1 Introduction

This study aims to elucidate the temporal evolution of policy themes by applying topic modeling using BERTopic (Grootendorst 2022) to the minutes of deliberation councils involved in the formation of Japanese environmental policy.

Japan's environmental policy spans diverse domains, including climate change and biodiversity. The Basic Environmental Plan, which sets these policy directions, is formulated through discussions by the General Policy Committee of the Central Environment Council under the Ministry of the Environment. As the Committee includes officials, experts, and public representatives, its more than 100 meetings provide useful material for tracing long-term policy trends and deliberative structure.

Discourse is a key element of policy formation (Majone 1989), and understanding deliberative opinion structures is therefore important (Fishkin 2009). As deliberative processes generate large-scale textual data, computational approaches have become effective tools for identifying thematic structures and dynamics (Parkinson et al. 2022). Building on these developments and the authors' previous BERTopic-based studies (Ishii / Kameda 2025a; Ishii / Kameda 2025b; Ishii / Kameda 2025c), this study seeks to provide a macroscopic understanding of the thematic composition of the meetings and their transitions by applying BERTopic to the complete set of minutes from the first session to the present.

2 Methodological Background

BERTopic has increasingly been used to analyze complex texts (Abuzayed / Al-Khalifa 2021; Egger / Yu 2022; Ishii / Kameda 2025a; Ishii / Kameda 2025b; Ishii / Kameda 2025c; Lee et al. 2023; Wang et al. 2024; Mokarrama et al. 2024; Li et al. 2024). By combining Sentence-BERT embeddings, UMAP, HDBSCAN, and c-TF-IDF, it can capture contextual information in text [1]. Previous studies show that this pipeline can extract fine-grained and interpretable topics even from Japanese policy minutes with layered contexts and uneven sentence length (Ishii / Kameda 2025; Li et al. 2024).

However, HDBSCAN tends to exclude noisy or compound sentences as outliers (labeled as -1). Rather than relying only on parameter tuning, this study used post-clustering probability distributions to reassess and selectively reassign outlier sentences. This made it possible to incorporate ambiguity while preserving topic robustness.

3 Data and Preprocessing

The data for analysis comprises the minutes of the Committee from the 1st to the 120th meeting (totaling 9,048 speeches). All texts, including secretariat explanations, committee member remarks, and Q&A sessions, were acquired via web scraping from the Ministry of the Environment's website.

As a preprocessing step, speeches were split into individual sentences. Short sentences with fewer than 20 characters, which contain sparse semantic content, were excluded, resulting in a total of 56,176 sentences for analysis. Each sentence was assigned a unique ID containing metadata such as the meeting number and speaker information, creating a structured dataset that enables time-series analysis and cross-referencing with speaker attributes.

4 Methods and Results

4.1 Topic Modeling

Sentence embeddings were generated using Sentence-BERT (sonoisa/sentence-bert-base-ja-mean-tokens-v2). Dimensionality was reduced to 5 dimensions using UMAP, followed by clustering using HDBSCAN with a minimum cluster size of 120. For topic representation, a count vectorizer and c-TF-IDF were applied. Noun extraction and stop-word definition were performed using MeCab with the custom dictionary mecab-ipadic-NEologd to capture contemporary terminology. As a result, 51 topics were extracted (Table 1: Major Topics). In the initial model, 12,498 sentences were classified into topics, while 43,677 sentences were labeled as outliers.

4.2 Reassignment of Outlier Topics

In BERTopic, the outlier label (-1) indicates that HDBSCAN did not assign a sentence to a sufficiently dense cluster with enough confidence; it does not necessarily mean that the sentence is semantically irrelevant. We therefore treated outliers as boundary cases and re-evaluated them using the topic-probability distribution returned by BERTopic's transform function. Sentences whose maximum topic probability exceeded a fixed threshold were reassigned to the most plausible topic, whereas sentences with near-competitive top probabilities were retained as "Ties" rather than being forced into a single category. Through this post-hoc procedure, 23,299 of the 43,677 initially outlying sentences were reassigned to non-negative topics, and 5,089 tie cases were identified. As a result, the number of topic-assigned sentences increased from 12,498 to 34,659, while 21,516 sentences remained unassigned.

These results suggest that a substantial portion of initially discarded sentences can be recovered efficiently by using information that BERTopic and HDBSCAN already provide, without rebuilding the clustering model itself. Exploratory inspection showed that the method is useful when several sentences from the same local discussion were labeled as outliers even though the block was clearly about a shared issue. In such cases, reassignment can bring those sentences back into the same topical neighborhood, making block-level aggregation and interpretation easier. In this sense, the procedure functions as a practical recovery layer for peripheral yet meaningful material that would otherwise remain outside thematic analysis.

At the same time, reassignment also acted as filtering. Among the 12,498 sentences that initially had non-outlier topic labels, 1,138 (9.11%) were reclassified as unassigned. Most of these were generic or procedural assignments, but some meaningful sentences were also removed, and some rescued sentences still appeared weak on close reading. Reassignment should therefore be treated not as a fully reliable automatic correction, but as a structured aid for interpretation that still requires selective human review.

4.3 Temporal Comparison of Thematic Composition and Speaker Analysis

Based on the reassigned data, five phases were defined using the revision dates of the Basic Environmental Plan as boundaries (Table 2).

Using this segmentation, we analyzed topic composition with a heatmap (Figure 2)

Figure 2 shows the topic distribution after reassignment. To facilitate comparison with the pre-reassignment results, supplementary heatmaps and detailed underlying data are provided in the GitHub repository: https://github.com/kou10111415/DH2026-repo.

and examined speaker attributes. Tie cases were retained during reassignment but apportioned equally across their tied topics for phase-level aggregation.

Several foundational topics, including Topic 12 ("Promotion of Environmental Education and ESD and Improvement of School Settings"), Topic 41 ("Information Sharing and Risk Communication"), Topic 45 ("Educational Institutions and the Position of Universities"), and Topic 50 ("Development and Social Implementation of Technological Innovation"), appeared throughout the period. Among them, Topic 50 rose sharply in Phase 5.

Phase-specific patterns show changing priorities. In Phase 1, Topic 37 ("Roles of NPOs and NGOs and Description of Their Activities") was especially prominent, reflecting strong interest in citizen participation. In Phase 4, Topic 0 ("Explanatory Discussion on Environment, ESG, and Financial Support") and Topic 38 ("Employment of Persons with Disabilities and Agricultural Support through Agriculture-Welfare Collaboration") were more prominent than in other phases, indicating a shift toward concrete social implementation. Topic 0 was strongly associated with regional banking and other financial practice. Topic 14 ("Division of Roles Between Municipalities and Prefectures and Regional Collaboration") also became prominent in Phase 4 and included speakers involved in regional renewable energy business, linking urban and regional planning to local decarbonization and project development. Topic 38 peaked in the 97th meeting because of testimony from an agriculture-, forestry-, and fisheries-related NPO witness.

External events were also reflected clearly in the distribution. Topic 20 ("Reconstruction and Environmental Restoration after the Great East Japan Earthquake") and Topic 40 ("Environmental Recovery and Monitoring of Radioactive Materials") became especially visible in Phases 2 and 3. These discussions included not only reports from government officials but also many contributions by researchers specializing in living environments, especially from the perspectives of reconstruction and urban planning. Policy instruments also shifted over time: Topic 7 ("Chemical Substance Management and Environmental Risk Reduction") and Topic 31 ("Consideration of the Global Warming Countermeasures Tax and Environmental Tax") were more frequent in Phases 1 and 2, whereas Topic 2 ("Regional Circular and Ecological Sphere and Decarbonized Energy") became more prominent in Phases 4 and 5.

Education-related topics followed a non-linear trajectory. In particular, Topic 12, which focused on environmental education, occupied a relatively large share in Phases 1 to 3 and remained visible afterward. The top speakers for Topic 12 and Topic 45 included former university presidents, and an elementary school principal also appeared among the major contributors. This suggests that education-related discussion was supported not only by administrative actors but also by participants with direct educational leadership experience.

In Phase 5, Topic 50 became one of the largest topics. Its leading speakers included an environmental economist and participants connected to labor-related perspectives, and their remarks repeatedly referred to the "well-being" orientation of the Sixth Basic Environmental Plan. Innovation was therefore discussed not only in technological terms but also in relation to quality of life and the integration of environmental, economic, and social values. Overall, topic weights varied substantially across phases and were linked to speakers whose expertise matched the issues under discussion.

5 Conclusion and Future Perspectives

This study clarified the temporal evolution of themes in environmental policy deliberation and changes in speaker structure by integrating topic modeling, outlier reassignment, and speaker information.

The analysis showed that changes in discussion were closely linked to changes in the expertise and roles required in each phase. Education, finance, and urban or regional planning topics were characterized by relatively large speaking weights among actors with specific backgrounds, while the rise of Topic 50 was closely associated with references to well-being in the Sixth Basic Environmental Plan. The proposed method usefully shows who spoke about which topic, and when, and should also be applicable to time-series text data in other policy fields.

However, technical challenges remain. Sentence-level analysis alone was insufficient to represent abbreviations such as “ESD” (Education for Sustainable Development), and fillers specific to spoken language occasionally affected embedding accuracy. Future work should introduce embeddings better suited to conversational text and broader contextual units.

Acknowledgment

This work was supported by NIHU Digital Humanities Projects, JSPS KAKENHI Grant Number 26KJ0683, Doctoral Program for World-leading Innovative & Smart Education (WISE Program), Chiba University: Applied Humanities Program for Cultivating Global Leaders, and Nippon Foundation HUMAI Program.