Opinion prediction device, opinion prediction method, and model generation device

Abstract

An opinion prediction device 30 inputs, to an important comment extraction model 100, data pertaining to a designated topic, data pertaining to comments made in the past on the designated topic by other users except for a designated user, and data pertaining to the designated user, thereby outputting data pertaining to a comment that is important for the designated user among the comments, and inputs, to an opinion prediction model 200, the data pertaining to the designated topic, the data pertaining to the designated user, and the data pertaining to the comment that is important for the designated user on the designated topic, thereby predicting and outputting the tendency of a comment to be made in the future on the designated topic by the designated user.

Description

Opinion prediction device, opinion prediction method, and model generation device

[0001] The present invention relates to an opinion prediction device, an opinion prediction method, and a model generation device.

[0002] INCORPORATION BY REFERENCE This application claims priority to Japanese Patent Application No. 2024-092389, filed June 6, 2024, the contents of which are incorporated herein by reference. In recent years, methods using internet-based services such as SNS (Social Networking Service) have become increasingly common for corporate information (product and service information, corporate strategy, internal policies, and marketing activities) to be disseminated, allowing information to reach a wide range of users. As a result, each user is increasingly susceptible to being influenced by the opinions and positions of other users. Therefore, it is becoming increasingly important for companies to analyze and predict what opinions each user has regarding disseminated information, taking into account such influence between users.

[0003] As a technique for predicting user opinions regarding such information dissemination, Non-Patent Document 1 discloses a technique for predicting a user's future opinions on a certain topic by combining the user's own past opinions with those of other users and utilizing a temporal point process model that takes into account the decay of influence over time.

[0004] Furthermore, as a technology that takes into consideration that each user has their own position, Non-Patent Document 2 discloses an RNN model that predicts a user's current position for a certain topic by combining the user's own past comments within a certain time period, other users' past comments within the same time period, and TOPIC information.

[0005] Furthermore, Patent Document 1 discloses a technology that combines a user's own past opinions and the past opinions of other users on a certain topic, proposes a neural network that controls psychological mechanisms, and predicts one's future opinions.

[0006] International Publication No. 2023 / 084640

[0007] De et al., "Learning and Forecasting Opinion Dynamics in Social Networks", 2016, https: / / arxiv.org / abs / 1506.05474.Zhu et al., "Neural opinion dynamics model for the prediction of user-level stance dynamics", 2019, https: / / doi.org / 10.1016 / j.ipm.2019.03.010.

[0008] However, Non-Patent Document 1 assumes that the user whose opinion is to be predicted has already expressed an opinion in the past, and therefore cannot predict the opinion if the user has not expressed an opinion in the past.Furthermore, it does not take into account that users have various positions.

[0009] Furthermore, Non-Patent Document 2 infers the current position of a specific user on the assumption that there are users who make comments from various positions, but does not predict opinions.

[0010] Although Patent Document 1 predicts opinions, it cannot make predictions unless the user to be predicted has posted an opinion in the past, and does not take into account the content of the topic that is the source of the opinion in the first place.

[0011] The present invention has been made in consideration of the above circumstances, and its purpose is to provide an opinion prediction device, an opinion prediction method, and a model generation device that are capable of predicting future comments on a topic, even for users who have not commented on that topic in the past, while taking into account the influence of comments made by other users.

[0012] One aspect of the present invention for solving the above problem is an opinion prediction device including: a storage device that stores an important comment extraction model that receives as input data on a topic, data on each comment made on the topic in the past, and user data, and outputs data on comments that are important to the user from among the comments; and an opinion prediction model that receives as input data on a topic, data on the user, and data on comments on the topic that are important to the user from among the comments, and outputs data on comments that the user will make on the topic in the future; and a computing device that executes: an important comment extraction process that accepts designation of a user and a topic, and inputs data on the designated topic, data on each comment made on the topic in the past by users other than the designated user, and data on the designated user into the important comment extraction model, thereby outputting data on comments that are important to the designated user from among the comments; and an opinion prediction process that inputs data on the designated topic, data on the designated user, and data on comments that are important to the designated user into the opinion prediction model, thereby outputting a tendency of comments that the designated user will make on the designated topic in the future.

[0013] Another aspect of the present invention for solving the above-mentioned problems is a model generation device including a computing device that generates an important comment extraction model that receives topic data, data on each comment made in the past on the topic, and user data and outputs data on comments that are important to the user from among the comments, by learning the relationship between topic data, data on each comment made in the past on the topic, and user data, and outputs data on comments that are important to the user from among the comments; and that receives topic data, user data, and data on comments that are important to the user from among the comments, by learning the relationship between topic data, user data, data on comments that are important to the user from among the comments, and a tendency of comments that the user will make on the topic in the future, and outputs an opinion prediction model that receives topic data, user data, and data on comments that are important to the user from among the topic, by learning the relationship between topic data, user data, data on comments that are important to the user from among the comments, and a tendency of comments that the user will make on the topic in the future.

[0014] According to the present invention, even for a user who has not commented on a topic in the past, it is possible to predict future comments that will be made on that topic while taking into account the influence of comments made by other users.

[0015] Configurations and effects other than those described above will become apparent from the following description of the embodiments.

[0016] 1 is a diagram illustrating an example of a configuration of an opinion prediction system according to an embodiment of the present invention. FIG. 1 is a diagram illustrating an example of functions provided in an opinion prediction device. FIG. 2 is a diagram illustrating an example of learning data for an important comment extraction model and an opinion prediction model. FIG. 3 is a diagram illustrating details of a data preprocessing unit. FIG. 4 is a diagram illustrating details of an attention construction unit. FIG. 5 is a diagram illustrating details of a model learning unit. FIG. 6 is a diagram illustrating details of an important comment extraction unit. FIG. 7 is a diagram illustrating details of an opinion prediction unit. FIG. 8 is a diagram illustrating an example of a hardware configuration provided in each information processing device in the opinion prediction system. FIG. 9 is a flow diagram illustrating an overview of opinion prediction processing. FIG. 10 is a diagram illustrating an example of a data input screen. FIG. 11 is a flow diagram illustrating details of topic data preprocessing. FIG. 12 is a flow diagram illustrating details of comment data preprocessing. FIG. 13 is a flow diagram illustrating details of user data preprocessing. FIG. 14 is a flow diagram illustrating an example of attention construction processing. FIG. 15 is a diagram illustrating a specific example of output processing of vector data of important comments using an important comment extraction model. FIG. 16 is a diagram illustrating an example of data generated by output processing of vector data of important comments using an important comment extraction model. FIG. 17 is a flow diagram illustrating details of model learning processing. FIG. 18 is a diagram illustrating an example of data generated by model learning processing. FIG. 19 is a flow diagram illustrating details of important comment extraction processing. FIG. 19 is a diagram illustrating an example of data generated by output processing of vector data of important comments using an important comment extraction model. FIG. 10 is a diagram illustrating an example of a prediction result display screen.

[0017] An embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing an example of the configuration of an opinion prediction system 1 according to this embodiment. The opinion prediction system 1 includes an information distribution system 10 that accepts transmission of certain information (hereinafter referred to as a topic) from businesses such as companies and also accepts transmission of responses to the transmitted information (hereinafter referred to as a comment), a user terminal 20 used by each user (e.g., a customer of the company) who transmits comments to the information distribution system 10, an opinion prediction device 30, and a business device 40 used by a business (e.g., a company) that transmits topics to the information distribution system 10 and predicts trends in comments on the transmitted topics using the opinion prediction device 30. The topic is, for example, information related to services, products, internal policies, and marketing activities provided by the business.

[0018] The information distribution system 10, the user terminal 20, the opinion prediction device 30, and the business device 40 are communicatively connected to each other via a wired or wireless communication network 5 such as the Internet, a LAN (Local Area Network), a WAN (Wide Area Network), or a dedicated line.

[0019] The information distribution system 10 is, for example, a server of a news site that transmits news, or a server that realizes a social networking service (SNS). The information distribution system 10 receives text data of topics (which may include images, videos, etc.) from the business operator device 40 and transmits the received topics to each user terminal 20. The information distribution system 10 also receives text data of comments on the topics (impressions, criticisms, etc.) from each user terminal 20 and transmits the received topics to each user terminal 20. Note that the user terminal 20 may transmit topics to the information distribution system 10, and the business operator device 40 may transmit comments to the information distribution system 10.

[0020] The business operator device 40 distributes topics that the business operator wants to post to each user terminal 20 via the information distribution system 10. Here, the opinion prediction device 30 predicts what kind of comments users will post in the future on the topics distributed by the business operator. Specifically, the business operator device 40 predicts what kind of opinions users who have not posted any comments on the topic to the information distribution system 10 so far will post in the future, taking into account the influence of comments posted in the past by users other than the user.

[0021] 2 is a diagram showing an example of functions provided in the opinion prediction device 30. The opinion prediction device 30 stores topic data 310, user data 320, and comment data 330.

[0022] The topic data 310 stores data (text data) on topics previously transmitted to the information distribution system 10 .

[0023] The user data 320 stores attribute information of users who have registered accounts in the information distribution system 10. The attribute information includes, for example, the user's age, hobbies, and values.

[0024] The comment data 330 stores data (text data) of comments sent to the information distribution system 10 .

[0025] The opinion prediction device 30 has the following functional units: a data preprocessing unit 340 , a learning unit 350 , and a prediction unit 360 .

[0026] The data preprocessing unit 340 includes a topic data preprocessing unit 341 , a user data preprocessing unit 342 , and a comment data preprocessing unit 343 .

[0027] The topic data preprocessing unit 341 converts each piece of text data in the topic data 310 into data in a predetermined format (vector data in this embodiment) that represents the characteristics of the topic. The user data preprocessing unit 342 converts each piece of text data in the user data 320 into data in a predetermined format (vector data in this embodiment) that represents the characteristics of the user. The comment data preprocessing unit 343 converts each piece of text data in the comment data 330 into data in a predetermined format (vector data in this embodiment) that represents the characteristics of the comment.

[0028] The learning unit 350 generates an important comment extraction model and an opinion prediction model through machine learning.

[0029] The important comment extraction model inputs data on a topic, data on each comment made on that topic in the past, and user data, and outputs data on comments (hereinafter referred to as important comments) that are important to the user (have influence on the user's future comment postings).

[0030] The opinion prediction model inputs topic data, user data, data on comments on the topic that are important to the user, the time the important comment data was posted, and data on the period since the topic was posted, and predicts and outputs the tendency of comments that the user will make on the topic in the future (after the period has passed) and the time when those comments will be posted.

[0031] Specifically, the learning unit 350 includes an attention construction unit 351 and a model learning unit 352 .

[0032] The attention construction unit 351 inputs data on the topic, data on each comment made in the past on that topic, and user attribute data into an (existing) important comment extraction model, and outputs data on comments that are important to the user among the individual comments.

[0033] The model learning unit 352 performs machine learning to update the parameters of the important comment extraction model and the opinion prediction model by comparing the trends in comments made by the user on the topic over the elapse of the period and the times at which the comments were posted, which are obtained by inputting the data on the topic, the data on the user, the data on the important comments output by the attention building unit 351, the time at which the important comment data was posted, and data on the period since the topic was posted, into the opinion prediction model, with the trends in comments actually made by the user and the times at which the comments were posted.

[0034] In this embodiment, the tendency of a comment is expressed as a numerical value that indicates the positivity of the comment content (the higher the numerical value, the more positive the comment content), but this is not intended to be limiting. Note that in this specification, this tendency of a comment is also referred to as an "opinion."

[0035] The prediction unit 360 includes an important comment extraction unit 361 and an opinion prediction unit 362 .

[0036] The important comment extraction unit 361 accepts designation of a user whose opinion is to be predicted (a user to be predicted) and a topic whose opinion is to be predicted (a topic to be predicted). The important comment extraction unit 361 then inputs data on the topic to be predicted, data on comments made in the past by users other than the user to be predicted on the topic to be predicted, and data on the user to be predicted into an important comment extraction model, and outputs data on comments (important comments) that are important to the user to be predicted from among the above comments.

[0037] The opinion prediction unit 362 inputs data on the topic to be predicted, data on the user to be predicted, and data on comments (important comments) on the topic to be predicted that are important to the user to be predicted into the opinion prediction model, and thereby predicts and outputs the tendency of comments that the user to be predicted will make on the topic to be predicted.

[0038] Of the above functional units, a device having only the functions of the data preprocessing unit 340 and the learning unit 350 may be made independent as a model generation device.

[0039] (Learning Data) FIG. 3 is a diagram showing an example of learning data (topic data 310, user data 320, and comment data 330) for the important comment extraction model and the opinion prediction model.

[0040] Topic data 310 includes the text of each topic.

[0041] The user data 320 is data associated with each topic, and includes an ID 321 of each user who posted a comment on the topic, and attributes 322 of each user.

[0042] Comment data 330 is data linked to each topic and user, and includes the text of each comment 331 of the user on the topic, the tendency of the comment 332, the time when the comment was posted (comment time 333), and the period when the comment was posted (comment period 334).

[0043] Comment time 334 is data that indicates the length of the period from when the corresponding topic is posted to when the corresponding comment is posted. In this embodiment, a numerical value corresponding to the length of the period (the larger the value, the longer the period) is set, but this is not intended to be limiting.

[0044] 4 is a diagram illustrating the details of the data preprocessing unit 340. The topic data preprocessing unit 341 converts the topic data 310 into vector data 311 by inputting the topic data 310 into a predetermined model (for example, a neural network such as Word2Vec (Word to Vector)). The user data preprocessing unit 342 converts the user data 320 into vector data 321 by inputting the user data 320 into a predetermined model (for example, a neural network such as Word2Vec (Word to Vector)). The comment data preprocessing unit 343 converts the comment data 330 into vector data 335 by inputting the comment data 330 into a predetermined model (for example, a neural network such as Word2Vec (Word to Vector)).

[0045] 5 is a diagram illustrating details of the attention construction unit 351. The attention construction unit 351 inputs the vector data 321 of user data, the vector data 311 of topic data, and the vector data 335 of comment data into the important comment extraction model 100, and outputs, for each topic and each user, vector data 101 related to important comments among the comments posted in the past by the user. This vector data 101 is input to the model learning unit 352, which will be described next.

[0046] The important comment extraction model 100 is a trained model based on an attention mechanism (attention block) described later, and is, for example, a mathematical model using Transformer.

[0047] 6 is a diagram illustrating details of the model learning unit 352. The model learning unit 352 predicts the tendency of comments made by the user on the topic by inputting the vector data 321 of the user data, the vector data 311 of the topic data, the vector data 101 related to important comments output by the attention construction unit 351, and the comment time data 102 into the opinion prediction model 200. The comment time data 102 is data including the comment time 333 and the comment time 334 of important comments.

[0048] Then, the model learning unit 352 calculates the value of a predetermined function (loss function) that indicates the deviation between the opinion prediction result by the opinion prediction model 200 and the correct opinion, based on the predicted trend (opinion) and trend 332 related to comments that have actually been posted and registered in the comment data 330. The model learning unit 352 updates each parameter in the opinion prediction model 200 and the important comment extraction model 100 based on the calculated value of the loss function, thereby learning the opinion prediction model 200 and the important comment extraction model 100.

[0049] 7 is a diagram illustrating details of the important comment extraction unit 361. The important comment extraction unit 361 inputs vector data 323 of user data 320 of the user to be predicted, vector data 312 of topic data 310 of the topic to be predicted, and vector data 336 of comment data 330 related to comments previously made on the topic to be predicted into the important comment extraction model 100, and outputs vector data 103 of important comments related to the user to be predicted from among comments previously posted on the target topic. This vector data 103 is input to the opinion prediction unit 362, which will be described next.

[0050] 8 is a diagram illustrating the details of the opinion prediction unit 362. The opinion prediction unit 362 inputs vector data 323 of the user data 320 of the user to be predicted, vector data 312 of the topic data 310 of the topic to be predicted, vector data 103 of important comments (and the comment times) related to the user to be predicted output by the important comment extraction unit 361, and data 104 of the prediction target period into an opinion prediction model, thereby predicting and acquiring the tendency 202 (opinions) of comments made by the user to be predicted on the topic to be predicted during the prediction target period and the comment times. Note that the prediction target period refers to the period from the publication of the topic to be predicted to the comment being made (corresponding to comment period 334).

[0051] 9 is a diagram showing an example of a hardware configuration of each information processing device in the opinion prediction system 1. Each information processing device includes a calculation device 91 such as a CPU (Central Processing Unit), a main storage device 92 such as a RAM (Random Access Memory) or a ROM (Read Only Memory), an external storage device 93 such as a HDD (Hard Disk Drive) or an SSD (Solid State Drive), an input device 94 such as a keyboard, a mouse, or a touch panel, an output device 95 such as a display or a touch panel, and a communication device 96 configured from a NIC (Network Interface Card), a wireless communication module, a USB (Universal Serial Interface) module, a serial communication module, or the like.

[0052] The functions of the functional units of each information processing device in the opinion prediction system 1 described above are realized by the arithmetic unit 91 of each information processing device reading a program from the main storage device 92 or the external storage device 93. Each program can be recorded on, for example, a portable or fixed recording medium and distributed. All or part of each program in each information processing device may be realized using virtual information processing resources provided using virtualization technology, process space separation technology, or the like, such as a virtual server provided by a cloud system. All or part of these programs may be realized by a service provided by the cloud system via an API (Application Programming Interface), for example. Next, the processing performed in the opinion prediction system 1 will be described.

[0053] 10 is a flow diagram illustrating an outline of the opinion prediction process performed in the opinion prediction system 1. The opinion prediction process is started, for example, when the business operator device 40 transmits a predetermined start instruction to the opinion prediction device 30.

[0054] The opinion prediction device 30 executes topic preprocessing s10, comment data preprocessing s20, and user data preprocessing s30 to generate vector data 311, 321, and 335 of each topic posted to the information distribution system 10, each comment on each topic, and the attributes of the users who posted each comment.

[0055] The opinion prediction device 30 inputs the vector data 311, 321, 335 generated in the topic preprocessing s10, the comment data preprocessing s20, and the user data preprocessing s30 into the important comment extraction model 100, and executes the attention construction process s40 that outputs the vector data 101 of important comments.

[0056] Then, the opinion prediction device 30 executes a model learning process s50 to learn the opinion prediction model 200 and the important comment extraction model 100 based on the vector data 311, 321 generated in the topic pre-processing s10 and the user data pre-processing s30, the vector data 101 of the important comments output in the attention construction process s40, and the correct answer data of the opinions, etc.

[0057] The attention construction process s40 and the model learning process s50 are repeated multiple times.

[0058] Thereafter, the opinion prediction device 30 receives input of data on the user to be predicted from the business operator device 40 in order to predict opinions (s60).

[0059] 11 is a diagram showing an example of a data input screen 1100 displayed on the opinion prediction device 30. The data input screen 1100 has an age input field 1101 that receives input of age, which is an attribute of the user to be predicted, from the business operator device 40, a hobby input field 1102 that receives input of hobby, which is an attribute of the user to be predicted, from the business operator device 40, a value input field 1103 that receives input of a value type, which is an attribute of the user to be predicted, from the business operator device 40, a topic selection field 1104 that receives selection of a topic to be predicted from the business operator device 40, and a prediction target period selection field 1105 that receives selection of a prediction target period from the business operator device 40.

[0060] The prediction target user is a user who has not previously commented on the prediction target topic.

[0061] Next, as shown in FIG. 10 , the opinion prediction device 30 inputs the user to be predicted and the topic to be predicted, etc., input in s60 into the important comment extraction model 100, and executes an important comment extraction process s70 to extract important comments related to the user to be predicted from past comments on the topic to be predicted.

[0062] Then, the opinion prediction device 30 executes an opinion prediction process s80 that predicts the tendency (opinions) of future comments made by the user to be predicted on the topic to be predicted by inputting the important comments, etc. extracted in the important comment extraction process s70 into the opinion prediction model 200.

[0063] The opinion prediction device 30 displays a screen showing the execution result of the opinion prediction process s80 on the screen of the business entity device 40 (screen display process s90). This completes the opinion prediction process.

[0064] Each process will be described in detail below.

[0065] 12 is a flow diagram illustrating the details of the topic data preprocessing step s10. The topic data preprocessing unit 341 first reads a neural network (such as word2vec) (s11).

[0066] The topic data preprocessing unit 341 also acquires text data of all topics in the topic data 310 (s12).

[0067] The topic data preprocessing unit 341 inputs each piece of text data acquired in s11 into a neural network, and outputs vector data 311 corresponding to the text data (s13).

[0068] The topic data preprocessing unit 341 stores each of the output vector data 311 (s14).

[0069] <Comment Data Preprocessing> Fig. 13 is a flow diagram illustrating the comment data preprocessing step s20 in detail. The comment data preprocessing unit 343 first reads a neural network (such as word2vec) (s21).

[0070] Furthermore, the comment data preprocessing unit 343 acquires the text data of all comments in the comment data 330 (s22).

[0071] The comment data preprocessing unit 343 inputs each piece of text data acquired in s22 into a neural network, and outputs vector data 335 corresponding to the comment data (s23).

[0072] The comment data preprocessing unit 343 stores each of the output vector data 335 (s24).

[0073] <User Data Pre-Processing> FIG. 14 is a flow diagram illustrating the details of the user data pre-processing s30.

[0074] The user data preprocessing unit 342 acquires data of all user attributes in the user data 320 (s31).

[0075] The user data preprocessing unit 342 determines whether the data for each user attribute acquired in s31 is text data (s32). If the data for the user attribute acquired in s31 is text data (s32: YES), the user data preprocessing unit 342 executes the process of s33 for the data for that user attribute. If the data for the user attribute acquired in s31 is not text data (s32: NO), the user data preprocessing unit 342 executes the process of s36 for the data for that user attribute.

[0076] In s33, the user data preprocessing unit 342 reads a neural network (such as word2vec).

[0077] Then, the user data preprocessing unit 342 inputs the data of the user attributes into the neural network, thereby outputting vector data 321 corresponding to the user data (s34).

[0078] The user data pre-processing unit 342 stores the output vector data 321 (s35). This completes the user data pre-processing s30.

[0079] Meanwhile, in s36, the user data preprocessing unit 342 stores the user attribute data as data corresponding to the vector data 321. This completes the user data preprocessing s30.

[0080] 15 is a flow diagram illustrating an example of the attention building process s40. The attention building unit 351 reads the vector data 311 of each topic generated in the topic data preprocessing s10 (s41).

[0081] The attention building unit 351 also reads the vector data 335 of each comment posted in the past for each of the topics generated in the comment data preprocessing s20 (s42).The attention building unit 351 also reads the vector data 321 of each user's attributes generated in the user data preprocessing s30 (s43).

[0082] Then, the attention construction unit 351 inputs each of the data read in steps s41-s43 into the important comment extraction model 100 (s44), and outputs vector data 101 of the important comments of each of the users posted on each of the topics (s45).

[0083] 16 is a diagram illustrating a specific example of output processing of vector data 101 of important comments by the important comment extraction model 100. The attention construction unit 351 acquires vector data 1601 of a user's attributes from the vector data of each user's attributes, and also acquires vector data 1602 of a topic's attributes from the vector data of each topic. The attention construction unit 351 combines these vector data and inputs them into a predetermined neural network (fully connected layer) to generate a query vector 1603 related to the user's attributes and the topic (s441). The query vector 1603 is query data for identifying a comment to be acquired from each comment, which will be described later.

[0084] The attention construction unit 351 also inputs vector data 1604 of each past comment on the above topic into a predetermined neural network (fully connected layer) to generate a key vector 1605 (data representing the proximity of the query to the data to be acquired) for each past comment (s442). The attention construction unit 351 also inputs vector data 1604 of each past comment on the above topic into a predetermined neural network (fully connected layer) to generate a value vector 1606 of each comment (vector data of the comment output based on the key vector, as described below) (s443).

[0085] The attention construction unit 351 generates an attention vector 1607 based on the query vector 1603 generated in s441 and the key vector 1605 generated in s442 (s444).

[0086] Specifically, the attention construction unit 351 calculates the dot product of a query vector 1603 related to the user attribute and topic and a key vector 1605 of each past comment, thereby calculating the relevance of each user attribute and topic to each comment.The attention construction unit 351 then inputs each dot product value into a predetermined function (e.g., a softmax function) and normalizes each dot product value to calculate an attention vector 1607 (a vector made up of weight parameters (attention weight) that indicate the degree of attention the query pays to each past comment).

[0087] Furthermore, the attention construction unit 351 calculates the dot product of the attention vector 1607 and the generated value vector 1606 of each comment, thereby extracting a comment vector from the vector of each past comment according to the attention weight (s445), and inputs the extracted vector into a predetermined neural network (fully connected layer) to convert it into vector data 1608 representing important comments on the topic related to the user. The attention construction unit 351 generates and stores data related to the posting time of the important comment and the comment timing in association with the converted vector data 1608 of the important comment.

[0088] 17 is a diagram showing an example of data 1700 generated by the output process of the vector data 101 of important comments using the important comment extraction model 100. This data 1700 is data including, for each topic, vector data 1701 (1608) of important comments on that topic, the time at which the important comment was posted (comment time 1702), and the comment period 1703 related to the important comment.

[0089] <Model Learning Process s50> FIG. 18 is a flowchart illustrating the details of the model learning process s50.

[0090] The model learning unit 352 acquires vector data for a certain topic A from the vector data 311 for each topic (s51). The model learning unit 352 also acquires vector data for important comments on the topic A from the vector data 101 generated in the attention building process s40 (s52). The model learning unit 352 also acquires vector data for the attributes of a certain user U from the vector data 321 for each user's attributes (s53).

[0091] Furthermore, the model learning unit 352 acquires the comment time 1702 and comment period 1703 of the important comment from the data 1700 (s54).

[0092] The model learning unit 352 inputs the data acquired in s51 to s54 into the opinion prediction model 200 (s55), thereby predicting the tendency and comment time of comments made by user U on topic A in the comment period 334 after the posting of topic A (s56). The model learning unit 352 generates and stores data including the prediction result of s56.

[0093] Meanwhile, the model learning unit 352 acquires the tendency (opinion) and comment time 333 of the comment actually posted by the user U on the topic A from the tendency 332 of the comment data 330 (s57).

[0094] The model learning unit 352 calculates the degree of deviation between the comment tendency (opinion) and comment time predicted in s55 and the comment tendency (opinion) and comment time 333 acquired in s56. The degree of deviation is calculated, for example, using a loss function that calculates a mean squared error (MSE).

[0095] The model learning unit 352 updates the hyperparameters of the important comment extraction model 100 and the opinion prediction model 200 in accordance with the calculation result of s58 (s58). This completes the model learning process s50.

[0096] 19 is a diagram showing an example of data generated by the model learning process s50. This data 1900 includes a topic 1901, a user ID 1902, attributes 1903 of the user, vector data 1904 of comments predicted to be posted by the user on the topic, a tendency 1905 of the comments, a predicted posting time 1906 of the comment, and a commenting period 1907 of the comment.

[0097] 20 is a flow diagram illustrating the details of the important comment extraction process s70. The important comment extraction unit 361 reads the vector data of the prediction target topic specified in s60 from the vector data 311 generated in the topic data preprocessing s10 (s71).

[0098] The important comment extraction unit 361 reads the vector data of each comment posted in the past on the topic to be predicted (note that the user to be predicted has never posted any of these comments) from the vector data 335 generated in the comment data pre-processing s30 (s72). Furthermore, the important comment extraction unit 361 reads the vector of the user to be predicted specified in s60 from the vector data 321 generated in the user data pre-processing s20 (s73).

[0099] Then, the important comment extraction unit 361 inputs each of the vector data read in steps s71-s73 into the important comment extraction model 100 (s74), and outputs vector data of important comments related to the user to be predicted from among the comments posted on the topic to be predicted (s75).

[0100] 21 is a diagram showing an example of data 2100 generated by the process of outputting vector data of important comments using the important comment extraction model 100. This data 2100 includes vector data 2101 of a topic to be predicted, vector data 2102 of each important comment corresponding to the vector to be predicted, comment time 2103 of each important comment, and comment period 2104 of each important comment.

[0101] <Opinion Prediction Processing> FIG. 22 is a flowchart illustrating the opinion prediction processing s80 in detail.

[0102] The opinion prediction unit 362 acquires vector data of the topic to be predicted (s81). The opinion prediction unit 362 also acquires vector data of important comments on the topic to be predicted from the data 2100 generated in the important comment extraction process s70 (s82). The opinion prediction unit 362 also acquires vector data of the attributes of the user to be predicted (s83).

[0103] Furthermore, the opinion prediction unit 362 acquires the comment time and comment period of the important comment from the data 2100 (s84).

[0104] The opinion prediction unit 362 inputs the data acquired in s81 to s84 into the opinion prediction model 200 (s85), thereby predicting the type of comment (opinion) and the comment time that the user to be predicted will make on the topic to be predicted during the comment period after the posting of the topic to be predicted (s86). The opinion prediction unit 362 generates and stores data including the prediction result of s86.

[0105] 23 is a diagram showing an example of data generated in the opinion prediction process s80. This data 2300 includes vector data 2301 of the topic to be predicted, a user to be predicted 2302, attributes 2303 of the user to be predicted, a vector 2304 of a comment that is predicted to be posted by the user to be predicted on the topic to be predicted, a tendency 2305 (opinion) of the comment, a comment time 2306 that is predicted to be posted of the comment, and a comment period 2307 of the comment.

[0106] 24 is a diagram showing an example of a prediction result display screen 2400 displayed in the screen display process s90. The prediction result display screen 2400 has a graph display screen 2410 that displays a time series graph 2403 of opinions that shows the relationship between the period (horizontal axis 2401) from the time the topic to be predicted is posted to the time the user to be predicted posts a comment on the topic to be predicted and the opinions (trends) of comments predicted to be made at each time during that period (vertical axis 2402), and an important comment display field 2420 that displays the important comments extracted in the important comment extraction process s70.

[0107] The horizontal axis 2401 may display the comment period as well as the specific predicted time of the comment.

[0108] Although the prediction result display screen 2400 displays a time series graph, the time series of opinions may be displayed in other forms of graphics (such as bar graphs).

[0109] As described above, the opinion prediction device 30 of this embodiment outputs comments that are important to the user to be predicted (important comments) from among the above comments by inputting data on the topic to be predicted, data on each comment made in the past by users other than the user to be predicted, and attribute data of the user to be predicted into the important comment extraction model 100. Then, the opinion prediction device 30 inputs data on the topic to be predicted, attribute data of the user to be predicted, and the above important comments that are important to the user to be predicted into the opinion prediction model 200, thereby predicting and outputting comments that the user to be predicted will make in the future on the topic to be predicted.

[0110] That is, the opinion prediction device 30 of this embodiment identifies comments that are important to the user to be predicted, i.e., that will have an impact on the user to be predicted, based on comments made in the past by users other than the user to be predicted, for the topic to be predicted, and predicts the comments that the user to be predicted will make on the topic to be predicted using those comments as parameters.

[0111] In this way, according to the opinion prediction device 30 of this embodiment, even if a user has not commented on a topic in the past, it is possible to predict future comments that the user will make on that topic, taking into account the influence of comments made by other users.

[0112] Furthermore, the opinion prediction model 200 of this embodiment further receives input of the time when the data of the important comment was posted and data on the period from the posting of the topic (the period to be predicted), and predicts and outputs the tendency of comments that users will make on the topic over the elapse of the period and the posting time of the comment.

[0113] This makes it possible to predict opinions taking into account the timing of posting of important comments, and also to predict comments posted by users to be predicted on topics to be predicted, along with the posting times of those comments.

[0114] Furthermore, the opinion prediction model of this embodiment outputs to a prediction result display screen 2400 the relationship between the period from the publication of the topic to be predicted to the publication of a predicted comment on the topic to be predicted and the tendency (opinion) of the predicted comment.

[0115] This allows us to check the time series changes in the impact that the topic to be predicted has on the comments of the predicted user.

[0116] Moreover, the opinion prediction device 30 (model generation device) of this embodiment learns the relationship between topic data, data on each comment made on the topic in the past, and between user data and data on comments that are important to the user among the above comments, thereby generating the important comment extraction model 100. Furthermore, the opinion prediction device 30 learns the relationship between topic data, user data, data on comments on the topic that are important to the user, and the tendency of the user to make comments on the topic in the future, thereby generating the opinion prediction model 200.

[0117] By using the important comment extraction model 100 and opinion prediction model 200 generated in this way, it is possible to predict future comments on a topic, even for a user who has not commented on that topic in the past, while taking into account the influence of comments made by users other than that user.

[0118] The present invention is not limited to the above-described embodiments, and can be implemented using any components within the scope of the present invention. The above-described embodiments and modifications are merely examples, and the present invention is not limited to these contents as long as the features of the invention are not impaired. Furthermore, although various embodiments and modifications have been described above, the present invention is not limited to these contents. Other aspects conceivable within the scope of the technical idea of ​​the present invention are also included within the scope of the present invention.

[0119] For example, part of the hardware provided in each device of this embodiment may be provided in another device.

[0120] Furthermore, each program of each device may be provided in another device, a program may consist of multiple programs, or multiple programs may be integrated into one program.

[0121] Furthermore, in this embodiment, the important comment extraction model is an extraction model for an important comment from among a plurality of comments, but it may also be a model for extracting an important part of a certain comment.

[0122] 30 Opinion prediction device, 100 Important comment extraction model, 200 Opinion prediction model

Claims

1. An opinion prediction device comprising: a storage device that stores an important comment extraction model that receives input of topic data, data on each comment made on the topic in the past, and user data, and outputs data on comments that are important to the user from among the comments; an opinion prediction model that receives input of topic data, user data, and data on comments on the topic that are important to the user from among the comments, and outputs data on comments that the user will make in the future from the topic; and a computing device that executes an important comment extraction process that accepts designation of a user and a topic, and inputs data on the designated topic, data on each comment made in the past on the topic by users other than the designated user, and data on the designated user into the important comment extraction model, thereby outputting data on comments that are important to the designated user from among the comments; and an opinion prediction device that inputs data on the designated topic, data on the designated user, and data on comments that are important to the designated user into the opinion prediction model, thereby outputting data on comments that the designated user will make in the future from the designated topic.

2. The opinion prediction device according to claim 1, wherein the storage device further inputs the time when the important comment data was posted and data on the period since the topic was posted, and stores an opinion prediction model that outputs the tendency of comments made by the user on the topic over the period and the time when the comment was posted; and the calculation device, in the opinion prediction process, further inputs the time when the important comment data was posted and data on the period since the specified topic was posted into the opinion prediction model, thereby outputting the tendency of comments made by the specified user on the specified topic over the period and the time when the comment was posted.

3. The opinion prediction device according to claim 1, wherein the calculation device outputs to an output device a graphic representing the relationship between the period from the posting of the specified topic to the posting of the future comment and the tendency of the future comment.

4. An opinion prediction method using an information processing device comprising: a storage device that stores an important comment extraction model that receives input of topic data, data on each comment made on the topic in the past, and user data, and outputs data on comments that are important to the user from among the comments; and an opinion prediction model that receives input of topic data, user data, and data on comments on the topic that are important to the user from among the comments, and outputs tendencies of comments that the user will make on the topic in the future, and a computing device, wherein the computing device executes: an important comment extraction process that accepts designation of a user and a topic, and inputs data on the designated topic, data on each comment made in the past on the topic by users other than the designated user, and data on the designated user into the important comment extraction model, thereby outputting data on comments that are important to the designated user from among the comments; and an opinion prediction process that inputs data on the designated topic, data on the designated user, and data on comments that are important to the designated user into the opinion prediction model, thereby outputting tendencies of comments that the designated user will make on the designated topic in the future.

5. The opinion prediction method of claim 4, wherein the storage device further inputs the time when the important comment data was posted and data on the period since the topic was posted, and stores an opinion prediction model that outputs the tendency of comments made by the user on the topic over the period and the time when the comment was posted; and the calculation device, in the opinion prediction process, further inputs the time when the important comment data was posted and data on the period since the specified topic was posted into the opinion prediction model, thereby outputting the tendency of comments made by the specified user on the specified topic over the period and the time when the comment was posted.

6. The opinion prediction method according to claim 4, wherein the calculation device outputs to an output device a graphic representing the relationship between the period from the posting of the specified topic to the posting of the future comment and the tendency of the future comment.

7. A model generation device comprising: a computing device that generates an important comment extraction model that receives topic data, data on each comment made in the past on the topic, and user data and outputs data on comments that are important to the user from among the comments, by learning the relationship between topic data, data on each comment made in the past on the topic, and user data, and outputs data on comments that are important to the user from among the comments; and a computing device that receives topic data, user data, and data on comments that are important to the user from among the topic, and generates an opinion prediction model that receives topic data, user data, and data on comments that are important to the user from among the topic, by learning the relationship between topic data, user data, data on comments that are important to the user from among the topic, and predicts and outputs trends in comments that the user will make on the topic in the future.

Images