JP7808066B2 - Apparatus, method and program for providing personalized updates on AI models - Google Patents

Description

The present invention relates to an apparatus, method, and program that analyzes information (information processing, content, etc.) related to AI models and individually provides personalized update information to users who have an interest in the AI models.

The system and method described in Patent Document 1 detect changes (updates) in online content and determine whether the updates are meaningful. Whether a content update is meaningful is determined based on the nature of the content itself (for example, whether it is a meaningful part, like a news headline, or a meaningless part, like an advertisement) and how users have interacted with the content in the past, such as by clicking or zooming on the content. Users can also specify parts of the content to give them meaning.

The system and method described in Patent Document 1 are expected to use technology that automatically analyzes content. If a change in content is determined to be significant, the system notifies the user of what has changed and to what extent the content has changed. The user receives a notification on their client device. The notification may include a portion of the content or a visual design to attract the user's attention.

Non-Patent Document 1 proposes a semi-automated approach to help users understand the differences between two versions of software release notes (notes showing changes made when upgrading software). This approach collects relevant information about the release and automatically creates a list of changes. It then generates release notes targeted at specific readers depending on the information required (e.g., by filtering the information to reduce the amount of information). The data displayed can also be based on the general needs of specific readers, such as focusing on technical information for technical users and non-technical information for non-technical users. For example, it is possible to group user stories by episode, or to hide information about user-generated bug reports.

U.S. Patent No. 9,830,400

Sebastian Klepper, Stephan Krusche & Bernd Bruegge, “Semi-Automatic Generation of Audience-Specific Release Notes”, Proceedings of the International Workshop on Continuous Software Evolution and Delivery (CSED), May 2016, pp. 19-22

The system and method of Patent Document 1 determine whether meaningful changes have occurred in the content based on the user's interactions with the content (clicking, zooming, etc.). However, this system and method does not analyze why the user performed that interaction. Furthermore, while the system and method may allow the user to specify something meaningful to them, the user cannot specify information that does not yet exist, and therefore does not handle completely new information within the content. Furthermore, while the user is notified of changes to the content, the user's interaction that led to the notification is unknown.

The approach in Non-Patent Document 1 provides information at the group level, rather than at the individual user level. As a result, many users in the group receive information, but that information may not necessarily be relevant to each user. Also, as in Patent Document 1, the data displayed to the user only includes what has changed, not the reason why the user is being notified.

The present invention aims to provide an apparatus, method, and program that can analyze user interests and AI model information related to AI models, and, when an AI model is updated, can individually notify users who are interested in the AI model of personalized reports.

The information provision device of the present invention analyzes user interests in an AI model and information related to the AI model, and when the AI model information is updated, creates and provides personalized reports individually to users who are interested in the AI model based on the user interests related to the updated AI model information.

The information providing device may include a first module that generates user interest categories based on the user's interests in an AI model and assigns the generated interest categories to the user's interests; a second module that generates AI model information topics based on information about the AI model and the user's interest categories, assigns the interest categories corresponding to the generated AI model information topics, and triggers a report to the user in response to an update of the AI model information; and a third module that, when the report to the user is triggered, identifies interested users based on the interest categories corresponding to the updated AI model information topics and the user's interests corresponding to the interest categories, and creates a personalized report for the identified users.

In addition, the information provision method of the present invention analyzes user interests in an AI model and information related to the AI model, and when the AI model information is updated, creates and provides personalized reports individually to users who are interested in the AI model based on the user interests related to the updated AI model information.

The information provision method may include the steps of generating user interest categories based on the user's interests in an AI model and assigning the generated interest categories to the user's interests; generating AI model information topics based on information about the AI model and the user's interest categories, assigning the interest categories corresponding to the generated AI model information topics, and triggering a report to the user in response to an update of the AI model information; and, when the report to the user is triggered, identifying interested users based on the interest categories corresponding to the updated AI model information topics and the user's interests corresponding to the interest categories, and creating a personalized report for the identified users.

Furthermore, the information provision program according to the present invention may be a program for causing a computer to function as the aforementioned information provision device.

The present invention makes it possible to individually provide personalized reports regarding updates to AI model information to each user who has an interest in an AI model, and also enables users who receive the reports to easily understand the relevance of the reports to their interests (i.e., why the report was sent to them).

1 is a diagram illustrating a configuration of an information providing apparatus according to an embodiment. FIG. 1 is a diagram illustrating a process flow in a first module (module 1) for generating user interest categories according to an embodiment. FIG. 10 is a diagram showing the flow of the initial processing in the second module (module 2) that generates AI model information topics in the embodiment. FIG. 10 illustrates a stage in which the AI model information is updated and module 2 is re-executed in the second module (module 2) that generates AI model information topics in an embodiment. FIG. 10 illustrates a process flow for a third module (Module 3) that provides a report to a user in an embodiment. FIG. 10 is a diagram showing the process flow (continuation of FIG. 5) in a third module (module 3) that provides a report to a user in an embodiment.

Hereinafter, an example of an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a diagram showing the configuration of an information providing apparatus 10 according to an embodiment.
The information providing device 10 receives user interests and AI model information as input data, and outputs (transmits) personalized reports regarding updates to the AI model to individual users who have interests in the AI model. As can be seen in Figure 1, the information providing device 10 is an information processing device (computer) that includes a control unit 20, an input unit 30, an output unit 40, a storage unit 50, and repositories 61 to 64.

The control unit 20 controls the entire information providing device 10, and performs processing in each module described below by appropriately reading and executing various programs stored in the memory unit 50. The control unit 20 is not particularly limited, but may be a CPU.

The storage unit 50 is a storage area for various programs and data that enable the hardware group to function as the information providing device 10, and may be, but is not limited to, ROM, RAM, flash memory, a hard disk drive (HDD), a solid state drive (SSD), etc.

A repository means a storehouse or repository, and since the object of storage in this invention is data, it is a type of storage device. Note that while four repositories 61-64 are shown in Figure 1, this is for the sake of convenience of explanation only, and four physical repositories are not a required configuration; each repository may be virtual (logical) (for example, included in storage unit 50).

From the viewpoint of the data processing flow, the information providing device 10 is composed of at least three software modules (modules 1 to 3 shown in FIGS. 2, 3, and 5). Each module performs processes such as data generation, allocation, and storage, which will be described later. A program describing commands for these processes is stored in the storage unit 50 of the computer, and the CPU of the control unit 20 reads and executes the program.
The processing contents (flow) in the three modules 1 to 3 will be described in detail below.

<Module 1>
2 is a diagram showing the processing flow of a first module (module 1) that generates user interest categories. Module 1 performs the following processes: analyzing user interests to generate (define) interest categories; storing (registering) the generated interest categories in the category repository 61; assigning user interests to the generated interest categories to generate a first table; and storing the generated first table in the first repository 62.

The input to Module 1 is user interests regarding the AI model, such as questions, feedback, comments, etc., from the user related to the AI model.
In the initial stage, Module 1 analyzes the input user interests and generates (defines) initial interest categories. This analysis can be performed manually, automatically, or in a hybrid manner. The automatic analysis can be performed, for example, by applying natural language processing technology, without any particular limitation.

For example, as shown in Figure 2, suppose User A comments to an AI model that generates images, "Interesting app. However, when I select the 'fictional character' style, it sometimes generates a naked human no matter what I do. When I enter 'animal' in the description, it sometimes ends up with three eyes, five legs, or two tails." When this comment is entered into Module 1, it is analyzed as the user's interest, and an interest category called "inappropriate images" is generated (defined) in relation to "naked humans no matter what I do" (User A's Interest 1), and an interest category called "dysmorphic" is generated (defined) in relation to "three eyes, five legs, or two tails" (User A's Interest 2).

Similarly, when User B comments, "Even when I enter the description of a person as 'non-white,' it often ends up being a light-skinned person. Sometimes the person's face appears distorted, and the text is difficult to read," Module 1 generates (defines) "Racial issues" in relation to "Even when I enter the description of a person as 'non-white,' it ends up being a light-skinned person" (User B's Interest 1), "Dysmorphic features" in relation to "The person's face appears distorted" (User B's Interest 2), and "Difficult to read text" in relation to "The text is difficult to read" (User B's Interest 3).

The module 1 stores (registers) the generated (defined) interest categories in the category repository 61 .
Once the interest categories are defined, module 1 assigns the corresponding interest categories to user A's interests 1 and 2 and user B's interests 1 to 3, respectively, to generate a table of user interests x interest categories (first table), and stores the generated first table in the first table repository 62. An example of the first table is shown in the lower right of FIG.

Module 1 can also be configured to re-run as new user interests are accumulated and entered, to generate new interest categories or update existing interest categories. The re-run phase involves the same processing as the initial phase described above.

The interest categories are defined according to the usage of the AI model.
For example, in the case of a health-related AI model that diagnoses diseases, many users may have questions about how the AI model handles elderly patients, resulting in the creation of an interest category called "elderly patients."
Also, for an AI model that generates images, many users may want to know whether it can generate images with readable text, which may result in the creation of an interest category called "text readability."

Furthermore, user interests that may appear similar at first glance but have some differences may generate one or more interest categories. For example, in the case of an AI model that generates images, many users may want to know whether it can generate images with readable text, but some users may refer to English and others to Japanese. As a result, the interest categories "readability of text," "readability of English text," and "readability of Japanese text" may be generated.

<Module 2>
3 is a diagram showing the flow of the initial processing in the second module (module 2) that generates AI model information topics. Module 2 performs the following processes: generating AI model information topics from AI model information, storing (registering) the generated AI model information topics in the topic repository 63, generating a second table by assigning which AI model information topic deals with which interest category (which topic corresponds to which category), and storing the generated second table in the second table repository 64.

The inputs to Module 2 are AI model information and interest categories. Here, AI model information refers to information about the AI model, such as the AI model's performance, training data, allowable uses, limitations, biases, etc. The interest categories are defined in Module 1 and registered in Category Repository 4.

In the initial stage, Module 2 analyzes existing AI model information and generates (defines) initial AI model information topics. Here, the AI model information topics are defined according to the use cases of the AI model and user interest categories.

For example, in the case of a health-related AI model that diagnoses diseases, the AI model information may be categorized into general topics such as "performance" and "application," but in the case of AI model information that specifies "performance" related to the elderly, a more specific AI model information topic such as "performance for the elderly" may be generated (defined) in accordance with the interest category of "elderly."
Also, in the case of an AI model that generates images, the AI model information may be categorized into general topics such as "performance" and "limitations," but if the information includes a reference to generating text, a "text" topic may be created (defined).

In addition, analysis of AI model information, like analysis of user interests, can be performed manually, automatically, or a hybrid of both. Automatic analysis is not particularly limited, but can be performed, for example, by applying natural language processing technology.

For example, as shown in Figure 3, assume that the following information is input regarding restrictions on AI model information: "This model cannot output readable text. This model was trained on a dataset containing adult, violent, and sexual content." When this information is input into Module 2, it is analyzed as restriction information on the AI model, and an AI model information topic called "Text output" is generated (defined) in relation to "cannot output readable text," and an AI model information topic called "Violent and sexual content" is generated (defined) in relation to "trained on a dataset containing adult, violent, and sexual content."

Similarly, with regard to bias in AI model information, when the information "This model may lack consideration for communities and cultures that use minority languages. In addition, Western culture is set as the default, which may affect the output," is input, Module 2 generates (defines) the AI model information topic "Language is limited" in relation to "Lack of consideration for communities and cultures that use minority languages," and the AI model information topic "Western is the default" in relation to "Western culture is set as the default."

Module 2 stores (registers) the generated (defined) AI model information topics in topic repository 63. Here, module 2 stores the registered AI model information topics in topic repository 63, including information on which part of the AI model information the topics correspond to.

When an AI model information topic is registered in the topic repository 63, the determination as to whether the topic repository shown in Figure 3 has been updated is "YES," and the process proceeds to the next step. In the next step, which AI model information topic deals with which interest category (which topic corresponds to which category) is assigned, and a report to the user is triggered (except in the initial stage).

For example, in the case of the aforementioned health-related AI model that diagnoses diseases, the AI model information topic "Elderly Performance" can be assigned to the corresponding interest category "Elderly." Note that an AI model information topic may correspond to zero interest categories, and vice versa. Also, one or more AI model information topics can be generated from the same piece of AI model information.

Once the AI model information topics are defined in this way, module 2 assigns which AI model information topics correspond to which interest categories, generates a table of AI model information topics x interest categories (second table), and stores the generated second table in the second table repository 64. An example of the second table is shown in the lower right corner of Figure 3.

Module 2 can also be configured to be re-executed when a new version of the AI model is released (when the AI model information is updated) to generate new AI model information topics or update existing AI model information topics. The re-execution phase also performs the same processing as the initial phase described above, but the difference is that in the initial phase, even if the determination of whether the topic repository has been updated is "YES," a report to the user is not triggered, whereas in the re-execution phase, a report to the user is triggered.

Figure 4 illustrates the stage when the AI model information has been updated and module 2 has been re-executed. Comparing it with Figure 3, it can be seen that the AI model information, AI model information topics, and second table have each been updated.

For example, we can see that the "restriction" that was in the AI model information at the initial stage (Figure 3), "This model cannot output easy-to-read text," has been removed and replaced with the information, "This model can output easy-to-read text." We can also see that a new "restriction" has been added to the AI model information, "Faces and people in general may not be generated properly."
As a result, in the AI model information topics, "Text output" has been replaced with "Text output (updated),""Face generation (new)" has been added, and the corresponding part of the second table has been updated.
It can also be seen that a report to the user has been triggered.

<Module 3>
5 is a diagram showing the process flow of the third module (module 3) that provides a report to the user. Module 3 starts processing when a report to the user is triggered in module 2.

First, the second table (a table of AI model information topics x interest categories) is read from the second table repository 64 and referenced. In this example embodiment, it determines that the "Text Output" topic has been updated and that the "Face Generation" topic has been registered, and then determines the interest categories corresponding to these topics, namely, "Hard to Read Documents" and "Dysmorphic Forms."

Next, a process is performed to read and reference the first table (a table of user interests x interest categories) from the first table repository 62. In the example embodiment, it is determined that the interest category "Hard to Read Documents" is interest 3 for user B, and that the interest category "Urban Disfigurement" is interest 2 for user A, and users A and B are identified as targets for the report.
That is, module 3 determines which user to report to and what the user's interests are based on the interest categories affected by the updated or newly registered AI model information topics.

Figure 6 shows the processing flow (continuation of Figure 5) in the third module (Module 3) that provides reports to users. Module 3 creates a personalized report for each user based on the user's interests identified in the previous process and the parts of the AI model information that relate to updated or newly registered AI model information topics. As mentioned above, Module 2 stores in the topic repository 63 information indicating which parts of the AI model information correspond to the AI model information topics to be registered. Therefore, the parts of the AI model information that relate to updated or newly registered AI model information topics can be identified by referring to the topic repository 63.

Finally, module 3 sends the report to the identified users at the appropriate time.
An example of a report to be sent to a user is shown in the lower right of Fig. 6. A report is created for User A stating, "The AI model information has been updated, particularly the text in the section corresponding to the topic 'Face Generation'. This update addresses your 'Interest 2'," and a report is created for User B stating, "The AI model information has been updated, particularly the text in the section corresponding to the topic 'Text Output'. This update addresses your 'Interest 3'," and both reports are sent to each user.

Through the above processing, the information providing device 10 can individually provide personalized reports regarding updates to AI model information to each of user A and user B, who have interests in the AI model. Furthermore, this has the excellent effect of allowing user A and user B, who receive the reports, to easily understand the relevance of the reports to their interests (i.e., why the report was sent to them).

In addition, this embodiment also makes it possible to provide personalized reports regarding updates to AI model information to users who have an interest in AI models, thereby contributing to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), which is to "Build resilient infrastructure, promote sustainable industrialization and foster innovation."

Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments. Furthermore, the effects described in the above-described embodiments are merely a list of the most favorable effects resulting from the present invention, and the effects of the present invention are not limited to those described in the embodiments.

REFERENCE SIGNS LIST 1 Module for generating user interest categories 2 Module for generating AI model information topics 3 Module for providing reports to users 10 Information providing device 20 Control unit 30 Input unit 40 Output unit 50 Memory unit 61 Category repository 62 First table repository 63 Topics repository 64 Second table repository

Claims (3)

An information providing device that analyzes user interests in an AI model and AI model information related to the AI model, and when the AI model information is updated, individually creates and provides a personalized report to a user who is interested in the AI model based on the user interests related to the updated AI model information,
a first module for generating user interest categories based on user interests for an AI model and assigning the generated interest categories to the user interests;
a second module that generates AI model information topics based on the AI model information and the user's interest categories, assigns the generated AI model information topics to the corresponding interest categories, and triggers a report to the user according to updates of the AI model information;
a third module for identifying an interested user based on interest categories corresponding to the updated AI model information topics and the user's interests corresponding to the interest categories when a report to the user is triggered, and creating a personalized report for the identified user;
An information providing device comprising : 1. A computer-implemented information provision method that analyzes user interests in an AI model and AI model information related to the AI model, and when the AI model information is updated, individually creates and provides a personalized report to a user who is interested in the AI model based on the user interests related to the updated AI model information,
generating user interest categories based on the user's interests for the AI model and assigning the generated interest categories to the user's interests;
Generating AI model information topics based on the AI model information and the user's interest categories, allocating the generated AI model information topics to the corresponding interest categories, and triggering a report to the user according to updates of the AI model information;
When a report to the user is triggered, identifying an interested user based on interest categories corresponding to the updated AI model information topics and the user's interests corresponding to the interest categories, and creating a personalized report for the identified user;
Methods of providing information, including : An information providing program for causing a computer to function as the information providing device according to claim 1 .