WO2021145228A1 - Dispositif de traitement d'informations et procédé de traitement d'informations - Google Patents

Dispositif de traitement d'informations et procédé de traitement d'informations Download PDF

Info

Publication number
WO2021145228A1
WO2021145228A1 PCT/JP2021/000017 JP2021000017W WO2021145228A1 WO 2021145228 A1 WO2021145228 A1 WO 2021145228A1 JP 2021000017 W JP2021000017 W JP 2021000017W WO 2021145228 A1 WO2021145228 A1 WO 2021145228A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
information processing
candidate
input
user
Prior art date
Application number
PCT/JP2021/000017
Other languages
English (en)
Japanese (ja)
Inventor
悠希 武田
Original Assignee
ソニーグループ株式会社
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 ソニーグループ株式会社 filed Critical ソニーグループ株式会社
Publication of WO2021145228A1 publication Critical patent/WO2021145228A1/fr

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
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri

Definitions

  • the present technology relates to an information processing device and an information processing method, and more particularly to an information processing device and an information processing method capable of efficiently collecting data desired by a user.
  • Patent Document 1 discloses an inter-vocabulary relationship estimation device that defines a relationship between a plurality of vocabularies to be processed by machine learning using existing inter-vocabulary relationship data. This technique can also be applied to tasks such as collecting closely related and similar data.
  • This technology was made in view of such a situation, and enables the user to efficiently collect the desired data.
  • the information processing apparatus of the present technology includes a control unit that outputs candidate data that is a candidate for giving the same label as the input data input by the user based on the learning model, and the control unit is included in the candidate data.
  • This is an information processing device that outputs further the candidate data based on the learning model learned by the user using the selection data selected as the target for assigning the label.
  • the information processing apparatus outputs candidate data that can be given the same label as the input data input by the user based on the learning model, and the user selects the candidate data from the candidate data.
  • candidate data that is a candidate for assigning the same label as the input data input by the user is output, and the candidate data is selected by the user as the target for assigning the label. Further candidate data is output based on the learning model trained using the selected data.
  • the word data handled in this technology includes "words” such as "cat"-"Animal”, “apple”-"Fruit”, “festival”-”Event”, etc. It is described by a pair of “labels” that represent the concept to which the word belongs.
  • word data as shown in FIG. 1 is indispensable. However, collecting such word data is time consuming and costly.
  • Collecting word data here means giving a label of the concept to which each word belongs to a large number of words collected by some method.
  • an algorithm that infers the concept to which the word belongs and assigns the corresponding label is required.
  • the required word data structure may differ depending on the type of voice agent you want to create. For example, as shown in FIG. 2, when creating a voice agent 11 that supports cooking, the word “fish” should be labeled with “ingredients”. On the other hand, when creating a voice agent 12 for guidance in an aquarium, the word “fish” should be labeled as "creature”.
  • candidate data close to the input data input by the user is output, and further candidate data is output by learning using the selection data selected by the user from the candidate data. It enables the user to efficiently collect the desired data.
  • FIG. 3 is a block diagram showing a configuration example of an information processing device to which the present technology is applied.
  • the information processing device 100 of FIG. 3 is composed of a personal computer operated by a user, a tablet terminal, a computer terminal such as a smartphone, and the like.
  • the information processing device 100 is composed of an input unit 110, a presentation unit 120, a storage unit 130, and a control unit 140.
  • the input unit 110 is composed of buttons, a keyboard, a mouse, a microphone, etc., and accepts user operations.
  • the input information according to the user's operation is input to the control unit 140.
  • the presentation unit 120 is composed of an organic EL (Electro-Luminescence) display, a liquid crystal display, a speaker, and the like, and presents data and the like output by the control unit 140.
  • the input unit 110 may be composed of a touch panel formed integrally with the display as the presentation unit 120.
  • the storage unit 130 is composed of a non-volatile memory such as a flash memory, and stores a program and various data read by the control unit 140.
  • the control unit 140 is composed of a processor such as a CPU (Central Processing Unit) and a memory.
  • the control unit 140 includes a learning unit 141.
  • the learning unit 141 is realized by executing a predetermined program read from the storage unit 130 by the CPU constituting the control unit 140.
  • a predetermined number of word data P1s labeled with "Fruit" are input to the learning model 141a as training data from an existing word data list.
  • the input of the word data P1 is accepted by the input unit 110.
  • word data N1 with the label "Animal”, word data N2 with the label “Vegetable”, and the like are word data.
  • the same number as P1 is randomly extracted.
  • the word data N1 and the word data N2 are also input to the learning model 141a as learning data.
  • the learning unit 141 learns a learning model 141a that embeds words (converts words into vectors) using an algorithm such as word2vec, using word data P1 as a positive example and word data N1 and word data N2 as negative examples. Let me.
  • the learning unit 141 classifies the unlabeled word data WD extracted from the vocabulary database VDB using the binary classifier 141b based on the feature amount (vector of each word) which is the learning result of the learning model 141a.
  • the unlabeled word data WD classified on the positive example side is output to the presentation unit 120 as a candidate data CD, and is presented by the presentation unit 120.
  • the candidate data CD is word data having a vector close to the word data P1.
  • the learning unit 141 causes the learning model 141a to input the selection data SD selected by the user from the candidate data CDs presented by the presentation unit 120 as regular learning data.
  • the learning unit 141 repeats the above process until instructed by the user.
  • the learning unit 141 learns the input data (word data) supplied from the input unit 110 using the learning model, and outputs candidate data (unlabeled word data) close to the input data to the presentation unit 120. do. Further, the learning unit 141 learns again using the candidate data selected by the user from the candidate data presented to the presentation unit 120, and outputs new candidate data to the presentation unit 120.
  • step S11 the input unit 110 accepts the input of word data to which the label of the attribute that the user wants to collect is attached.
  • step S12 the control unit 140 (learning unit 141) provides the presentation unit 120 with candidate data that is a candidate for giving the same label as the input word data based on the learning model trained using the input data. Output.
  • step S13 the presentation unit 120 presents the candidate data output from the control unit 140.
  • step S14 the control unit 140 determines whether or not the instruction to end the collection of word data has been received by the user's operation on the input unit 110.
  • step S15 When the input unit 110 is not instructed to end the collection of word data, but is accepted to select a predetermined candidate data as a label assignment target from the candidate data presented by the presentation unit 120. , The process proceeds to step S15.
  • step S15 the control unit 140 (learning unit 141) adds the selected candidate data, which is the selected data, to the positive example, and the non-selected data, which is the non-selected candidate data, to the negative example.
  • step S16 the control unit 140 (learning unit 141) trains the learning model using the selected data as a positive example and the non-selected data as a negative example. After that, the process returns to step S12, and the subsequent processes are repeated.
  • 6 to 8 are diagrams showing an example of presentation of candidate data in the presentation unit 120.
  • the screens shown in FIGS. 6 to 8 are provided with a word data input area 151, a send button 152, and a candidate data display area 153-1, 153-2, 153, ...
  • word data normal learning data
  • word data normal learning data
  • FIG. 6 “cake”, “pudding”, and “cookie” labeled as “confectionery” are input as word data in the word data input area 151.
  • the selection buttons 162P and 162N are displayed.
  • the selection button 162P is a button for using the candidate data 161 as selection data (normal learning data).
  • the selection button 162N is a button for using the candidate data 161 as non-selection data (learning data of a negative example).
  • “chocolate”, “ramune”, and “curry” are used as candidate data 161 for assigning a “confectionery” label to each of the candidate data display areas 153-1, 153-2, and 153-3. Is displayed. The score of the binary classification is shown on the right side of each of the candidate data 161. The closer this score is to 1.0, the more the corresponding candidate data 161 is classified on the positive example side.
  • one of the selection buttons 162P and 162N is selected in each of the candidate data display areas 153-1, 153-2, 153-3 from the state of FIG.
  • the selection button 162P having "chocolate” of the candidate data 161 as the selection data is selected, and in the candidate data display area 1532, the "ramune” of the candidate data 161 is selected.
  • the selection button 162P with "" as the selection data is selected.
  • the selection button 162N having "curry" of the candidate data 161 as non-selection data is selected.
  • the word data input area 151 is newly labeled with “confectionery” in addition to the “cake”, “pudding”, and “cookie” input in the state of FIG.
  • the pudding "chocolate” and “ramune” are entered.
  • step S14 it is determined that the instruction to end the collection of the word data has been received, such as by determining that the user has collected a satisfactory amount of word data.
  • the process ends.
  • the labeled selection data, that is, the collected word data is stored in, for example, the storage unit 130.
  • candidate data close to the word data input by the user is presented, and further candidate data is presented by learning using the selection data selected from the candidate data.
  • the time required to output word data can be shortened. This makes it possible to improve the accuracy of word data proposals by interactive learning without stressing the user.
  • Second embodiment weighting for each learning model>
  • the context in which word data should be collected may differ. For example, when creating a voice agent that introduces news, you want to handle many hard-expressing words used in news, but when creating a voice agent that is a friendly conversation partner, you want to use buzzwords and broken expressions. I want to handle many words.
  • steps S21, S23 to S26 in the flowchart of FIG. 9 is basically the same as the processing of steps S11, S13 to S16 in the flowchart of FIG. 5, so the description thereof will be omitted.
  • step S22 the control unit 140 (learning unit 141) provides candidate data as a candidate to be given the same label as the input word data for each learning model based on a plurality of learning models prepared in advance. , Is output to the presentation unit 120. At this time, the same number of candidate data is output for each learning model.
  • step S27 the control unit 140 (learning unit 141) changes the weighting of the number of candidate data output for each learning model according to the selected candidate data (selected data). After that, the process returns to step S22, and the subsequent processes are repeated.
  • FIG. 10 is a diagram illustrating weighting of the number of candidate data for each learning model.
  • a word2vec model 210 learned from the contents of a newspaper article and a word2vec model 220 learned from the contents of an SNS post are prepared.
  • the same number (for example, 5) of candidate data 231 is output from each of the word2vec models 210 and 220.
  • the weighting of the number of candidate data output from each of the word2vec models 210 and 220 is changed. Specifically, for the word2vec model in which more selection data is output, the weighting is changed so that more candidate data is output.
  • seven candidate data 232 are output from the word2vec model 210, and three candidate data 232 are output from the word2vec model 220.
  • the weighting of the number of candidate data output for each learning model is changed according to the selection data selected by the user, so that the word data in the context desired by the user can be collected. Is possible.
  • word data of one language for example, Japanese
  • word data of one language for example, Japanese
  • FIG. 11 is a block diagram showing a configuration example of the information processing apparatus of the present embodiment.
  • the information processing device 100 of FIG. 11 is different from the information processing device 100 of FIG. 3 in that the control unit 140 includes a translation processing unit 311 in addition to the learning unit 141.
  • the translation processing unit 311 translates the input data (word data) input by the user and the candidate data (selected data) to which the same label as the input data is given into a predetermined language.
  • Language translation is performed, for example, by using a dictionary of the corresponding language stored in the storage unit 130.
  • steps S31 to S36 in the flowchart of FIG. 12 is basically the same as the processing of steps S11 to S16 in the flowchart of FIG. 5, so the description thereof will be omitted.
  • step S37 the translation processing unit 311 selects the regular word data (that is, the input data and the selection data). , Translate into the given language.
  • the word data 330 when word data 330 belonging to a fruit is collected as regular word data, the word data 330 includes Chinese word data 330C, English word data 330E, and Russian. Is translated into the word data 330R of.
  • the man-hours for collecting multilingual word data can be shortened, and the cost and time required for developing a voice agent corresponding to multiple languages can be reduced.
  • the user selects the word data on the screen as shown in FIGS. 6 to 8 to give the word data a label.
  • word data may be selected on the screen 350 for providing the quiz game as shown in FIG. 14, and a label may be given to the word data.
  • buttons 361 to 364 for selecting "Dorian”, “Tomato”, “Table”, and “Chair”, which are candidates for the word of the attribute (fruit), are displayed.
  • the user selects the button of the word that is considered to be a fruit so that the user answers the question "Which fruit?", And the selected word is given a fruit label.
  • word data is input as input data, but data in other formats may be input.
  • image data may be input as input data.
  • candidate data candidate image data to be given the same label as the input image data is presented, and further image data is presented by learning using the image data selected from the candidate data.
  • image data close to the input image data can be efficiently collected, and for example, the accuracy of similar image search can be improved.
  • sensor data obtained from various sensors can be input.
  • the acceleration data obtained from the acceleration sensor or the angular acceleration data obtained from the gyro sensor may be input as the input data.
  • candidate data sensor data that gives the same label as the input sensor data is presented as acceleration or rotation in, for example, VR (Virtual Reality) goggles. Further sensor data is presented by learning using the sensor data selected from them.
  • FIG. 15 is a block diagram showing a functional configuration example of an information processing system in which this technology is applied to cloud computing.
  • the information processing system of FIG. 15 is composed of a terminal device 400 and a cloud server 500.
  • the terminal device 400 and the cloud server 500 are connected by a network NW such as the Internet, and can communicate with each other.
  • NW such as the Internet
  • the terminal device 400 is composed of a computer terminal such as a PC, a tablet terminal, or a smartphone operated by a user, and includes at least an input unit 410 and a presentation unit 420.
  • the input unit 410 and the presentation unit 420 correspond to the input unit 110 and the presentation unit 120 of FIG. 3, respectively.
  • the cloud server 500 is composed of a large computer device, and includes at least a storage unit 510 and a control unit 520.
  • the control unit 520 includes a learning unit 521.
  • the storage unit 510, the control unit 520, and the learning unit 521 correspond to the storage unit 130, the control unit 140, and the learning unit 141 of FIG. 3, respectively.
  • the storage unit 510 may be provided in a database server or the like configured separately from the cloud server 500.
  • the terminal device 400 transmits the input data to the cloud server 500.
  • the cloud server 500 transmits candidate data that is a candidate to be given the same label as the input data from the terminal device 400 to the terminal device 400.
  • the terminal device 400 presents candidate data, and one of them is selected by the user as a label addition target.
  • the selected candidate data (selected data) is transmitted to the cloud server 500.
  • the cloud server 500 transmits further candidate data to the terminal device 400 based on the learning model learned using the candidate data (selected data) selected in the terminal device 400.
  • FIG. 16 is a block diagram showing a configuration example of computer hardware that executes the above-mentioned series of processes programmatically.
  • the information processing device 100 and the cloud server 500 described above are realized by the computer 1000 having the configuration shown in FIG.
  • the CPU 1001, ROM 1002, and RAM 1003 are connected to each other by the bus 1004.
  • An input / output interface 1005 is further connected to the bus 1004.
  • An input unit 1006 including a keyboard and a mouse, and an output unit 1007 including a display and a speaker are connected to the input / output interface 1005.
  • the input / output interface 1005 is connected to a storage unit 1008 including a hard disk and a non-volatile memory, a communication unit 1009 including a network interface, and a drive 1010 for driving the removable media 1011.
  • the CPU 1001 loads the program stored in the storage unit 1008 into the RAM 1003 via the input / output interface 1005 and the bus 1004 and executes the program, thereby executing the series described above. Processing is done.
  • the program executed by the CPU 1001 is recorded on the removable media 1011 or provided via a wired or wireless transmission medium such as a local area network, the Internet, or digital broadcasting, and is installed in the storage unit 1008.
  • a wired or wireless transmission medium such as a local area network, the Internet, or digital broadcasting
  • the program executed by the computer 1000 may be a program in which processing is performed in chronological order in the order described in this specification, or at a required timing such as in parallel or when a call is made. It may be a program that is processed by.
  • the present technology can have the following configuration.
  • a control unit that outputs candidate data that is a candidate to be given the same label as the input data input by the user.
  • the control unit is an information processing device that outputs further candidate data based on the learning model learned by using the selected data selected by the user as the target for assigning the label from the candidate data.
  • the control unit trains the learning model using the selected data as a positive example and non-selected data not selected by the user as a target for assigning the label from the candidate data as a negative example (1).
  • the information processing device described.
  • (3) The information processing device according to (1) or (2), wherein the control unit outputs the candidate data for each of the learning models based on the plurality of learning models.
  • the control unit The same number of candidate data is output for each learning model, and the same number of candidate data is output.
  • the information processing apparatus according to (6) or (7), further comprising a translation processing unit that translates the input data and the selected data to which the label is attached into a predetermined language.
  • the information processing device wherein the input data is image data.
  • the information processing device according to (1), wherein the input data is sensor data.
  • the information processing device is acceleration data.
  • the sensor data is angular acceleration data.
  • a presentation unit that presents the output candidate data to the user, The information processing apparatus according to any one of (1) to (12), further comprising an input unit that accepts selection of the selected data from the presented candidate data.
  • Information processing device Based on the learning model, output candidate data that can be given the same label as the input data input by the user, and output the candidate data.
  • 100 information processing device 110 input unit, 120 presentation unit, 130 storage unit, 140 control unit, 141 learning unit, 311 translation processing unit, 500 cloud server

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

La présente invention concerne un dispositif de traitement d'informations et un procédé de traitement d'informations qui permettent une collecte efficace de données souhaitées par un utilisateur. Selon la présente invention, une unité de commande qui délivre, sur la base d'un modèle d'apprentissage, des données candidates servant de candidat pour l'attribution de la même étiquette en tant qu'entrée de données d'entrée par un utilisateur, et délivre des données candidates supplémentaires sur la base d'un modèle d'apprentissage qui a appris à l'aide de données sélectionnées sélectionnées en tant que cible pour l'attribution d'une étiquette par l'utilisateur, parmi les données candidates. La présente technologie peut être appliquée à un terminal informatique utilisé dans la collecte de données de mots, par exemple.
PCT/JP2021/000017 2020-01-17 2021-01-04 Dispositif de traitement d'informations et procédé de traitement d'informations WO2021145228A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2020-005937 2020-01-17
JP2020005937A JP2021114082A (ja) 2020-01-17 2020-01-17 情報処理装置および情報処理方法

Publications (1)

Publication Number Publication Date
WO2021145228A1 true WO2021145228A1 (fr) 2021-07-22

Family

ID=76863997

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2021/000017 WO2021145228A1 (fr) 2020-01-17 2021-01-04 Dispositif de traitement d'informations et procédé de traitement d'informations

Country Status (2)

Country Link
JP (1) JP2021114082A (fr)
WO (1) WO2021145228A1 (fr)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2017194782A (ja) * 2016-04-19 2017-10-26 ソニー株式会社 情報処理装置及び情報処理方法
JP2019117437A (ja) * 2017-12-26 2019-07-18 大日本印刷株式会社 物品特定装置、物品特定方法、及びプログラム
WO2019235335A1 (fr) * 2018-06-05 2019-12-12 住友化学株式会社 Système, procédé et programme d'aide au diagnostic

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2017194782A (ja) * 2016-04-19 2017-10-26 ソニー株式会社 情報処理装置及び情報処理方法
JP2019117437A (ja) * 2017-12-26 2019-07-18 大日本印刷株式会社 物品特定装置、物品特定方法、及びプログラム
WO2019235335A1 (fr) * 2018-06-05 2019-12-12 住友化学株式会社 Système, procédé et programme d'aide au diagnostic

Also Published As

Publication number Publication date
JP2021114082A (ja) 2021-08-05

Similar Documents

Publication Publication Date Title
US11037048B2 (en) Virtual conversation method or system
AU2018260853B2 (en) Customized virtual reality learning environment
EP1986175A2 (fr) Procédé, interface et système pour obtenir une entrée d'utilisateur
CA3109113A1 (fr) Systemes et procedes d'entrainement de classificateur a intelligence artificielle
US20200285353A1 (en) Apparatus for vision and language-assisted smartphone task automation and method thereof
US10885444B2 (en) Application tool recommendations
JP6986978B2 (ja) 情報処理装置、情報処理方法、及び情報処理プログラム
WO2015187584A1 (fr) Équipes d'assistants virtuels
CN109940627A (zh) 一种面向绘本阅读机器人的人机交互方法和系统
US20100075289A1 (en) Method and system for automated content customization and delivery
KR102035088B1 (ko) 스토리텔링 기반 멀티미디어 무인 원격 1:1 맞춤형 교육시스템
CN114339285B (zh) 知识点的处理方法、视频处理方法、装置及电子设备
JP2021073590A (ja) 検索用資料情報記憶装置
US11037460B2 (en) Cognitive robotic assistance for education
CN204425398U (zh) 智能早教系统及装置
CN116932733A (zh) 基于生成式大语言模型的信息推荐方法及相关装置
WO2021145228A1 (fr) Dispositif de traitement d'informations et procédé de traitement d'informations
CN110032561A (zh) 基于语义的表单构建方法和系统
KR102272567B1 (ko) 음성 인식 수정 시스템
WO2021132088A1 (fr) Système de traitement d'informations, dispositif de traitement d'informations et procédé de traitement d'informations
JP2022529543A (ja) 脳科学基盤の学習記憶方法、システム、記録媒体
JP2020140674A (ja) 回答選択装置及びプログラム
JP7359196B2 (ja) 情報処理装置、情報処理方法、及び、情報処理プログラム
JPWO2018147435A1 (ja) 学習支援システム及び方法、並びにコンピュータプログラム
WO2023079712A1 (fr) Dispositif de traitement, procédé de traitement et programme de traitement

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: 21741852

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: 21741852

Country of ref document: EP

Kind code of ref document: A1