01. 연구 (2) 원전 정책 담론 연구 03 연구 방법론

이 글은 이전 문서(00. 개요)의 후속 내용으로,

해당 연구 설계 및 구현 구조를 보다 구체적으로 정리했다.

 

블로그에 공개한 01–06 문서는 당시 컨택 과정에서 제출했던 18페이지 분량의 영어 연구계획서 가운데

방법론과 연구 구현에 해당한다. 

최종 제출본에는 분량 제한 때문에 일부만 반영되었지만,

방법론 부분이 가장 오래 수정 및 검토되었다.

 

연구계획서를 작성하는 일련의 과정 속에서,

나는 연구 과정을 스스로 설명할 수 있어야 한다는 점을 중요하게 여겼다.

 

예상과 다른 결과가 나오면 원인을 다시 확인했고, 구현 과정이 바뀌면 방법론도 함께 수정했다.

연구 질문과 분석 구조가 맞지 않는다고 판단한 부분은 여러 번 다시 설계했다.

 

그래서 블로그에 공개한 문서에는 완성된 결과보다 시행착오와 수정 과정, 

그리고 연구를 진행하면서 마주했던 한계와 고민이 더 많이 남아 있다.

 

본론

---

03. Research Methodology

---

가장 오래 붙잡고 수정했던 문서다. 

구현을 반복할수록 논문 몇 줄로 설명할 수 있는 수준이 아니라는 것을 체감했고, 

데이터 구축과 분석 절차도 여러 차례 바뀌었다. 

특히 Data Reduction, Core Keyword, Analysis Layer를 어떻게 구분해야 

분석 과정이 연구자의 임의적 판단으로 보이지 않을지 가장 많이 고민했다.

 

다만, 블로그에서는 모든 과정을 공개하지 않게 되었고,

연구 과정에서 정리한 방법론 문서를 공개용으로 재구성했다.

 

실제 연구에 적용한 분석 환경과 구현 과정은 일부 생략하였으며,

연구 질문을 어떠한 분석 구조로 설계했는지에 초점을 맞추었다.

 

연구 노트에 시행착오와 구현 기록을 별도로 정리할 예정이며,

본 문서는 연구 설계의 흐름을 공유하기 위한 목적으로 공개한다.

 

이와 함께 연구계획서의 방법론적 배경이 된 논문과

그 과정에서 형성된 연구 질문의 흐름도 함께 제시하고자 한다.

 

(여러 논문을 읽으며 연구 질문이 구체화되었지만, 

이 글에서는 올해 초 소개했던 DSH 논문과

실제 Research Proposal로 이어진 과정을 정리한 글만 함께 소개한다.)

 

▶ [올해 초 읽었던 DSH 논문, 그리고 실제 내 Research Proposal까지 이어진 연구의 흐름]

https://uourstar.tistory.com/40

디지털 인문학 공부 : 올해 초 읽었던 DSH 논문, 그리고 실제 내 Research Proposal까지 이어진 연구의

 

---

1. Research Design

This study conducts a text-based policy analysis using news data to examine changes in social discourse surrounding South Korea’s radioactive waste policy from 1994 to 2025.

To integrally analyze the relationship between policy events and media discourse, this study adopts a Policy–Discourse Integrated Analysis Framework, which combines the following three approaches:

• Event-based Policy Analysis
• News Data–Driven Text Mining
• Keyword-based Discourse Structure Analysis

Furthermore, this study develops a web-based Nuclear Policy–Discourse Analysis System that integrates policy, news, and event data, thereby establishing an environment for data collection, processing, analysis, and visualization.

Notably, this system functions not merely as a visualization tool, but as an Analytical Execution Environment, in which data processing logic (e.g., Data Reduction, Filtering, TF-IDF analysis) is integrated at the code level.

Meanwhile, this study does not aim to verify a direct causal relationship between policy events and discourse changes. Although it is possible to interpret discourse changes before and after specific events based on temporal ordering, such interpretation remains at an explanatory and exploratory level rather than constituting causal verification.

Accordingly, the findings of this study are limited to the interpretation of correlation and co-variation structures, and do not extend to causal claims that policy events “trigger” discourse changes.

This limitation arises from the methodological condition that the study is based on observational data rather than an experimental or quasi-experimental design, which is explicitly acknowledged.

The research methodology is grounded in a mixed-methods approach that combines quantitative content analysis with qualitative discourse analysis. Quantitative analysis examines the distribution and change of discourse through keyword frequency and TF-IDF–based importance measures, while qualitative analysis complements this by interpreting semantic structures within textual contexts.

These two approaches are integrated through a complementary triangulation structure, whereby quantitative results are not interpreted independently but in conjunction with qualitative context. This design helps mitigate potential methodological bias that may arise from relying on a single analytical approach.

---

2. Data Source

This study collects data using BigKinds, a news big data platform operated by the Korea Press Foundation.

BigKinds provides a structured database of articles from major domestic media outlets, including metadata such as article titles, body text, media source, publication date, and URL, as well as core keywords and feature values (i.e., word importance). It also supports functionalities such as core keyword extraction, inter-article relationship analysis, and Named Entity Recognition (NER) based on text analysis algorithms.

In this study, BigKinds is treated as a data infrastructure rather than a data generator. Accordingly, the keywords and importance values generated by the platform are used only as auxiliary inputs during the preprocessing stage, rather than as final analytical outputs.

The analysis period spans from 1994 to 2025, based on the following considerations:

• Data Stability: reflecting the period after the mid-1990s, when BigKinds news data began to be systematically accumulated
• Policy Importance: corresponding to the period during which radioactive waste policy emerged as a major social issue in South Korea
• Analytical Continuity: ensuring sufficient continuity for long-term time-series analysis

News data is used as the primary medium through which social discussions on policy issues are produced and disseminated; in this study, it is treated as observational data for public discourse. However, it is interpreted as a media-mediated discourse structure, rather than as a direct representation of overall societal perception.

Additionally, the keywords and importance values generated by the platform’s internal algorithms are understood as estimated indicators rather than perfectly objective measures. Therefore, this study uses them as auxiliary inputs and avoids over-interpretation.

At the same time, by ensuring that similar datasets can be reconstructed when data are repeatedly collected using identical search queries, this study secures procedural reproducibility in the data collection stage.

3. Unit of Analysis

The basic Unit of Analysis in this study is the News Article. A news article is regarded as the fundamental textual unit through which social discourse on policy issues is produced and disseminated. In particular, news media serves as a key medium that conveys social conflicts, policy-making processes, and policy discourse related to policy issues to the public sphere, and thus article-level analysis is widely used in policy research.

In existing research as well, the approach of setting individual news articles as the basic unit of analysis has been extensively adopted. For example, Hopmann et al. (2021) pointed out that “existing studies treat individual news articles as the basic unit of analysis,” thereby presenting the academic validity of article-based analysis.

---

In this study, the ‘Unit of Analysis’ and the ‘Level of Analysis’ are conceptually distinguished.

The Unit of Analysis refers to the physical object (news article) in which data is actually observed and collected, whereas the Level of Analysis refers to the interpretive layer in which collected data are combined to derive meaning.

In this regard, Trilling and van Hoof (2020) emphasized that attention should be paid not to individual articles themselves, but to sets of articles dealing with specific events, suggesting the necessity of reconfiguring article units into event-level units depending on the purpose of analysis.

---

Accordingly, in this study, an ‘Event’ is defined not as a new unit of analysis, but as an aggregation rule applied to the single unit of analysis (news articles). In other words, this study does not adopt a multi-unit structure; rather, it follows a single-unit, multi-level analysis structure, in which a single unit of analysis is aggregated at the event level and interpreted accordingly.

This multi-level analytical system is structured as follows:

Analysis Layer Analysis Target Description

1st Analysis (Micro)	News Article	생략
2nd Analysis (Meso)	Article Keywords & Attributes	생략
3rd Analysis (Macro)	Event-based News Corpus	Discourse structure aggregated 생략

---

Through this structure, this study constructs an Event-based Discourse Analysis Framework that analyzes the relationship between policy events and discourse structures based on news article data. At this stage, the relationship between events and discourse is interpreted not as a direct causal relationship, but as co-variation and structural association of discourse changes following the development of policy events.

---

References

• Hopmann, D. N., Shehata, A., & Strömbäck, J. (2021). Journalism & Mass Communication Quarterly, 98(3), 735–755.
• Trilling, D., & van Hoof, M. (2020). Digital Journalism, 8(10), 1317–1337.

---

4. Search Query Design

To simultaneously ensure both comprehensiveness and accuracy in news data collection, this study designed a two-stage search strategy based on the concepts of Recall and Precision from Information Retrieval (IR) theory.

---

4.1 Conceptual Application

In IR theory, Recall refers to the degree to which relevant documents are comprehensively retrieved, while Precision refers to the accuracy of the retrieved results; these two indicators generally exist in a trade-off relationship.

In this study, these concepts are not used as formula-based evaluation metrics, but rather as a methodological framework to explain the balance between comprehensiveness and accuracy in the data collection process.

---

4.2 Primary Search: Recall-Oriented Comprehensive Collection

The primary search is an exploratory collection stage aimed at maximizing Recall in order to minimize the omission of policy-related discourse.

• Strategy: Constructing search queries centered on general policy terms and core topic keywords
• Example (HLW Act, 2013–2019): (High-level OR radioactive waste) AND (public deliberation OR social consensus OR management policy)
• (고준위 OR 방사성폐기물) AND (공론화 OR 사회적합의 OR 관리정책)
• Role: Forming the reference set of the overall discourse and minimizing selection bias

---

4.3 Secondary Search: Precision-Oriented Refinement

Since the primary search results may include data with low direct relevance to policy issues, a Precision enhancement stage reflecting the context of policy events is applied.

• Strategy: Filtering based on core keywords such as locations, actors, and institutional elements for each event
• Example 1 (Buan Conflict 2003–2005): Buan AND (radioactive waste OR nuclear waste) AND (protest OR clash OR referendum)
• `부안 AND (방폐 OR 핵폐기물) AND (시위 OR 충돌 OR 주민투표)`
• Example 2 (Gyeongju Site Selection 2005–2008): Gyeongju AND (radioactive waste) AND (hosting OR acceptability OR referendum)
• 경주 AND (방폐 OR 방사성폐기물) AND (유치 OR 수용성 OR 주민투표 OR 지원)
• Core Design:
  Rather than relying on broad keywords, AND conditions are restructured around core nouns that constitute the event, thereby strengthening contextual relevance
• Effect:
  Removing non-related data such as general controversy reports and securing high-density discourse directly related to policy events

---

4.4 Dataset Construction and Validity

This study ensures analytical accuracy by constructing a dataset refined through the secondary search.

Meanwhile, the primary search results are not discarded but maintained as a reference validation set, used to verify whether the analysis dataset sufficiently reflects the overall discourse context.

The subsequent Data Reduction stage functions as a quantitative refinement process, applying statistical importance based on TF, weight, and Level, grounded in these qualitative criteria (search query design).

---

4.5 Design Principles and Methodological Control

The search query structure in this study was not adjusted ex-post, but rather pre-defined following literature review and policy event definition.

Furthermore, the composition of search queries is not based on arbitrary choices by the researcher, but on a structural design reflecting key actors, spaces, and institutional elements that constitute policy events.

Through this multi-stage structure, the study distributes the influence of query subjectivity across stages, thereby ensuring methodological control in the data collection process.

---

4.6 Hierarchical Structure of Constructed BigKinds Search Queries (Unit of Analysis)

5. Dataset Construction

The data construction process of this study is composed of a multi-stage preprocessing pipeline, which includes Data Reduction following search-query-based data collection.

This structure is designed to systematize the entire process from data collection to analysis in a step-by-step manner and to clearly distinguish the function of each stage.

Each stage is configured as an independent processing unit, while simultaneously being integrated into an interconnected pipeline structure, thereby ensuring consistency and reproducibility of the analytical process.

---

5.0 Overall Data Construction Procedure

Literature and Theoretical Review
↓
Search Query Design (Primary & Secondary)
↓
News Data Collection (BigKinds)
↓
Data Reduction (TF, Weight, Keyword Level-based)
↓
Data Cleaning (Deduplication and Structuring)
↓
Core Keyword System Construction (Normalization + Intersection)
↓
News Corpus Construction
↓
Analytical Filtering
↓
Discourse Analysis (TF-IDF and Time-series Analysis)

The above procedure follows a hierarchical structure consisting of:

Collection → Reduction → Cleaning → Conceptualization → Analysis

---

5.1 Data Collection and Analytical Environment

This study uses news articles and metadata collected through the BigKinds platform as the foundational corpus.

The collected data is transformed into an analyzable dataset through a multi-stage preprocessing procedure.

The analytical environment is configured as follows:

• 생략
• Web: Visualization support for analytical results

This environment is not a simple combination of tools; rather, it is designed as an execution pipeline that guarantees identical input–identical output, thereby ensuring reproducibility.

---

5.2 Data Reduction

(1) Stage Definition and Tool Development (Pipeline & Tooling)

In this study, a Python-based data preprocessing and keyword optimization tool was developed and applied in order to resolve issues of excessive data volume and noise arising from search-query-based news collection, and to ensure analytical efficiency.

The tool is designed as a rule-based processing system executed in a GUI environment, automating a series of processes including data merging, keyword importance calculation, and article filtering.

-생략-

---

(2) Integrated Importance Calculation Algorithm (Method Formalization)

Keyword importance is evaluated not through a single metric but via a hierarchical ranking structure.

First, the Level value is set as the primary criterion to reflect conceptual hierarchy. Within the same Level, relative importance is calculated through normalized TF and Weight values.

The core formulation is as follows:

• -생략-

This structure represents a combination of Level-based priority ordering and TF/Weight normalization.

---

(3) Data Filtering and Sorting (Pipeline Detail)

The derived top-ranked keywords are directly used for article filtering.

For each article, a text field is constructed by combining title, body, and keyword metadata (keywords_raw), and article inclusion is determined based on whether core keywords appear within the text.

-생략-

---

(4) Data Scale Optimization Rules (Targeting Rules)

This study does not apply a single threshold for data reduction. Instead, adaptive targeting rules are applied depending on the size of the source dataset.

Source Data Size (n) Retained Data

-생략-

These variable rules are designed to control data noise during periods of excessive media coverage of policy events, while preventing information loss during low-coverage periods, thereby ensuring continuity of discourse across the entire analysis period.

---

(5) Methodological Justification

The Data Reduction methodology of this study is justified in the following respects.

First, reproducibility is ensured.

The Python-based scripts and rule-based processing structure are designed so that identical inputs generate identical outputs, ensuring consistency.

Second, clear distinction of indicators is maintained.

TF used in the Data Reduction stage refers to frequency derived from BigKinds metadata and is not identical to TF used in TF-IDF analysis at the later stage, as both differ in calculation method and purpose.

Third, analytical efficiency and noise reduction are achieved.

Stopword removal, exclusion of low-ranked keywords, and priority-based article selection reduce noise in advance, thereby improving the accuracy of subsequent Core Keyword-based semantic analysis.

However, this process is explicitly based on a researcher-defined rule-based design, and reflects methodological decisions regarding keyword thresholds, stopword definitions, and matching rules.

---

5.3 Data Cleaning

(1) Stage Definition (Pipeline)

The data cleaning process includes the following preprocessing steps:

• Removal of duplicate articles
• Removal of missing data
• Sorting by publication date
• Structuring of article metadata

The cleaned dataset is structured for integration into an SQL-based relational database for subsequent analysis.

---

(2) Method and Structure

The criteria for data cleaning (duplicates, missing values, etc.) are based on predefined rule-based filtering, and are designed to be consistently applicable under identical conditions.

In addition, the SQL database is not merely a storage system, but functions as an analytical support layer, enabling:

• -생략-

---

5.4 Core Keyword Construction

(1) Stage Definition (Pipeline)

Core Keywords are defined as the intersection of the following two sets:

• Policy keywords derived from search queries (policy concepts)
• Data-driven top keywords (based on TF, weight, and Level)

The intersection is derived following a concept normalization process (concept normalization).

---

(2) Method Definition

The search-query-based policy keywords used in this study constitute a set of secondary search query keywords, designed through literature review and theoretical validation. These represent a conceptual framework defined a priori by the researcher, rather than data-derived outputs.

In contrast, data-driven keywords reflect actual expressions in news texts, derived through BigKinds metadata and Python-based preprocessing.

Accordingly, Core Keyword construction functions as a process of aligning theory-based concepts (query keywords) with data-based expressions (textual keywords).

In this process, concept normalization goes beyond simple word replacement and involves a process of conceptual categorization, whereby heterogeneous expressions are converged into higher-level policy concepts.

Core Keywords are constructed based on top-ranked keywords from Data Reduction. TF-IDF is not directly used in their derivation but is utilized as a supplementary indicator in subsequent discourse analysis.

Thus, Data Reduction reduces data scale, while Core Keyword construction ensures conceptual consistency and semantic alignment.

---

(3) Derivation Mechanism (Pipeline + Method)

The Core Keyword derivation process is structured as follows:

Search Query Keywords (Policy Concepts)

↓

TF-based Top Keywords (Data Expressions)

↓

Concept Normalization (Expression → Concept Transformation)

↓

Concept Integration

↓

Intersection Derivation

↓

Final Core Keyword Set

Intersection is derived based on:

• Data keyword set generated via Python (automatic extraction)
• Search query keyword set defined by the researcher (theoretical framework)

Matching is first performed computationally via Python to generate candidate sets, and final confirmation is conducted through manual validation by the researcher.

This process aligns heterogeneous textual expressions with unified conceptual structures.

---

(4) Methodological Definition (Justification)

Search query keywords do not function as variables that directly generate analytical outcomes. Instead, they serve as a reference framework for alignment with data-driven keywords.

Core Keywords function as a structural alignment mechanism that bridges gaps between empirical data expressions and theoretical policy concepts, while also serving as a validation layer for interpretability.

Therefore, Core Keywords do not operate as restrictive filters; rather, they function as a structural supplement that enhances interpretability.

This process is semi-structured rather than fully automated, and is based on a controlled validation procedure, not arbitrary researcher subjectivity.

---

5.5 Theoretical Foundation

(1) Keyword Extraction

BigKinds extracts core keywords using morphological analysis and Named Entity Recognition (NER).

However, these outputs are algorithm-generated estimates and therefore cannot be considered fully objective indicators.

---

(2) TF-IDF-based Importance Theory

TF-IDF is a representative method for measuring relative word importance.

However, it does not fully capture contextual meaning; therefore, in this study, it is triangulated with qualitative discourse analysis.

TF is distinguished into two forms:

• Data Reduction stage: TF based on BigKinds metadata
• Analysis stage: TF used in TF-IDF computation

These are not identical indicators in terms of purpose or calculation.

---

(3) Concept Normalization Theory

Concept normalization refers to the process of integrating heterogeneous textual expressions into standardized conceptual units.

Since news discourse tends to express identical policy objects through multiple linguistic forms, such integration is essential.

In this study, concept normalization is not based on researcher intuition but is performed through iterative validation using literature, policy documents, and contextual news analysis.

This ensures reduction of noise and improves analytical validity.

---

5.6 Practical Example of Core Keyword Derivation

This study applies a dual alignment structure combining policy concepts (search queries) and data-driven expressions (textual keywords).

Policy Event Search Keywords Data-driven Keywords Concept Normalization Core Keywords

일부 생략
HLW Special Act (2024)	HLW, radioactive waste, bill, parliament, passage, negotiation	Special Act, HLW Act, spent fuel, Kim Seok-ki	Special Act→bill / spent fuel→radioactive waste	HLW, radioactive waste, bill, parliament
고준위 특별법 (2024)	고준위, 방사성폐기물, 법안, 국회, 통과, 협상	특별법, 준위방폐물법, 사용후핵연료, 김석기 의원	특별법/준위방폐물법→법안 / 사용후핵연료→방사성폐기물	고준위, 방사성폐기물, 법안, 국회

 

---

5.7 Corpus Construction and Database Finalization (Data Finalization)

(1) Stage Definition (Pipeline)

After completion of all Data Reduction and Data Cleaning processes, the final dataset is constructed as a CSV-based corpus and structured into an SQL relational database.

This corpus is finalized as the Original Corpus, serving as the foundational dataset for all subsequent analyses.

---

(2) Method and Significance

This stage is not merely data storage but a structural process ensuring that identical queries produce identical outputs during analysis.

Through this, reproducibility is guaranteed, and the corpus serves as the reference baseline for all subsequent filtering and analytical procedures.

6. Analytical and Visualization Environment (System Layer)

This study constructed the following technical environment in order to efficiently analyze the established reference corpus and interpret the results.

---

6.1 Analytical Environment

An integrated analytical environment was established by linking Python libraries with an SQL database to perform statistical analysis and text mining.

---

6.2 Visualization Layer

A web-based visualization tool is utilized to visually represent time-series changes and keyword importance in order to support the interpretation of analytical results.

This constitutes the final stage of an integrated analytical flow in which textual data is transformed into visual information through processes of quantification and structuring.

---

7. Analysis Data Preprocessing and Sample Structure (Analysis Layer)

This study defines a stage in which a sub-dataset is selected from the finalized corpus in accordance with analytical objectives in order to ensure stability of analysis and consistency of interpretation.

This stage is not a simple data cleaning process, but a structural preprocessing stage involving the determination of analytical units and the selection of analysis targets.

---

7.1 Sample Structure Definition

The sample in this study is not defined as data used for quantitative analysis itself, but as a qualitative analysis target for interpreting quantitative results.

The sampling and keyword filtering conducted in this stage are distinguished from the Data Reduction stage, which is responsible for controlling data scale and noise.

Instead, this stage corresponds to a research design phase that defines analytical units consistent with the research questions.

This study applies an importance-based sampling strategy to select analytical targets from the finalized corpus.

The sample size is not based on statistical inference criteria but is used as a design rule for determining the analytical method.

---

(1) Sampling Rationale

In this study, approximately 100 documents with the highest importance among candidate documents were selected as core analytical targets.

This criterion is grounded in the following studies:

• Text mining allows identification of core structures even in small corpora due to repetitive patterns and word importance (Qaiser & Ali, 2018; Lee & Kim, 2009).
• TF-IDF analysis focuses on information content per word rather than dataset size, enabling meaningful results even with relatively small document sets (Aggarwal, 2018).

생략

---

(2) Analysis Method by Sample Size

Number of Articles (N) Analysis Method Description

생략

---

(3) Methodological Rationale

The N-based analytical structure defined in this study has the following methodological validity:

생략

---

7.2 Keyword Filtering-Based Selection of Analytical Dataset (Filtering Pipeline)

Since the finalized corpus may still contain noise not directly related to policy, a multi-stage keyword filtering process is applied to select the final analytical dataset.

The keyword filtering conducted in this stage is a process of generating a sub-dataset optimized for specific research questions, while preserving the integrity of the original corpus (CSV).

---

(1) Policy Keyword Filtering

In the first stage, news data directly related to policy is selected based on core policy keywords.

---

(2) Discourse Keyword Filtering

In the second stage, discourse-related keywords (e.g., policy conflict, acceptance, safety discourse, international norms) are applied to extract texts directly connected to the structure of policy discourse.

This is not a simple relevance filter, but a process for extracting texts directly linked to policy discourse structures.

---

(3) Optional Refinement Filtering

If the dataset is excessively large and contains high levels of noise, additional filtering is optionally applied based on predefined thresholds.

This is not part of the data construction stage, but a selective process within the analytical stage that reconstructs sub-datasets according to analytical objectives.

In other words, it functions as an analytical decision-making process addressing both quantitative (data scale reduction) and qualitative (discourse selection) issues.

---

7.3 Integrated Significance of Preprocessing Stage

Analytical preprocessing in this study consists of the integration of the following two components:

• Sample structure definition: determination of analytical method based on data size (N)
• Keyword filtering: definition of analytical dataset aligned with research objectives

---

This process reflects a rule-guided structure, in which researcher design is incorporated.

By applying filtering criteria consistently, the study minimizes subjective bias and ensures validity of analysis.

---

8. Text Mining Method

This study conducts TF-IDF-based word importance analysis for news text analysis.

TF-IDF is a representative text mining technique that calculates the relative importance of words in a document by combining term frequency (TF) and inverse document frequency (IDF).

---

8.1 TF-IDF Theory and Conceptual Limitations

Although TF-IDF is useful for quantitatively measuring word importance, it has limitations in that it does not fully capture contextual meaning.

Therefore, in this study, TF-IDF results are not used as an independent interpretive tool but are combined with qualitative discourse analysis through triangulation.

TF-IDF results are interpreted not as direct semantic representations but as indicators of keyword distribution structures constituting discourse.

Accordingly, TF-IDF values are treated not as explanatory evidence, but as exploratory signals that guide interpretation.

---

8.2 Definition of TF

(see Section 5.5)

---

8.3 TF-IDF Analysis Design

TF-IDF analysis in this study is conducted in a Python environment based on an SQL-structured corpus, following these steps:

• Tokenization of news text
• Stopword removal
• Generation of TF-IDF matrix
• Calculation of keyword importance

This constitutes a quantitative analytical stage aimed at deriving core word structures that form policy discourse.

---

8.4 Execution of TF-IDF Analysis

In this stage, TF-IDF analysis is performed on the final corpus after data reduction and cleaning.

This enables the derivation of keyword importance and discourse structure changes by policy event, which are subsequently used as foundational data for time-series and comparative analysis.

---

9. Event-based Policy Discourse Analysis

This study applies an Event-based Policy Discourse Analysis design that focuses on changes in news discourse centered on policy events.

Policy events are defined as policy turning points that influence discourse structures, including policy decisions, social conflicts, and changes in international norms.

Representative events include:

• Gulupdo nuclear waste controversy (1994–1997)
• Buan nuclear waste conflict (2003–2005)
• Gyeongju site selection for nuclear waste facility (2005)
• Fukushima nuclear accident (2011)
• High-level radioactive waste special act discussions (2013–2025)

This event-based analysis is designed to compare patterns of changes in media coverage and discourse structures based on policy turning points.

Event selection is not determined post hoc based on results, but is pre-defined based on theoretical criteria regarding policy turning points.

Furthermore, event classification is not treated as a unique or fixed truth but as a result of analytical framework selection.

Importantly, the comparison of pre- and post-event discourse is not designed for causal inference.

Rather, it is based on observational data analysis, aiming to explore temporal ordering and structural co-variation in discourse changes.

Therefore, this study does not test whether policy events directly cause discourse changes, but instead examines temporal sequences and structural co-variation patterns between events and discourse evolution.

Additionally, event selection is not intended for representativeness but corresponds to theoretical case selection for explaining policy discourse structures.

---

10. Summary

The data processing structure of this study is a dual-structured methodology that aligns concept-based policy representations with data-driven keyword expressions to ensure analytical coherence.

This constitutes a hybrid analytical design that integrates theory-driven and data-driven approaches.

Furthermore, this study does not separate automated text analysis from researcher interpretation, but instead combines them within a unified structure, proposing an integrated analytical model.

Through this design, the study minimizes reliance on a single analytical technique and ensures methodological consistency and structural stability across the entire analytical process.

The analytical processes described in this document are supported by the system architecture outlined in the Research Infrastructure document.

Detailed database schemas and implementation specifications are provided in Appendix A. Research Infrastructure.

For system-level execution and visualization, refer to the System Implementation and Demonstration.