Other languages
The YoronChizu(Public Opinion Map) is a service released on November 18, 2024, by Mielka, a Japanese nonprofit organization. Founded in 2016, Mielka aims to enhance democracy in Japan by visualizing politics, primarily led by young people. Since 2017, it has operated JAPAN CHOICE, one of Japan's largest election information platforms, which served 1.6 million users during the 2021 House of Representatives election.
The Public Opinion Map launched as an experimental feature of JAPAN CHOICE, allows users to anonymously vote on specific issues extracted from party manifestos during the 2024 House of Representatives election period. The voting data was used to visualize opinion distributions.
Over about two weeks, 4,403 unique users participated. This repository provides open access to that voting data.
The CSV files and images in this repository are licensed under CC BY 4.0. For usage instructions, please refer to README.ipynb.
Although new votes are no longer being recorded, the Public Opinion Map remains accessible for exploration: Public Opinion Map (Beta) - JAPAN CHOICE
The service uses Polis, a deliberation tool developed under the leadership of Colin Megill, on the server side. Polis gained recognition after its adoption in Taiwan's civic engagement project, vTaiwan.
The front end was redeveloped from scratch to suit Japan's context. Analysis of JAPAN CHOICE's user access revealed that 85% of the 3 million unique users accessed election information via mobile devices, necessitating optimization for small smartphone screens.
Additional features include:
- Visualization of political party stances using party icons
- AI-driven explanations of opinion clusters
- Rapid updates of user opinion position on the map
The Public Opinion Map extracted voting issues directly from party manifestos. Each party’s stance —support, opposition, or neutral— on each issue was encoded into data. This information was used to overlay political parties stances on the user opinion distribution scatterplot.
Polis-generated clusters can be difficult for non-expert users to interpret. To address this, the data on "opinions characterizing clusters" generated by Polis was passed to an LLM, which provided a description of the opinions held by each group. It also summarized these descriptions into concise group names, which were displayed on the scatterplot.
By embedding projection matrix information from PCA on the client side, user positions on the scatterplot could be computed without querying the server. This enabled real-time screen updates in response to user voting actions.
As a result, even though the Polis computation server is now inactive, the YoronChizu can still be experienced. This approach also allows the service to be demonstrated using the front end alone after the election period.
This figure visualizes opinion groups on economic policy. The lower-right cluster lacks party icons, indicating that this group's views are not well-represented by existing political parties. This suggests that public opinion in this group is not adequately reflected in national policy discussions. AI-generated analysis of this group is as follows:
Team Yellow: Skeptics on Price Pass-Through Measures
🤖 This group is relatively cautious about wage increases through measures supporting price pass-through for small businesses. They also oppose reductions or abolishment of consumption tax and are skeptical about providing subsidies to low-income households.
Observations from previous Polis results raised concerns that opinion distributions might converge over time into two obvious opposing structures. This occurs when increasing data points lead PCA results to form an ellipsoid lacking clear boundaries, causing k-means clustering to split clusters along the first principal component axis.
However, in this experiment, even with 4,400 votes, four clusters remained distinct. The voting issues in this experiment were drawn from party manifestos, and no new issues were added during the voting period. This indicates that the convergence problem is more likely caused by the addition of voting issues and higher dimensionality, rather than increased user data.
Statistics were recorded hourly during the experiment. The chart below shows the relationship between the number of users and the number of clusters. It took approximately 2,500 users for the clusters to stabilize at four.
On digital democracy—a topic underrepresented in party manifestos—JAPAN CHOICE team set voting issues. To avoid problems caused by insufficient data, party icons were intentionally excluded from this map.
Team Red: Ad Regulation & AI Skeptics
🤖 This group supports regulating targeted internet ads and remains cautious about AI-driven policy-making and automation. They prioritize regulations and cautious adoption over technological convenience.
Team Blue: Advocates for Digital Integration & Efficiency
🤖 This group strongly supports digital integration through My Number utilization and online administrative services to enhance efficiency. They also back adopting wearable cameras and AI technologies.
Team Yellow: AI Utilization & Personalized Optimization
🤖 This group actively supports automating policy decisions and visualizing public opinion using AI. They emphasize individualized approaches, such as enhanced tracking of performance metrics, to enable efficient decision-making.
Team Green: Privacy Protection Advocates
🤖 This group values safeguarding individual control over personal information and strengthening privacy laws to European standards. They support regulating online advertising and democratic decision-making through participatory agenda-setting systems.
The process of extracting unbiased opinions from party manifestos required careful human effort. While LLMs could assist in this process, their potential for hallucination poses significant risks, necessitating cautious experimentation. In tests with LLM-generated content on digital democracy, the output quality was deemed insufficient, and the feature was not utilized.
Concerns about hallucinated misinformation disadvantaging specific parties during the election led to the implementation of human reviews before publication. Updates were released three times during the election, sacrificing the real-time nature of Polis.
One challenge noted was that LLMs often misinterpreted clusters with complex opinion distributions. For instance, a cluster characterized by "70% support, 10% opposition within the cluster, but 90% support, 3% opposition outside" might be better described as "relatively higher opposition," yet LLMs tended to interpret this as "majority support." Further refinement in data representation for prompts is needed.
For similar reasons, user-submitted opinions were not accepted. While the ideal of [opening agenda-setting power to the people] remains desirable, real-time acceptance introduces challenges, including quality assurance, misinformation, and higher dimensionality with increased voting issues. Further research is required to address these challenges.