to vJ) with a user feature amount matrix Fnj (N Row J).
　ユーザ特徴要素は、一般的に、ユーザのプロファイル情報（性別、年齢、職業など）であってもよい。特に、ユーザの性格（認識能力、身体能力、知能力、人格力、忍耐力など）に基づくものであることが好ましい。ユーザ特徴要素は、用途に応じた、ユーザの特性を識別するパラメータ要素であればよい。
　また、ユーザ特徴要素は、ユーザｓnに対する複数の質問に基づくものであってもよい。その場合、ユーザ特徴量は、質問に対するユーザｓnの回答に基づくものとなる。 The user characteristic element may generally be user profile information (gender, age, occupation, etc.). In particular, it is preferably based on the personality of the user (recognition ability, physical ability, intellectual ability, personality ability, patience, etc.). The user characteristic element may be a parameter element that identifies the characteristics of the user according to the application.
The user characteristic element may be based on a plurality of questions for the user sn. In this case, the user feature amount is based on the answer of the user sn to the question.
　図４は、ユーザ特徴量行列を表す説明図である。 FIG. 4 is an explanatory diagram showing a user feature matrix.
　図４によれば、被験者となる各ユーザが、複数の質問に対して回答する。質問の内容としては、ユーザの共感性に関するものが好ましい。これによって、ユーザｓn毎に各質問ｑmに対する回答行列（N行M列）が得られる。
　この回答行列は、特異値分解（Singular Value Decomposition）によって、各ユーザの回答がJ次元に圧縮される．圧縮された特徴ベクトルを全てのユーザｓnでまとめた行列Ｕを、ユーザ特徴量行列とする。
　特異値分解とは、線形代数学における、複素数又は実数を成分とする行列に対する行列分解方法である（例えば非特許文献８参照）。階数rの行列Ｘの分解が存在する。
　　　Ｘ＝ＵΣＶT　->　Ｕ＝Ｘ・Ｒ　　Ｒ＝Ｖ・（ＶT・Ｖ）-1・Σ-1
　　　Ｕ：ユーザ特徴量行列 According to FIG. 4, each user who becomes a test subject answers a plurality of questions. The contents of the question are preferably those related to user empathy. As a result, an answer matrix (N rows and M columns) for each question qm is obtained for each user sn.
In this answer matrix, each user's answer is compressed to J dimension by Singular Value Decomposition. A matrix U in which compressed feature vectors are collected by all users sn is defined as a user feature amount matrix.
Singular value decomposition is a matrix decomposition method for a matrix having components of complex numbers or real numbers in linear algebra (see, for example, Non-Patent Document 8). There is a decomposition of the matrix X of rank r.
X = UΣVT-> U = X ・ R R = V ・ (VT ・ V) -1 ・ Σ-1
U: User feature matrix
［項目反応理論算出部１１］
　項目反応理論算出部１１は、ユーザｓn毎に、各コンテンツｃlの共感度行列Ｔnlを入力し、項目反応理論ＩＲＴ(Item Response Theory)に基づく共感能力θnを算出する。 [Item Response Theory Calculation Unit 11]
The item response theory calculation unit 11 inputs the cosensitivity matrix Tnl of each content cl for each user sn, and calculates the empathy ability θn based on the item response theory IRT (Item Response Theory).
　項目反応理論とは、評価項目群へのユーザの応答に基づいて、ユーザ能力θ（及び識別度α・難易度β）を測定するための試験理論をいう（例えば非特許文献７参照）。これは、評価項目に対するユーザの離散的な回答から、確率論的に、ユーザ能力（及び識別度・難易度）を算出する。ユーザに対する試験項目のストックを保守し、複数の試験の難易度を同等とみなすコンピュータ適応型テスト(Computerized Adaptive Testing)に適する。
　　ユーザ能力θl：コンテンツlに対するユーザnの能力の大きさを表す実数値
（識別度ａl：コンテンツlがユーザの共感能力を識別する力を表す実数値）
　（難易度ｂl：コンテンツlの難しさを表す実数値）
　基本的に、ユーザ能力と識別度及び難易度との差を、ロジスティック曲線に当てはめて、回答の確率を求める。例えばコンテンツｃlがユーザｓnにとって共感度が高い場合、その確率は限りなく１に近づき、逆に共感度が低い場合、その確率は限りなく０に近づく。
　２母数ロジスティックモデルの場合、以下のように確率が算出される。
　　　ｐ(θ)＝１／(１＋ｅ－Ｄａ(θ－ｂ)) The item response theory refers to a test theory for measuring the user ability θ (and the identification degree α / the difficulty level β) based on the user's response to the evaluation item group (see, for example, Non-Patent Document 7). This calculates the user ability (and the degree of identification / difficulty) stochastically from the user's discrete answers to the evaluation items. It is suitable for computerized adaptive testing, which maintains a stock of test items for users and regards the difficulty of multiple tests as equivalent.
User ability θl: a real value representing the magnitude of the ability of user n for content l (identification degree al: a real value representing the ability of content l to identify the user's empathy ability)
(Difficulty level bl: real value representing difficulty of content l)
Basically, the difference between the user ability and the degree of discrimination and difficulty is applied to the logistic curve to obtain the probability of the answer. For example, when the content cl has high co-sensitivity for the user sn, the probability approaches 1 as much as possible, and conversely, when the co-sensitivity is low, the probability approaches 0 as much as possible.
In the case of the 2-parameter logistic model, the probability is calculated as follows.
p (θ) = 1 / (1 + e−Da (θ−b))
［学習機能の機械学習エンジン１２］
　学習機能の機械学習エンジン１２は、ユーザｓn毎に、ユーザ特徴量行列Ｆnjと共感能力θnとを対応付けた教師データとして学習する。
　機械学習エンジン１２は、線形回帰、サポートベクターマシン(Support Vector Machine)、又は、深層学習の回帰学習に基づくものであってもよい。 [Machine learning engine 12 with learning function]
The machine learning engine 12 of the learning function learns as teacher data in which the user feature matrix Fnj and the sympathy ability θn are associated with each user sn.
The machine learning engine 12 may be based on linear regression, support vector machine, or deep learning regression learning.
＜運用機能＞
　運用機能は、未知のユーザｓlについて、コンテンツに対する共感能力θlを推定する。 <Operational function>
The operation function estimates the sympathy ability θl for the content for the unknown user sl.
［ユーザ特徴量行列算出部１３］
　ユーザ特徴量行列算出部１３は、推定対象ユーザｓEについて、所定の各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆE1～ｆEJ)を対応付けたユーザ特徴量行列ＦEを算出する。
　ここで、ユーザ特徴量行列算出部１３は、前述したユーザ特徴量行列記憶部１０３と全く同じ方法で、ユーザ特徴量行列ＦEを算出する。 [User feature matrix calculation unit 13]
The user feature quantity matrix calculation unit 13 calculates a user feature quantity matrix FE in which feature quantities (fE1 to fEJ) of predetermined user feature elements (v1 to vJ) are associated with the estimation target user sE.
Here, the user feature amount matrix calculation unit 13 calculates the user feature amount matrix FE in exactly the same manner as the user feature amount matrix storage unit 103 described above.
［運用機能の機械学習エンジン１２］
　運用機能の機械学習エンジン１２は、推定対象ユーザ（未知）ｓEのユーザ特徴量行列ＦEを入力し、共感能力θEを出力する。 [Machine learning engine 12 for operational functions]
The machine learning engine 12 of the operational function inputs the user feature matrix FE of the estimation target user (unknown) sE, and outputs the empathy ability θE.
　図５は、コンテンツに対してユーザが持つ認知度及び情動度を表す説明図である。 FIG. 5 is an explanatory diagram showing the degree of recognition and emotion of the user with respect to the content.
　心理学の知見によれば、共感度は、「認知度」と「情動度」とによって表される。
　　　「共感度」：他者の感情を理解し、他者と同じ感情を持つ人の性質
　　　「認知度」：他者の感情（心理状態）を理解する度合い
　　　「情動度」：他者と同じ感情（代理的な反応の強さ）を持つ度合い
　ここでの「他者」とは、コンテンツの登場人物又は著作者を意味することとなる。 According to the knowledge of psychology, co-sensitivity is expressed by “recognition” and “emotion”.
“Sensitivity”: The nature of a person who understands the emotions of others and has the same feeling as others. “Recognition”: The degree of understanding of others' emotions (psychological state). “Emotion”: The same emotions as others. Degree of having (the strength of proxy response) The “other” here means the character or author of the content.
　図５は、多数のユーザ（被験者）に同一の記事を閲覧してもらい、各ユーザが「興味を引く話題か？」「心に響く話題か？」を質問して得られた、主観評価としての回答結果を示す。被験者からの回答は、以下のとおり「認知度」および「情動度」で表される。
　　　興味を引く話題か？：引かない＝認知度が低い
　　　　　　　　　　　　：引く　　＝認知度が高い
　　　心に響く話題か？　：響かない＝情動度が低い
　　　　　　　　　　　　：響く　　＝情動度が高い
　図５の表によれば、同一の記事に対して、「興味を引く」と感じる一方で「心に響かない」と感じるユーザがおり、また、「興味を引かない」と感じる一方で「心に響く」と感じるユーザもいる。 FIG. 5 shows a subjective evaluation obtained by asking many users (subjects) to read the same article and asking each user “a topic that attracts interest” or “a topic that resonates”. The answer result is shown. The answers from the subjects are expressed by “recognition” and “emotional” as follows.
Is it an interesting topic? : Don't pull = Low recognition: Pull = High recognition Is it a reverberating topic? : Not resonating = Low emotional level: Resonating = High emotional level According to the table in Fig. 5, there are users who feel "not interested" but feel "not resonate" for the same article. There are also users who feel “not interested” but feel “resonate”.
　図６は、共感度として認知度及び情動度を適用した説明図である。
　図７は、共感度行列として認知度及び情動度を適用した説明図である。 FIG. 6 is an explanatory diagram in which recognition and emotion are applied as co-sensitivity.
FIG. 7 is an explanatory diagram in which recognition and emotion are applied as the co-sensitivity matrix.
　図６に示されるように、共感度行列記憶部１０２は、図２と比較して、共感度行列Ｔとして、認知度Ｔ1と情動度Ｔ2とが別々に記憶されている。
　図７に示されるように、被験者群 (ｓ1～ｓN)のN人にそれぞれに、各コンテンツ(ｃ1～ｃL)を明示し、その認知度及び情動度について例えば３段階(w＝3)の回答を得る。これによって、評価項目数＝2の共感度行列Ｔ1nl（ｔ111～ｔ11L,～,ｔ1N1～ｔ1NL)及びＴ2nl（ｔ211～ｔ21L,～,ｔ2N1～ｔ2NL)が、項目反応理論算出部１１へ入力される。 As shown in FIG. 6, the co-sensitivity matrix storage unit 102 stores the recognition degree T1 and the emotional degree T2 separately as the co-sensitivity matrix T as compared with FIG.
As shown in FIG. 7, each content (c1 to cL) is clearly shown to each of N persons in the subject group (s1 to sN), and for example, three levels (w = 3) of responses to the degree of recognition and emotion Get. Accordingly, the co-sensitivity matrices T1nl (t111 to t11L,..., T1N1 to t1NL) and T2nl (t211 to t21L,..., T2N1 to t2NL) with the number of evaluation items = 2 are input to the item response theory calculation unit 11.
　これに対し、項目反応理論算出部１１は、２母数ロジスティックモデルとして、ユーザｓn(ｓ1～ｓN）について、以下のように共感能力θを算出する。
　　　認知度の共感能力：θ11～θ1N
　　　情動度の共感能力：θ21～θ2N
　尚、一部のユーザが、ランダムに回答することができる場合、項目反応理論算出部１１は、３母数ロジスティックモデルとして算出するものであってもよい。 On the other hand, the item response theory calculation unit 11 calculates the empathy ability θ as follows for the user sn (s1 to sN) as a two-parameter logistic model.
Awareness of awareness: θ11 ～ θ1N
Emotional empathy ability: θ21 ～ θ2N
In addition, when a part of users can answer at random, the item reaction theory calculation part 11 may be calculated as a 3 parameter logistic model.
　機械学習エンジン１２は、コンテンツｃl毎に、特徴量行列Ｆlと、認知度に対する識別度α11及び難易度β11と、情動度に対する識別度α12及び難易度β12とを対応付けた教師データによって学習する。そして、以下のように学習モデルを構築する。
　　　認知度の共感能力θ11に対する学習モデル：Ｍ1
　　　情動度の共感能力θ12に対する学習モデル：Ｍ2 The machine learning engine 12 learns, for each content cl, teacher data in which the feature quantity matrix Fl, the identification degree α11 and the difficulty level β11 for the recognition degree, and the identification degree α12 and the difficulty level β12 for the emotion degree are associated with each other. Then, a learning model is constructed as follows.
Learning model for awareness of sympathetic ability θ11: M1
Learning model for emotional empathy ability θ12: M2
　次に、機械学習エンジン１２は、推定対象ユーザｓEの特徴量行列ＦEを入力すると、識別度に対する共感能力θ1Eと、情動度に対する共感能力θ2Eとを出力する。 Next, when the feature quantity matrix FE of the estimation target user sE is input, the machine learning engine 12 outputs the sympathy ability θ1E for the discrimination degree and the sympathy ability θ2E for the emotion degree.
　図８Ａ、図８Ｂ、図８Ｃおよび図８Ｄは、コンテンツ毎の項目反応カテゴリ特性曲線を表す説明図である。 8A, 8B, 8C, and 8D are explanatory diagrams showing item response category characteristic curves for each content.
　項目反応カテゴリ特性曲線（ＩＲＣＣＣ(Item Response Category Characteristic Curve)）は、以下のように作成される。
　　　横軸：共感能力値（左側：共感度が低い、右側：共感度が高い）
　　　　　　　　　　　（左側：識別度が低い、右側：識別度が高い）
　　　　　　　　　　　（左側：情動度が低い、右側：情動度が高い）
　　　縦軸：確率ｐ
　　　曲線：各選択肢（「共感しない」「わからない」「共感する」）
　　　　　　　　　　（「興味を引かない」「わからない」「興味を引く」）
　　　　　　　　　　（「心に響かない」「わからない」「心に響かない」） An item response category characteristic curve (IRCCC) is created as follows.
Horizontal axis: empathy ability value (left side: low sensitivity, right side: high sensitivity)
(Left side: low discrimination, right side: high discrimination)
(Left side: low emotion level, right side: high emotion level)
Vertical axis: probability p
Curve: each choice ("do not sympathize""do not know""sympathize")
("I'm not interested", "I don't know", "I'm interested")
("I don't resonate""Idon'tknow""Idon'tresonate")
　図８Ａ～図８Ｄの項目反応カテゴリ特性曲線は、100件のコンテンツ(ｃ1～ｃ100)について、300名のユーザ(ｓ1～ｓ300)に対する認知度及び情動度の共感度行列Ｔ1及びＴ2と、８次元のユーザ特徴量行列Ｆとから得られたものである。
　難易度ｂは、共感能力θと同じスケール上にある。識別度ａは、項目反応カテゴリ特性曲線の傾きを決定する。曲線の傾きが大きいほど、難易度ｂと共感能力θとの差が大きいほど、回答がくっきり分かれることを表す。 The item response category characteristic curves of FIGS. 8A to 8D show the co-sensitivity matrices T1 and T2 of the recognition degree and emotion degree for 300 users (s1 to s300) and eight dimensions for 100 contents (c1 to c100). Obtained from the user feature matrix F.
The difficulty level b is on the same scale as the empathy ability θ. The degree of discrimination a determines the slope of the item response category characteristic curve. The greater the slope of the curve, the clearer the answer is, the greater the difference between the difficulty level b and the empathy ability θ.
　図８Ａ～図８Ｄによれば、４件のコンテンツ（ID36,ID72,ID82,ID94）について、認知度及び情動度の共感度行列それぞれの項目反応カテゴリ特性曲線が表されている。図８Ａに示されたコンテンツID82及び図８Ｂに示されたコンテンツID94は、図８Ｄに示されたコンテンツID72と比較して、認知度が高いことがわかる。また、図８Ａに示されたコンテンツID82は、図８Ｂに示されたコンテンツID94及び図８Ｄに示されたコンテンツID72と比較して、情動度が高いことがわかる。
　これによって、図８Ａに示されたコンテンツID82は、認知度が高く且つ情動度も高いことがわかる。また、図８Ｂに示されたコンテンツID94及び図ＤＡに示されたコンテンツID72は、情動度が低いことがわかる。
　尚、あるコンテンツについて共感度が低いと推定された図８Ｄに示されたコンテンツID72についても、ユーザの共感能力が高い場合、高い共感度が得られる。 According to FIGS. 8A to 8D, the item response category characteristic curves of the co-sensitivity matrices of the recognition degree and the emotion degree are represented for the four contents (ID36, ID72, ID82, ID94). It can be seen that the content ID 82 shown in FIG. 8A and the content ID 94 shown in FIG. 8B have higher recognition than the content ID 72 shown in FIG. 8D. Further, it can be seen that the content ID 82 shown in FIG. 8A has higher emotion than the content ID 94 shown in FIG. 8B and the content ID 72 shown in FIG. 8D.
Accordingly, it can be seen that the content ID 82 shown in FIG. 8A has high recognition and high emotion. Further, it can be seen that the content ID 94 shown in FIG. 8B and the content ID 72 shown in FIG.
Note that the content ID 72 shown in FIG. 8D in which the co-sensitivity of a certain content is estimated to be low can also be obtained when the user's empathy capability is high.
　以上、詳細に説明したように、本発明のプログラム、推定装置及び方法によれば、コンテンツに対するユーザの共感能力を推定することができる。 As described above in detail, according to the program, the estimation apparatus, and the method of the present invention, it is possible to estimate the user's sympathy ability with respect to the content.
　前述した本発明の種々の実施形態について、本発明の技術思想及び見地の範囲の種々の変更、修正及び省略は、当業者によれば容易に行うことができる。前述の説明はあくまで例であって、何ら制約しようとするものではない。本発明は、特許請求の範囲及びその均等物として限定するものにのみ制約される。 For the various embodiments of the present invention described above, various changes, modifications, and omissions in the technical idea and scope of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.
　１　推定装置
　１０１　コンテンツ蓄積部
　１０２　共感度行列記憶部
　１０３　ユーザ特徴量行列記憶部
　１１　項目反応理論算出部
　１２　機械学習エンジン
　１３　ユーザ特徴量行列算出部
　２　Ｗｅｂサーバ
　３　端末 DESCRIPTION OF SYMBOLS 1 Estimation apparatus 101 Content storage part 102 Cosensitivity matrix memory | storage part 103 User feature-value matrix memory | storage part 11 Item reaction theory calculation part 12 Machine learning engine 13 User feature-value matrix calculation part 2 Web server 3 Terminal

Claims (7)

  1. 　コンテンツｃに対する推定対象ユーザｓEのユーザの共感能力θEを推定する装置に搭載されたコンピュータを機能させるプログラムであって、前記プログラムは、
    　学習機能を実現するために、前記コンピュータを
    　ユーザｓn(ｓ1～ｓN)毎に、各コンテンツｃl(ｃ1～ｃL)の共感度(ｔ11～ｔ1L,～,ｔN1～ｔNL)を対応付けた共感度行列Ｔnlを記憶する共感度行列記憶手段と、
    　前記ユーザｓn毎に、所定の各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆ11～ｆ1J,～,ｆN1～ｆNJ)を対応付けたユーザ特徴量行列Ｆnjを記憶するユーザ特徴量行列記憶手段と、
    　前記ユーザｓn毎に、前記各コンテンツｃlの前記共感度行列Ｔnlを入力し、項目反応理論に基づく共感能力θnを算出する項目反応理論算出手段と、
    　前記ユーザｓn毎に、前記ユーザ特徴量行列Ｆnjと前記共感能力θnとを対応付けた教師データとして学習する機械学習エンジンと
    して機能させ、
    　運用機能を実現するために、前記コンピュータを
    　前記推定対象ユーザｓEについて、所定の前記各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆE1～ｆEJ)を対応付けたユーザ特徴量行列ＦEを算出するユーザ特徴量行列算出手段として機能させると共に、
    　前記機械学習エンジンは、前記推定対象ユーザｓEの前記ユーザ特徴量行列ＦEを入力し、共感能力θEを出力するように機能させるプログラム。     A program for causing a computer installed in a device for estimating the user's sympathy ability θE of the estimation target user sE to the content c to function,
    In order to realize the learning function, the computer is associated with the sensitivity (t11 to t1L,..., TN1 to tNL) of each content cl (c1 to cL) for each user sn (s1 to sN). A co-sensitivity matrix storage means for storing Tnl;
    User feature value matrix storage means for storing a user feature value matrix Fnj in which feature values (f11 to f1J,..., FN1 to fNJ) of predetermined user feature elements (v1 to vJ) are associated with each user sn. ,
    Item response theory calculation means for inputting the empathy matrix Tnl of each content cl for each user sn and calculating the sympathy ability θn based on the item response theory;
    For each user sn, function as a machine learning engine that learns as teacher data in which the user feature matrix Fnj and the sympathetic ability θn are associated,
    In order to realize the operation function, the computer calculates a user feature matrix FE in which the feature quantities (fE1 to fEJ) of the predetermined user feature elements (v1 to vJ) are associated with the estimation target user sE. While functioning as a user feature matrix calculation means,
    The machine learning engine is a program that inputs the user feature matrix FE of the estimation target user sE and functions to output the empathy ability θE.
  2. 　前記共感度は、
    　　前記コンテンツの登場人物又は著作者の感情を理解する認知度ｔRと、
    　　前記コンテンツの登場人物又は著作者と同じ感情を生じる情動度ｔEと
    からなり、
    　前記項目反応理論算出手段は、２母数ロジスティックモデルとして前記共感能力θnを算出するようにコンピュータを機能させる請求項１に記載のプログラム。     The co-sensitivity is
    A degree of recognition tR for understanding the emotion of the characters or authors of the content;
    The emotional level tE that produces the same emotion as the character or author of the content,
    The program according to claim 1, wherein the item response theory calculation means causes a computer to calculate the empathy ability θn as a two-parameter logistic model.
  3. 　前記ユーザ特徴要素は、前記ユーザｓnに対する複数の質問に基づくものであり、
    　前記特徴量は、前記質問に対する前記ユーザｓnの回答に基づくものである
    ようにコンピュータを機能させる請求項１又は２に記載のプログラム。     The user characteristic element is based on a plurality of questions for the user sn;
    The program according to claim 1 or 2, which causes a computer to function so that the feature amount is based on an answer of the user sn to the question.
  4. 　前記コンテンツは、不特定多数の第三者からＷｅｂサーバに投稿されたものであり、
    　前記コンテンツ影響力Ｍは、ユーザがＷｅｂサーバの当該コンテンツを閲覧した際に受ける感情に基づくものである
    ようにコンピュータを機能させる請求項１から３のいずれか１項に記載の質問回答プログラム。     The content is posted to a web server from an unspecified number of third parties,
    The question answering program according to any one of claims 1 to 3, wherein the content influence M causes the computer to function based on an emotion received when the user browses the content on the Web server.
  5. 　前記機械学習エンジンは、線形回帰、サポートベクターマシン、又は、深層学習の回帰学習に基づくものであるようにコンピュータを機能させる請求項１から４のいずれか１項に記載のプログラム。 The program according to any one of claims 1 to 4, wherein the machine learning engine causes a computer to function based on linear regression, support vector machine, or deep learning regression learning.
  6. 　コンテンツｃに対する推定対象ユーザｓEのユーザの共感能力θEを推定する装置であって、
    　学習機能として、
    　ユーザｓn(ｓ1～ｓN)毎に、各コンテンツｃl(ｃ1～ｃL)の共感度(ｔ11～ｔ1L,～,ｔN1～ｔNL)を対応付けた共感度行列Ｔnlを記憶する共感度行列記憶手段と、
    　前記ユーザｓn毎に、所定の各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆ11～ｆ1J,～,ｆN1～ｆNJ)を対応付けたユーザ特徴量行列Ｆnjを記憶するユーザ特徴量行列記憶手段と、
    　前記ユーザｓn毎に、前記各コンテンツｃlの前記共感度行列Ｔnlを入力し、項目反応理論に基づく共感能力θnを算出する項目反応理論算出手段と、
    　前記ユーザｓn毎に、前記ユーザ特徴量行列Ｆnjと前記共感能力θnとを対応付けた教師データとして学習する機械学習エンジンと
    を有し、
    　運用機能として、
    　前記推定対象ユーザｓEについて、所定の前記各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆE1～ｆEJ)を対応付けたユーザ特徴量行列ＦEを算出するユーザ特徴量行列算出手段を有し、
    　前記機械学習エンジンは、前記推定対象ユーザｓEの前記ユーザ特徴量行列ＦEを入力し、共感能力θEを出力する装置。     An apparatus for estimating the user's empathy ability θE of the estimation target user sE for the content c,
    As a learning function,
    A sensitivity matrix storage means for storing a sensitivity matrix Tnl in which the sensitivity (t11 to t1L,..., TN1 to tNL) of each content cl (c1 to cL) is associated for each user sn (s1 to sN);
    User feature value matrix storage means for storing a user feature value matrix Fnj in which feature values (f11 to f1J,..., FN1 to fNJ) of predetermined user feature elements (v1 to vJ) are associated with each user sn. ,
    Item response theory calculation means for inputting the empathy matrix Tnl of each content cl for each user sn and calculating the sympathy ability θn based on the item response theory;
    A machine learning engine that learns teacher data associating the user feature matrix Fnj and the sympathetic ability θn for each user sn;
    As an operation function,
    User feature value matrix calculating means for calculating a user feature value matrix FE in which feature values (fE1 to fEJ) of predetermined user feature elements (v1 to vJ) are associated with the estimation target user sE;
    The machine learning engine is an apparatus that inputs the user feature matrix FE of the estimation target user sE and outputs sympathetic ability θE.
  7. 　コンテンツｃに対する推定対象ユーザｓEのユーザの共感能力θEを推定する装置の推定方法であって、
    　学習機能として、
    　ユーザｓn(ｓ1～ｓN)毎に、各コンテンツｃl(ｃ1～ｃL)の共感度(ｔ11～ｔ1L,～,ｔN1～ｔNL)を対応付けた共感度行列Ｔnlを共感度行列記憶部に記憶し、
    　前記ユーザｓn毎に、所定の各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆ11～ｆ1J,～,ｆN1～ｆNJ)を対応付けたユーザ特徴量行列Ｆnjをユーザ特徴量行列記憶部に記憶し、
    　前記ユーザｓn毎に、前記各コンテンツｃlの共感度行列Ｔnlを入力し、項目反応理論に基づく共感能力θnを算出し、
    　機械学習エンジンを用いて、前記ユーザｓn毎に、前記ユーザ特徴量行列Ｆnjと前記共感能力θnとを対応付けた教師データとして学習し、
    　運用機能として、
    　前記推定対象ユーザｓEについて、所定の前記各ユーザ特徴要素(ｖ1～ｖJ)の特徴量(ｆE1～ｆEJ)を対応付けたユーザ特徴量行列ＦEを算出し、
    　前記機械学習エンジンが、前記推定対象ユーザｓEの前記ユーザ特徴量行列ＦEを入力し、共感能力θEを出力する
    　推定方法。     An apparatus estimation method for estimating a user's empathy ability θE of an estimation target user sE for content c,
    As a learning function,
    For each user sn (s1 to sN), a cosensitivity matrix Tnl in which the cosensitivities (t11 to t1L,..., TN1 to tNL) of the contents cl (c1 to cL) are associated is stored in the cosensitivity matrix storage unit.
    A user feature quantity matrix Fnj in which feature quantities (f11 to f1J,..., FN1 to fNJ) of predetermined user feature elements (v1 to vJ) are associated with each user sn is stored in a user feature quantity matrix storage unit. ,
    For each user sn, the sympathy matrix Tnl of each content cl is input, and the sympathy ability θn based on the item reaction theory is calculated,
    Using a machine learning engine, for each user sn, learning as teacher data in which the user feature matrix Fnj and the empathy ability θn are associated with each other,
    As an operation function,
    For the estimation target user sE, a user feature value matrix FE that associates feature values (fE1 to fEJ) of the predetermined user feature elements (v1 to vJ) is calculated,
    An estimation method in which the machine learning engine inputs the user feature matrix FE of the estimation target user sE and outputs a sympathetic ability θE.
PCT/JP2019/006283 2018-03-01 2019-02-20 Program, device and method for estimating empathetic influence of users with respect to content WO2019167748A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2018037001A JP6663945B2 (en) 2018-03-01 2018-03-01 Program, apparatus, and method for estimating user's empathy for content
JP2018-037001 2018-03-01

Publications (1)

Publication Number Publication Date
WO2019167748A1 true WO2019167748A1 (en) 2019-09-06

Family

ID=67804983

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2019/006283 WO2019167748A1 (en) 2018-03-01 2019-02-20 Program, device and method for estimating empathetic influence of users with respect to content

Country Status (2)

Country Link
JP (1) JP6663945B2 (en)
WO (1) WO2019167748A1 (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023209962A1 (en) * 2022-04-28 2023-11-02 日本電信電話株式会社 Training apparatus, estimation apparatus, training method, estimation method, training program, and estimation program

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20080083248A (en) * 2008-08-25 2008-09-17 이종운 Online item recommendation system built upon users' "like or dislike"choices - based on statistical similarity between users' choices
JP2009223423A (en) * 2008-03-13 2009-10-01 Fujitsu Ltd User state prediction system, user state estimation method, and user state prediction program
US20140156582A1 (en) * 2012-11-30 2014-06-05 Jayson Holliewood Cornelius Item Response Methods as Applied to a Dynamic Content Distribution System and Methods

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2009223423A (en) * 2008-03-13 2009-10-01 Fujitsu Ltd User state prediction system, user state estimation method, and user state prediction program
KR20080083248A (en) * 2008-08-25 2008-09-17 이종운 Online item recommendation system built upon users' "like or dislike"choices - based on statistical similarity between users' choices
US20140156582A1 (en) * 2012-11-30 2014-06-05 Jayson Holliewood Cornelius Item Response Methods as Applied to a Dynamic Content Distribution System and Methods

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
HASHIMA, TOMOHIRO: "Suggestion for Recommended Contents on Website with Item Response Theory", 4TH VOLUME OF GENERAL LECTURE PROCEEDINGS OF 6TH FORUM ON INFORMATION TECHNOLOGY (FIT)2007, NETWORK-SECURITY UBIQUITOUS-MOBILE COMPUTING EDUCATION-HUMANITIES INFORMATION SYSTEM, August 2007 (2007-08-01) *
MATSUMOTO, KAZUNORI ET AL.: "Modeling of Affection on User's Empathy Caused by Text Content", IEICE TECHNICAL REPORT, vol. 117, no. 509, 6 March 2018 (2018-03-06), pages 131 - 137 *
TOKUNAGA, MASAKAZU ET AL.: "The Availability of Item Response Theory in Recommendation System", PROCEEDINGS OF THE 4TH FORUM ON DATA ENGINEERING AND INFORMATION MANAGEMENT (THE 10TH ANNUAL CONFERENCE OF THE DATABASE SOCIETY OF JAPAN, vol. A5-4, 13 July 2012 (2012-07-13), XP055635084, Retrieved from the Internet <URL:http://db-event.jpn.org/deim2012/proceedings/final-pdf/a5-4.pdf> [retrieved on 20120830] *

Also Published As

Publication number Publication date
JP2019153014A (en) 2019-09-12
JP6663945B2 (en) 2020-03-13

Similar Documents

Publication Publication Date Title
Himelboim et al. Valence-based homophily on Twitter: Network analysis of emotions and political talk in the 2012 presidential election
Colleoni CSR communication strategies for organizational legitimacy in social media
Banerjee et al. Using supervised learning to classify authentic and fake online reviews
US20210157872A1 (en) Method and system to filter out unwanted content from incoming social media data
Kaiser et al. Mining consumer dialog in online forums
Stern et al. A network perspective on intermedia agenda-setting
US20230419347A1 (en) Systems and methods for machine learning predictions of the impact of digital content
Waddell Attribution practices for the man-machine marriage: How perceived human intervention, automation metaphors, and byline location affect the perceived bias and credibility of purportedly automated content
Keller et al. Recommender systems for museums: evaluation on a real dataset
Khan et al. Comparative analysis on Facebook post interaction using DNN, ELM and LSTM
Tran et al. Detection and prediction of users attitude based on real-time and batch sentiment analysis of facebook comments
WO2019167742A1 (en) Program, device and method for estimating empathetic influence of content among users
WO2019167748A1 (en) Program, device and method for estimating empathetic influence of users with respect to content
KR101730340B1 (en) Method for quantification of evaluation result about target of evaluation
Sayal et al. Visual sentiment analysis using machine learning for entertainment Applications
Mendes et al. Do users write about the system in use? An investigation from messages in Natural Language on Twitter
Li et al. Correlating stressor events for social network based adolescent stress prediction
Vaishanvi et al. Product Recommendation Using Sentiment Analysis
Kozyreva et al. Network-oriented modeling of multi-criteria homophily and opinion dynamics in social media
Weber et al. Prospect for knowledge in survey data: an artificial neural network sensitivity analysis
Vedenyapin et al. The disruptometer: an artificial intelligence algorithm for market insights
Alowisheq et al. Investigating the relationship between trust and sentiment agreement in arab twitter users
Wickramaarachchi et al. Influence of smart interactive advertising based on age and gender: A case study from sri lanka
Ariel et al. Online agenda-setting research: Challenges and dilemmas
Rodavia et al. AutoRec: A recommender system based on social media stream

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19761597

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19761597

Country of ref document: EP

Kind code of ref document: A1