상세 컨텐츠

본문 제목

01. 연구 (3) 북한 매체 담론 분석 01 연구계획서 (26ver.)

본문

이번 글에서는 2026년 5월에 수정한 연구계획서의 핵심 내용을 다룬다.

 

앞서 소개했던 2025년 12월 원본 연구계획서 당시의 문제의식과 한계를 그대로 담고 있었다면,

2026년 5월 수정본은 그 문제와 한계를 어떻게 보완했는지가 잘 드러나는 문서라고 할 수 있다.

 

본론에서는 먼저 2025년 12월 연구계획서와 2026년 5월 수정본을 비교한 뒤,

수정된 연구계획서 일부를 소개하고자 한다.

 

2025년 12월에 작성한 연구계획서는 일제강점기 담론 연구를 수행하기 전에,

분석 구조를 먼저 검증하기 위해 설계한 연구였다.

 

당시에도 연구 질문 자체는 지금과 크게 다르지 않았다.

 

다만 일제강점기 데이터셋을 바로 구축하기에는 OCR 오류, 국한문 혼용체, 형태소 분석 등 해결해야 할 문제가 많았기 때문에,

비교적 안정적으로 구축할 수 있는 현대 한국어 자료를 이용하여 분석 구조를 먼저 설계하고자 했다.

 

또한 이 블로그에서 여러 차례 언급한 SCV(Structural Covariation) 역시

원래는 이 연구에 포함시키고자 했던 개념이었다.

 

나는 4학년 2학기 「AI와 인간언어」 수업을 들으며 이 아이디어를 구상했고,

처음에는 일제강점기 신문 담론 연구에 적용하는 것을 목표로 하고 있었다.

 

4학년 2학기가 시작될 당시만 해도

학기 말까지 한 교수님께 연구계획서를 제출하기로 되어 있었기 때문에,

일부러 역사 데이터베이스 수업을 들으며 일제강점기 담론 연구를 위한 연구계획서를 함께 구상하고 있었다.

그러나 일제강점기 데이터셋 구축이 예상보다 훨씬 어려웠고,

결국 연구 대상을 현대 자료로 바꾸어 별도의 연구계획서를 작성하게 되었다.

 

11월까지만 해도 일제강점기 연구와 SCV를 함께 구현할 수 있을 것이라고 생각했지만,

12월이 되어서야 그것이 현실적으로 어렵다는 사실을 인정하게 되었다.

 

결국 2025년 12월 연구계획서는 SCV를 제외한 상태에서 먼저 작성되었고,

당시에는 머신러닝에 대한 이해도 충분하지 않았기 때문에

ML 관련 부분은 AI의 도움을 많이 받았다.

 

그래서 -지난 글에서도 이야기했듯-

머신러닝과 기대효과 부분만큼은 지금 봐도 AI 특유의 표현이 가장 많이 남아 있다.

 

이후 2026년 5월 원전 정책 담론 연구를 수행하면서

데이터베이스 구조와 분석 파이프라인을 실제로 구현하게 되었고,

그 과정에서 기존 연구계획서를 다시 검토하게 되었다.

 

원전 연구를 통해 실제 구현 경험을 쌓은 뒤,

이 경험을 다시 북한 담론 연구계획서에 반영한 것이다.

 

예측이나 인과를 전제하는 표현을 줄이고,

담론과 사건 사이의 시간적 공변 구조(time-dependent covariation)를 분석하는 방향으로

연구 질문을 다시 정리하였다.

 

또한 통계적 결과와 역사적 해석을 보다 명확하게 구분하기 위해,

그동안 구상만 해 두었던 SCV 개념을 연구 설계 안으로 포함하였다.

 

즉, 2026년 5월 수정본은 새로운 연구계획서라기보다,

동일한 연구 질문을 실제 구현 경험과 그동안의 고민을 반영하여 다시 설계한 버전이라고 이해하면 된다.

아래 표는 두 버전의 핵심 차이를 간략하게 정리한 것이다.

 

항목 기존 연구 (2025.12) 수정본 (2026.05)
한 줄 개요 로동신문 담론 변화가 북한의 군사행동을 예측·설명할 수 있다는 전제 하에 TIS·JI·II·DCPD 지표를 설계하고 머신러닝으로 검증하려 한 연구계획 동일한 코퍼스와 데이터베이스 구조를 유지하되, 예측이 아닌 시간적 연관 구조 분석으로 연구 질문을 재정의하고 SCV를 통해 통계 결과와 해석을 분리
목적의식 북한 담론 연구 자체가 중심 북한 자료를 일제강점기 담론 연구를 위한 분석 구조 검증 단계로 명확히 위치시킴
핵심 산출물 위협지수 기반 예측 프레임 재현 가능한 분석 파이프라인과 SCV 기반 해석 체계
연구 질문 담론 변화가 군사행동을 예측할 수 있는가 담론 구조와 사건 구조가 시간적으로 어떤 방식으로 함께 변화하는가
해석 방식 통계 결과를 정치적 의미로 직접 연결 통계적 결과와 역사적 해석을 분리
머신러닝 활용 모델 자체에 많은 의미 부여 모델은 분석 도구로만 사용
연구 범위 결과 중심 분석 구조와 재현성 중심

2026년 5월 수정본 일부 (SCV 설명 삭제 ver.)

 

Temporal Association Structures in North Korean Official Discourse

An NLP-Based Analysis of Rodong Sinmun Across Governance Intervals (2012–2025)

Research Proposal

Computational Social Science · Korean NLP · Political Communication · Digital History


Research Question

How do linguistic and topical features of North Korea’s official state discourse — as documented in Rodong Sinmun — vary across governance intervals, and what temporal association structures emerge between computationally derived discourse indicators and an independently reconstructed chronology of documented events?

The project asks how this documentary record becomes temporally associated with an independently reconstructed governance chronology across different intervals. The emphasis is on observable alignment patterns within the archive itself, rather than on predicting regime behavior or inferring political intent.


Methodological Contribution (Core Contribution Statement)

This project proposes a reproducible pipeline for modeling temporal discourse-event alignment using structured NLP and relational database architecture.


Literature Review

Previous studies on North Korean official statements have relied primarily on qualitative discourse analysis, attending to rhetorical patterns and the semantic content of specific formulations. While these studies have produced important interpretive insights, they remain difficult to compare systematically across time due to heterogeneous source selection and analytic frameworks.

Computational approaches have begun to establish the potential of quantitative analysis in this domain. O et al. (2020) developed a text-mining framework for governing discourse across policy sectors in the Kim Jong-un era, providing a methodological precedent for corpus-based structural analysis. Whang, Lammbrau, and Joo (2018) demonstrated that temporally structured patterns in North Korean datasets are computationally identifiable. Junianse and Jatmika (2024) applied critical discourse analysis to Kim Jong-un’s 2022 speeches in a foreign policy context.

What remains underdeveloped is a systematic, longitudinal extraction of discourse salience indicators aligned against an independently constructed governance event chronology. In this study, discourse–event relationships are treated as descriptive empirical structures of documentary alignment, rather than as predictive or inferential claims.


Research Rationale

A key empirical motivation for this project is the post-2023 lexical shift associated with Kim Jong-un’s “two hostile states” doctrine. In particular, unification-related vocabulary exhibits a measurable decline within Rodong Sinmun over time. This shift is observable as a longitudinal textual pattern independent of interpretive political inference.

More broadly, North Korean official discourse exhibits structured compositional regularities across governance intervals, including escalatory framing, justificatory structures, and shifts in topic salience relative to documented events. The central methodological claim is that such patterns can be formalized as structured, measurable indicators rather than purely interpretive categories.

By transforming discourse into structured variables, this project enables reproducible cross-interval comparison. The output is descriptive rather than inferential, emphasizing stability, comparability, and methodological transparency.


Research Objectives

This study aims to:

  • Construct a longitudinal Rodong Sinmun corpus covering 2012–2025, including an original dataset for post-2020 materials not publicly available in existing databases
  • Develop a structured NLP pipeline to extract discourse salience indicators (topic distributions, lexical salience, polarity measures) at monthly resolution
  • Model temporal alignment between discourse indicators and an independently reconstructed governance event chronology
  • Document post-2023 lexical shifts as a measurable longitudinal textual phenomenon
  • Enable reproducible storage of discourse–event alignment structures in a relational database system
  • Produce a scalable analytical framework applicable to other historical or political corpora

Research Design Overview

The dataset is divided into two independent structures:

  1. Discourse Corpus (Rodong Sinmun, 2012–2025)
  2. Governance Event Chronology (externally constructed)

These are constructed independently and aligned only in later analytical stages to avoid circularity.


Database Construction

The corpus is assembled in two phases:

  • Phase 1 (2012–2019): digitized corpora from KINU and research institution archives
  • Phase 2 (2020–2025): independently constructed dataset based on ICNK archival materials

Consultation with KINU research staff confirmed that no publicly available structured database exists for post-2020 materials. Therefore, original corpus construction is a necessary methodological contribution of this study.

The governance event chronology is independently constructed using:

  • KINU monographs
  • UN Panel of Experts documentation
  • South Korean government publications
  • Academic governance histories

Events are included only when independently verifiable across multiple source categories.


Key Reference

O, G-seop et al. (2020). Governing Discourse and Policy Change for Each Sector in the Kim Jong Un Regime: Analysis of Discourses and Speeches with the Use of Text Mining (KINU Research Monograph No. 20-26). Institute for National Unification.


Methodology

The dataset at this stage is insufficient for deep learning approaches. Therefore, the analytical framework is built on conventional machine learning and statistical NLP methods, with extensibility toward embedding-based models in later stages.


Analytical Pipeline

SQL → Python → NLP Feature Extraction → Temporal Alignment Analysis


SQL Schema (Core Structure)

 
CREATE TABLE documents (
    doc_id      INTEGER PRIMARY KEY,
    pub_date    DATE NOT NULL,
    source      VARCHAR(50),
    title       TEXT,
    raw_content TEXT,
    doc_type    VARCHAR(20)
);

CREATE TABLE keyword_metrics (
    metric_id   INTEGER PRIMARY KEY,
    doc_id      INTEGER,
    keyword     VARCHAR(100),
    frequency   INTEGER DEFAULT 0,
    tfidf_score NUMERIC(10,8),
    year        INTEGER,
    FOREIGN KEY (doc_id) REFERENCES documents(doc_id)
);

CREATE TABLE discourse_features (
    feature_id           INTEGER PRIMARY KEY,
    doc_id               INTEGER,
    external_ref_score   NUMERIC(5,4),
    urgency_marker_score NUMERIC(5,4),
    justif_frame_score   NUMERIC(5,4),
    conciliatory_score   NUMERIC(5,4),
    sentiment_polarity   NUMERIC(5,4),
    is_change_point      BOOLEAN,
    model_name           VARCHAR(100),
    FOREIGN KEY (doc_id) REFERENCES documents(doc_id)
);

CREATE TABLE governance_events (
    event_id    INTEGER PRIMARY KEY,
    event_date  DATE,
    event_type  VARCHAR(100),
    description TEXT,
    source_refs TEXT
);

CREATE TABLE association_results (
    result_id      INTEGER PRIMARY KEY,
    interval_label VARCHAR(20),
    feature_col    VARCHAR(100),
    lag_window     VARCHAR(20),
    correlation    NUMERIC(8,6),
    p_value        NUMERIC(8,6),
    model_config   TEXT
);
 

 

Discourse Feature Indices

This study defines four measurable discourse indicators derived from textual structure:

1. External Actor Reference Salience (EARS)

Measures frequency-weighted mentions of external actors (e.g., US, South Korea, Japan), adjusted by co-occurring framing context.

Method: TF-IDF weighting of named entity mentions; adversarial/cooperative context separation.


2. Justificatory Framing Score (JFS)

Measures presence of justificatory structures (sovereignty, self-defense, historical mission framing).

Method: topic modeling (LDA/NMF) + embedding-based classification.


3. Conciliatory Vocabulary Salience (CVS)

Measures presence of dialogue/negotiation vocabulary over time.

Method: topic weight decay + sentiment polarity trends.


4. Discourse Change Point Detection (DCPD)

Detects structural shifts in discourse distribution over time.

Method: Bayesian Change Point Detection + Kernel CPD.


Temporal Association Analysis

The core analytical procedure estimates lagged cross-correlation between:

  • governance event intensity series
  • discourse salience indicators

Permutation testing is used to construct a null distribution.

This evaluates whether observed alignment is statistically distinguishable from chance.

It does not establish causal direction, institutional intent, or behavioral inference.


Post-Estimation Classification

❌ SCV framework — omitted

  • State A: -omitted-
  • State B: -omitted-
  • State C: -omitted-

(Original SCV classification framework removed from main text for clarity; retained in internal appendix version only.)


Python Pipeline (Summary)

SQLAlchemy is used for structured record-level operations, while Pandas is used for batch-level analysis and aggregation.

TF-IDF-based vectorization is used as baseline representation, with extensibility toward KoBERT/mBERT embedding integration.


TF-IDF + Random Forest (Exploratory Stage)

Random Forest is used as a lexical contribution analysis tool within the corpus representation space.

Target variables are discourse-derived scores, not external event labels.

This stage is exploratory and does not represent predictive modeling of political events.


Research Limitations

  • Korean morphological preprocessing (KoNLPy/Mecab) not yet implemented
  • Embedding-based models not yet integrated
  • Change point detection not fully operational
  • Post-2020 corpus requires archival construction (ICNK)
  • Rodong Sinmun represents only a subset of official discourse

Improvement Directions

  1. Korean morphological preprocessing integration
  2. Embedding-based NLP (KoBERT / mBERT)
  3. Topic modeling (LDA / NMF)
  4. Change point detection pipeline integration
  5. Schema versioning and reproducibility upgrades

Expected Outcomes

  • A structured, reproducible NLP pipeline for longitudinal discourse analysis
  • A relational database of Rodong Sinmun (2012–2025)
  • Temporal alignment structures between discourse and governance events
  • A documented post-2023 lexical shift as a measurable structural phenomenon
  • A reusable computational framework for historical and political text analysis

Intended Contribution

This study contributes to digital history and computational social science by:

  • formalizing discourse as structured temporal data
  • separating event chronology from textual corpus construction
  • enabling reproducible discourse–event alignment analysis
  • providing a scalable relational database + NLP pipeline architecture

Publication Targets

  • Journal of Quantitative Description: Digital Media
  • Journal of East Asian Studies
  • International Journal of Korean History
  • Computational humanities / NLP workshop proceedings

Timeline

  • 2025–2026: technical training + schema design
  • 2026–2027: ICNK corpus construction
  • 2027–2028: NLP + temporal modeling
  • 2028: interpretation + publication drafting

References

정성윤. (2025, December 12). 북핵 정세 평가와 2026년 전망 (Online Series CO 25-24). Korea Institute for National Unification.

Barannikova, A. (2025). Nuclear Strategy of the DPRK in Modern Conditions. Journal for Peace and Nuclear Disarmament, 8(1), 112–129. https://doi.org/10.1080/25751654.2025.2475569

박희진. (2024, June). 북한은 왜 적대적 두 국가를 선언했는가. 황해문화, 36–54.

Scott, W., Genz, B., Elmasry, S., & Adewole, S. (2024). Words of War: Exploring the Presidential Rhetorical Arsenal with Deep Learning. arXiv. https://doi.org/10.48550/arXiv.2412.08868

Junianse, F. M., & Jatmika, S. (2024). Critical Discourse Analysis of North Korea’s Foreign Policy in Kim Jong Un’s Speech in 2022. Multidisciplinary Reviews, 7(8), 2024177. https://doi.org/10.31893/multirev.2024177

Park, J., & Park, J. (2022). North Korea’s Nuclear Use Scenario. Journal of Asian Military Studies, 5(2), 135–158. https://doi.org/10.37944/jams.v5i2.152

이현지, 이화준. (2021). UN 대북제재에 대한 북한의 위기관리전략 연구. 아세아연구, 64(3), 31–68.

Dalton, T. (2020). From Deterrence to Cooperative Security on the Korean Peninsula. Journal for Peace and Nuclear Disarmament, 3(1), 144–156.

O, G-seop et al. (2020). Governing Discourse and Policy Change for Each Sector in the Kim Jong Un Regime: Analysis of Discourses and Speeches with the Use of Text Mining (KINU Research Monograph No. 20-26). Institute for National Unification.

Whang, T., Lammbrau, M., & Joo, H. (2018). Detecting Patterns in North Korean Military Provocations: What Machine-Learning Tells Us. International Relations of the Asia-Pacific, 18(2), 193–220. https://doi.org/10.1093/irap/lcw016

하태경, & 김익환. (2011). 북한의 대남정책결정 메커니즘 분석. 사단법인 열린북한.

Shin, B. (2025). History as Strategy: A Text-Mining Study of Russian Diplomatic Discourse on Ukraine, 2004–2022. 슬라브학보, 40(3), 151–184.


추신.

돌이켜 보면 글 정리 방식 등도 12월보다 나아졌다.

 

 

관련글 더보기