๐Ÿงพ Combining topic modelling and citation network analysis to study case law from the European Court on Human Rights on the right to respect for private and family life(์‚ฌ์ƒํ™œ ๋ฐ ๊ฐ€์ •์ƒํ™œ ์กด์ค‘๊ถŒ์— ๊ด€ํ•œ ์œ ๋Ÿฝ์ธ๊ถŒ์žฌํŒ์†Œ ํŒ๋ก€ ๋ถ„์„์„ ์œ„ํ•œ ํ† ํ”ฝ ๋ชจ๋ธ๋ง๊ณผ ์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„์˜ ํ†ตํ•ฉ ์—ฐ๊ตฌ)_2024_Cornell University_M. Mohammadi, L. M. Bruijn, M. Wieling, M. Vols

[Original Paper(๋…ผ๋ฌธ ์›๋ณธ)]

https://arxiv.org/abs/2401.16429

[Paper Summary(๋…ผ๋ฌธ ์š”์•ฝ)]

This study demonstrates that integrating LDA-based topic modelling with citation network analysis provides a powerful and efficient method for organizing and retrieving ECtHR Article 8 case law. The combined approach outperforms either method used alone and is particularly effective in identifying eviction-related cases within a large and complex body of judicial decisions.

(๋ณธ ์—ฐ๊ตฌ๋Š” LDA ๊ธฐ๋ฐ˜ ํ† ํ”ฝ ๋ชจ๋ธ๋ง๊ณผ ์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„์„ ํ†ตํ•ฉํ•  ๊ฒฝ์šฐ, ์œ ๋Ÿฝ์ธ๊ถŒ์žฌํŒ์†Œ(ECtHR) ์ œ8์กฐ ํŒ๋ก€๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ  ํšจ์œจ์ ์œผ๋กœ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ ๋ฐฉ๋ฒ•๋ก ์ด ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์คŒ. ๋˜ํ•œ ์ด ๊ฒฐํ•ฉ์  ์ ‘๊ทผ๋ฒ•์€ ๊ฐ๊ฐ์˜ ๋ฐฉ๋ฒ•์„ ๋‹จ๋…์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์œผ๋ฉฐ, ํŠนํžˆ ๋ฐฉ๋Œ€ํ•˜๊ณ  ๋ณต์žกํ•œ ํŒ๋ก€ ์ง‘ํ•ฉ ๋‚ด์—์„œ ๊ฐ•์ œํ‡ด๊ฑฐ(eviction) ๊ด€๋ จ ์‚ฌ๊ฑด์„ ์‹๋ณ„ํ•˜๋Š” ๋ฐ ๋งค์šฐ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚จ)

Key Concepts(์ฃผ์š” ๊ฐœ๋…)

โ–  Topic Modelling(ํ† ํ”ฝ ๋ชจ๋ธ๋ง)

Topic modelling is a Natural Language Processing (NLP) technique used to identify hidden themes within large collections of texts. In this study, the authors used Latent Dirichlet Allocation (LDA) to automatically group ECtHR cases according to their thematic content.

(ํ† ํ”ฝ ๋ชจ๋ธ๋ง์€ ๋Œ€๊ทœ๋ชจ ํ…์ŠคํŠธ ์ž๋ฃŒ์—์„œ ์ˆจ๊ฒจ์ง„ ์ฃผ์ œ๋‚˜ ํŒจํ„ด์„ ์ž๋™์œผ๋กœ ์ถ”์ถœํ•˜๋Š” ์ž์—ฐ์–ด์ฒ˜๋ฆฌ(NLP) ๊ธฐ๋ฒ•์ž„. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ž ์žฌ ๋””๋ฆฌํด๋ ˆ ํ• ๋‹น(Latent Dirichlet Allocation, LDA) ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ์œ ๋Ÿฝ์ธ๊ถŒ์žฌํŒ์†Œ(ECtHR) ํŒ๋ก€๋ฅผ ์ฃผ์ œ๋ณ„๋กœ ์ž๋™ ๋ถ„๋ฅ˜ํ•˜์˜€๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ฐฉ๋Œ€ํ•œ ํŒ๋ก€ ์ง‘ํ•ฉ ์†์—์„œ ์ฃผ์š” ๋ฒ•์  ์Ÿ์ ๊ณผ ์ฃผ์ œ๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ์Œ)

โ–  Citation Network Analysis(์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„)

Citation network analysis examines the citation relationships among court decisions. Cases that frequently cite one another are likely to share similar legal issues or doctrinal foundations. The study used the Louvain community detection algorithm to identify clusters of interconnected cases.

(์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„์€ ํŒ๋ก€๋“ค ์‚ฌ์ด์˜ ์ธ์šฉ ๊ด€๊ณ„๋ฅผ ๋„คํŠธ์›Œํฌ ํ˜•ํƒœ๋กœ ๋ถ„์„ํ•˜๋Š” ๋ฐฉ๋ฒ•์ž„. ์„œ๋กœ ์ž์ฃผ ์ธ์šฉํ•˜๋Š” ํŒ๋ก€๋“ค์€ ์œ ์‚ฌํ•œ ๋ฒ•์  ์Ÿ์ ์ด๋‚˜ ๋ฒ•๋ฆฌ์  ๊ธฐ๋ฐ˜์„ ๊ณต์œ ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Œ. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” Louvain ์ปค๋ฎค๋‹ˆํ‹ฐ ํƒ์ง€ ์•Œ๊ณ ๋ฆฌ์ฆ˜(Louvain Community Detection Algorithm) ์„ ํ™œ์šฉํ•˜์—ฌ ์ธ์šฉ ๊ด€๊ณ„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์„œ๋กœ ๋ฐ€์ ‘ํ•˜๊ฒŒ ์—ฐ๊ฒฐ๋œ ํŒ๋ก€ ์ง‘๋‹จ(์ปค๋ฎค๋‹ˆํ‹ฐ)์„ ์‹๋ณ„ํ•จ.)

โ–  Hybrid (Textโ€“Citation) Approach(ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ ‘๊ทผ๋ฒ•)

The study combines textual information from topic modelling with citation relationships among cases. This integrated approach aims to improve the retrieval and organization of case law by considering both what cases discuss and how they are connected through citations.

(๋ณธ ์—ฐ๊ตฌ๋Š” ํ† ํ”ฝ ๋ชจ๋ธ๋ง์„ ํ†ตํ•ด ๋„์ถœ๋œ ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์ •๋ณด์™€ ํŒ๋ก€ ๊ฐ„ ์ธ์šฉ ๊ด€๊ณ„๋ฅผ ๊ฒฐํ•ฉํ•œ ํ†ตํ•ฉ์  ์ ‘๊ทผ๋ฒ•์„ ์ œ์•ˆํ•จ. ์ด ๋ฐฉ๋ฒ•์€ ํŒ๋ก€๊ฐ€ ๋‹ค๋ฃจ๋Š” ์ฃผ์ œ์  ๋‚ด์šฉ๊ณผ ํŒ๋ก€ ๊ฐ„์˜ ์—ฐ๊ฒฐ ๊ตฌ์กฐ๋ฅผ ๋™์‹œ์— ๊ณ ๋ คํ•จ์œผ๋กœ์จ, ๊ด€๋ จ ํŒ๋ก€๋ฅผ ๋ณด๋‹ค ์ •ํ™•ํ•˜๊ฒŒ ํƒ์ƒ‰ํ•˜๊ณ  ์ฒด๊ณ„์ ์œผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•จ. ์ฆ‰, ํ…์ŠคํŠธ ๋ถ„์„๊ณผ ๋„คํŠธ์›Œํฌ ๋ถ„์„์˜ ์žฅ์ ์„ ๊ฒฐํ•ฉํ•˜์—ฌ ํŒ๋ก€ ๊ฒ€์ƒ‰ ๋ฐ ์กฐ์งํ™”์˜ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•จ)

1. Research Objective(์—ฐ๊ตฌ ๋ชฉ์ )

The study seeks to develop and evaluate a computational method for organizing and retrieving large-scale legal case law from the European Court of Human Rights (ECtHR), specifically concerning Article 8 of the European Convention on Human Rights (ECHR), which protects the right to respect for private and family life, home, and correspondence.

(๋ณธ ์—ฐ๊ตฌ๋Š” ์œ ๋Ÿฝ์ธ๊ถŒํ˜‘์•ฝ(ECHR) ์ œ8์กฐ์™€ ๊ด€๋ จ๋œ ์œ ๋Ÿฝ์ธ๊ถŒ์žฌํŒ์†Œ(ECtHR)์˜ ๋Œ€๊ทœ๋ชจ ํŒ๋ก€๋ฅผ ๋ณด๋‹ค ํšจ๊ณผ์ ์œผ๋กœ ์กฐ์งํ™”ํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ„์‚ฐ์ (computational) ๋ถ„์„ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉ์ ์œผ๋กœ ํ•จ. ์ œ8์กฐ๋Š” ๊ฐœ์ธ์˜ ์‚ฌ์ƒํ™œ, ๊ฐ€์ •์ƒํ™œ, ์ฃผ๊ฑฐ ๋ฐ ํ†ต์‹ ์˜ ์กด์ค‘๊ถŒ(right to respect for private and family life, home, and correspondence) ์„ ๋ณด์žฅํ•˜๋Š” ๊ทœ์ •์ž„)

The specific objectives are(๊ตฌ์ฒด์ ์ธ ์—ฐ๊ตฌ ๋ชฉ์ ):

  1. To assess the effectiveness of topic modelling in organizing Article 8 case law.(์ œ8์กฐ ๊ด€๋ จ ํŒ๋ก€๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ๋ถ„๋ฅ˜ํ•˜๋Š” ๋ฐ ์žˆ์–ด ํ† ํ”ฝ ๋ชจ๋ธ๋ง์˜ ํšจ๊ณผ์„ฑ์„ ํ‰๊ฐ€)
  2. To evaluate the usefulness of citation network analysis for identifying related groups of cases.(๊ด€๋ จ ํŒ๋ก€ ์ง‘๋‹จ์„ ์‹๋ณ„ํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„์˜ ์œ ์šฉ์„ฑ์„ ๊ฒ€์ฆ)
  3. To investigate whether combining both methods improves case retrieval and classification.(ํ† ํ”ฝ ๋ชจ๋ธ๋ง๊ณผ ์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„์„ ๊ฒฐํ•ฉํ•  ๊ฒฝ์šฐ, ํŒ๋ก€ ๊ฒ€์ƒ‰ ๋ฐ ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š”์ง€๋ฅผ ๋ถ„์„)
  4. To test the methods using eviction-related cases as a focused legal issue.(๊ฐ•์ œํ‡ด๊ฑฐ(eviction) ๊ด€๋ จ ์‚ฌ๊ฑด์„ ์ฃผ์š” ์‚ฌ๋ก€ ์—ฐ๊ตฌ ๋Œ€์ƒ์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ ์ œ์•ˆ๋œ ๋ฐฉ๋ฒ•๋ก ์˜ ์‹ค์ œ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๊ฒ€์ฆ)

2. Research Methodology(์—ฐ๊ตฌ ๋ฐฉ๋ฒ•)

โ–  Data(์—ฐ๊ตฌ ์ž๋ฃŒ)

The dataset consisted of 9,777 ECtHR Article 8 cases, including(๋ถ„์„ ๋Œ€์ƒ์€ ์œ ๋Ÿฝ์ธ๊ถŒ์žฌํŒ์†Œ(ECtHR)์˜ ์œ ๋Ÿฝ์ธ๊ถŒํ˜‘์•ฝ(ECHR) ์ œ8์กฐ ๊ด€๋ จ ํŒ๋ก€ 9,777๊ฑด์œผ๋กœ ๊ตฌ์„ฑ):

  • 6,854 English-language cases(์˜์–ด ํŒ๋ก€: 6,854๊ฑด)
  • 2,923 French-language cases(ํ”„๋ž‘์Šค์–ด ํŒ๋ก€: 2,923๊ฑด)

To evaluate the methods, the authors manually identified 198 eviction-related cases, which served as a benchmark dataset.(์—ฐ๊ตฌ์ž๋“ค์€ ์ œ์•ˆ๋œ ๋ถ„์„ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ•์ œํ‡ด๊ฑฐ(eviction)์™€ ๊ด€๋ จ๋œ ํŒ๋ก€ 198๊ฑด์„ ์ˆ˜์ž‘์—…์œผ๋กœ ์„ ๋ณ„ํ•˜์˜€๋‹ค. ์ด ๋ฐ์ดํ„ฐ์…‹์€ ์—ฐ๊ตฌ ๋ฐฉ๋ฒ•์˜ ์ •ํ™•์„ฑ๊ณผ ํšจ์œจ์„ฑ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์ค€ ๋ฐ์ดํ„ฐ์…‹(benchmark dataset) ์œผ๋กœ ํ™œ์šฉ)

โ–  Experiment 1(์—ฐ๊ตฌ ๋ฐฉ๋ฒ• 1): Topic Modelling(ํ† ํ”ฝ ๋ชจ๋ธ๋ง)

The authors applied LDA topic modelling to the English-language cases(์—ฐ๊ตฌ์ง„์€ ์œ ๋Ÿฝ์ธ๊ถŒ์žฌํŒ์†Œ(ECtHR)์˜ ์˜์–ด ํŒ๋ก€ 6,854๊ฑด์„ ๋Œ€์ƒ์œผ๋กœ LDA(Latent Dirichlet Allocation) ๊ธฐ๋ฐ˜ ํ† ํ”ฝ ๋ชจ๋ธ๋ง์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ฐฉ๋Œ€ํ•œ ํŒ๋ก€ ๋ฐ์ดํ„ฐ์— ๋‚ด์žฌ๋œ ์ฃผ์š” ์ฃผ์ œ๋“ค์„ ์ž๋™์œผ๋กœ ์ถ”์ถœํ•˜๊ณ  ๋ถ„๋ฅ˜ํ•˜๊ณ ์ž ํ•จ).

Procedure(์—ฐ๊ตฌ ์ ˆ์ฐจ)

  • Text preprocessing (tokenization, stop-word removal, lemmatization)(ํ…์ŠคํŠธ ์ „์ฒ˜๋ฆฌ(ํ† ํฐํ™”, ๋ถˆ์šฉ์–ด ์ œ๊ฑฐ, ํ‘œ์ œ์–ด ์ถ”์ถœ))
  • Construction of a document-term matrix(๋ฌธ์„œ-๋‹จ์–ด ํ–‰๋ ฌ ๊ตฌ์ถ•(์ „์ฒ˜๋ฆฌ๋œ ํ…์ŠคํŠธ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐ ๋ฌธ์„œ์— ํฌํ•จ๋œ ๋‹จ์–ด์˜ ๋นˆ๋„๋ฅผ ์ˆ˜์น˜ํ™”ํ•˜์—ฌ ๋ถ„์„ ๊ฐ€๋Šฅํ•œ ํ˜•ํƒœ๋กœ ๋ณ€ํ™˜))
  • Application of LDA(LDA ์ ์šฉ(LDA ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ™œ์šฉํ•˜์—ฌ ํŒ๋ก€ ๋ฌธ์„œ๋“ค์— ์กด์žฌํ•˜๋Š” ์ž ์žฌ์  ์ฃผ์ œ(Topic)๋ฅผ ์ถ”์ถœ))

Outcome(์—ฐ๊ตฌ ๊ฒฐ๊ณผ)

  • The model identified 17 major topics, including housing and property, family life, healthcare, warfare, religion, criminal justice, and communication surveillance.(ํ† ํ”ฝ ๋ชจ๋ธ๋ง ๊ฒฐ๊ณผ, ์ด 17๊ฐœ์˜ ์ฃผ์š” ์ฃผ์ œ(Topics) ๊ฐ€ ๋„์ถœ, ์ฃผ์š” ์ฃผ์ œ: ์ฃผ๊ฑฐ ๋ฐ ์žฌ์‚ฐ๊ถŒ, ๊ฐ€์กฑ์ƒํ™œ, ์˜๋ฃŒ ๋ฐ ๊ฑด๊ฐ•๊ด€๋ฆฌ, ์ „์Ÿ ๋ฐ ๋ฌด๋ ฅ๋ถ„์Ÿ, ์ข…๊ต, ํ˜•์‚ฌ์‚ฌ๋ฒ•, ํ†ต์‹  ๊ฐ์‹œ ๋ฐ ํ”„๋ผ์ด๋ฒ„์‹œ)

โ–  Experiment 2(์—ฐ๊ตฌ ๋ฐฉ๋ฒ• 2): Citation Network Analysis(์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„)

The authors constructed a citation network where(์—ฐ๊ตฌ์ง„์€ ์œ ๋Ÿฝ์ธ๊ถŒ์žฌํŒ์†Œ(ECtHR) ํŒ๋ก€ ๊ฐ„์˜ ์ธ์šฉ ๊ด€๊ณ„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ธ์šฉ ๋„คํŠธ์›Œํฌ(Citation Network) ๋ฅผ ๊ตฌ์ถ•):

  • Nodes represented cases(๋…ธ๋“œ: ๊ฐœ๋ณ„ ํŒ๋ก€(Case)).
  • Edges represented citation relationships between cases.(์—ฃ์ง€: ํŒ๋ก€ ๊ฐ„ ์ธ์šฉ ๊ด€๊ณ„)

The Louvain algorithm was used to detect communities of cases based on citation patterns.(ํŒ๋ก€ ๊ฐ„ ์ธ์šฉ ํŒจํ„ด์— ๊ธฐ๋ฐ˜ํ•œ ์ปค๋ฎค๋‹ˆํ‹ฐ(์ง‘๋‹จ)๋ฅผ ํƒ์ง€ํ•˜๊ธฐ ์œ„ํ•ด Louvain ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์‚ฌ์šฉ)

The analysis focused on the largest connected component of the network, which contained(๋ถ„์„์€ ๋„คํŠธ์›Œํฌ ๋‚ด์—์„œ ๊ฐ€์žฅ ํฐ ์—ฐ๊ฒฐ ์„ฑ๋ถ„(largest connected component)์„ ์ค‘์‹ฌ์œผ๋กœ ์ˆ˜ํ–‰๋˜์—ˆ์œผ๋ฉฐ, ํ•ด๋‹น ๋„คํŠธ์›Œํฌ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ตฌ์„ฑ):

  • 7,234 cases(7,234๊ฑด์˜ ํŒ๋ก€)
  • 39,501 citation links(39,501๊ฐœ์˜ ์ธ์šฉ ๊ด€๊ณ„)

โ–  Experiment 3(์—ฐ๊ตฌ ๋ฐฉ๋ฒ• 3): Integration of Topic Modelling and Citation Network Analysis(ํ† ํ”ฝ ๋ชจ๋ธ๋ง๊ณผ ์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„์˜ ํ†ตํ•ฉ)

To combine textual and citation information(ํ…์ŠคํŠธ ์ •๋ณด์™€ ์ธ์šฉ ์ •๋ณด๋ฅผ ๊ฒฐํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ ˆ์ฐจ๋ฅผ ์ˆ˜ํ–‰):

  1. Topic vectors generated by LDA were compared using cosine similarity.(LDA๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ๋œ ํ† ํ”ฝ ๋ฒกํ„ฐ(topic vectors) ๋ฅผ ์ฝ”์‚ฌ์ธ ์œ ์‚ฌ๋„(cosine similarity) ๋ฅผ ์ด์šฉํ•˜์—ฌ ๋น„๊ต)
  2. Citation links were weighted according to topic similarity.(ํŒ๋ก€ ๊ฐ„ ํ† ํ”ฝ ์œ ์‚ฌ์„ฑ(topic similarity) ์— ๋”ฐ๋ผ ์ธ์šฉ ๊ด€๊ณ„์— ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌ)
  3. The Louvain algorithm was applied to the weighted network.(๊ฐ€์ค‘์น˜๊ฐ€ ์ ์šฉ๋œ ๋„คํŠธ์›Œํฌ์— Louvain ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ ์šฉ)

This approach strengthened links between topically similar cases and weakened links between unrelated cases, resulting in more coherent legal communities.(์ด๋Ÿฌํ•œ ์ ‘๊ทผ๋ฒ•์€ ์ฃผ์ œ์ ์œผ๋กœ ์œ ์‚ฌํ•œ ํŒ๋ก€๋“ค ๊ฐ„์˜ ์—ฐ๊ฒฐ์„ ๊ฐ•ํ™”ํ•˜๊ณ , ๊ด€๋ จ์„ฑ์ด ๋‚ฎ์€ ํŒ๋ก€๋“ค ๊ฐ„์˜ ์—ฐ๊ฒฐ์„ ์•ฝํ™”์‹œํ‚ด. ๊ทธ ๊ฒฐ๊ณผ, ๋ณด๋‹ค ์‘์ง‘๋ ฅ ์žˆ๊ณ  ์ผ๊ด€์„ฑ ์žˆ๋Š” ๋ฒ•๋ฅ  ์ปค๋ฎค๋‹ˆํ‹ฐ(ํŒ๋ก€ ์ง‘๋‹จ)๋ฅผ ํ˜•์„ฑํ•  ์ˆ˜ ์žˆ์—ˆ์Œ)

3. Key Findings(์ฃผ์š” ์—ฐ๊ตฌ ๊ฒฐ๊ณผ)

โ–  Finding 1(์—ฐ๊ตฌ ๊ฒฐ๊ณผ 1): Topic Modelling Provides a Broad Overview but Misses Specific Issues(ํ† ํ”ฝ ๋ชจ๋ธ๋ง์€ ์ „๋ฐ˜์ ์ธ ์ฃผ์ œ ํŒŒ์•…์—๋Š” ํšจ๊ณผ์ ์ด์ง€๋งŒ, ํŠน์ • ์Ÿ์  ์‹๋ณ„์—๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์Œ)

LDA successfully identified the major themes within Article 8 case law. However, eviction-related cases were scattered across multiple topics.(LDA๋Š” ์ œ8์กฐ ๊ด€๋ จ ํŒ๋ก€์— ํฌํ•จ๋œ ์ฃผ์š” ์ฃผ์ œ๋“ค์„ ์„ฑ๊ณต์ ์œผ๋กœ ์‹๋ณ„ํ•˜์˜€๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ฐ•์ œํ‡ด๊ฑฐ(eviction) ๊ด€๋ จ ์‚ฌ๊ฑด๋“ค์€ ์—ฌ๋Ÿฌ ํ† ํ”ฝ์— ๋ถ„์‚ฐ๋˜์–ด ๋‚˜ํƒ€๋‚จ)

Although many eviction cases appeared in the โ€œHousing and Propertyโ€ topic, others were located in unrelated topics such as โ€œWarfareโ€, particularly cases involving the destruction of homes during the Turkishโ€“Kurdish conflict.(๋งŽ์€ ๊ฐ•์ œํ‡ด๊ฑฐ ์‚ฌ๊ฑด์ด โ€˜์ฃผ๊ฑฐ ๋ฐ ์žฌ์‚ฐโ€™ ํ† ํ”ฝ์— ํฌํ•จ๋˜์—ˆ์ง€๋งŒ, ์ผ๋ถ€ ์‚ฌ๊ฑด์€ โ€˜์ „์Ÿโ€™ ๊ณผ ๊ฐ™์€ ๊ด€๋ จ์„ฑ์ด ๋‚ฎ์•„ ๋ณด์ด๋Š” ํ† ํ”ฝ์— ํฌํ•จ๋˜์—ˆ๋‹ค. ํŠนํžˆ ์ด๋Š” ํ„ฐํ‚ค-์ฟ ๋ฅด๋“œ ๋ถ„์Ÿ ๊ณผ์ •์—์„œ ๋ฐœ์ƒํ•œ ์ฃผํƒ ํŒŒ๊ดด ์‚ฌ๊ฑด๋“ค๊ณผ ๊ด€๋ จ๋œ ์‚ฌ๋ก€๋“ค ๋•Œ๋ฌธ)

As a result, topic modelling alone could not comprehensively identify all eviction-related cases.(๊ทธ ๊ฒฐ๊ณผ, ํ† ํ”ฝ ๋ชจ๋ธ๋ง๋งŒ์œผ๋กœ๋Š” ๋ชจ๋“  ๊ฐ•์ œํ‡ด๊ฑฐ ๊ด€๋ จ ์‚ฌ๊ฑด์„ ํฌ๊ด„์ ์œผ๋กœ ์‹๋ณ„ํ•˜๋Š” ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚จ)

โ–  Finding 2(์—ฐ๊ตฌ ๊ฒฐ๊ณผ 2): Citation Network Analysis Also Has Limitations(์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„ ์—ญ์‹œ ํ•œ๊ณ„๋ฅผ ์ง€๋‹˜)

Citation-based clustering revealed several communities containing eviction cases, but the results depended heavily on the resolution parameter used in the Louvain algorithm.(์ธ์šฉ ๊ด€๊ณ„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๊ตฐ์ง‘ํ™”(citation-based clustering)๋Š” ๊ฐ•์ œํ‡ด๊ฑฐ ์‚ฌ๊ฑด์„ ํฌํ•จํ•˜๋Š” ์—ฌ๋Ÿฌ ์ปค๋ฎค๋‹ˆํ‹ฐ๋ฅผ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทธ ๊ฒฐ๊ณผ๋Š” Louvain ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ์‚ฌ์šฉ๋œ ํ•ด์ƒ๋„ ๋งค๊ฐœ๋ณ€์ˆ˜(resolution parameter) ์— ํฌ๊ฒŒ ์˜์กดํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚จ)

Additionally(๋˜ํ•œ):

  • Some eviction cases were located outside the giant citation network.(์ผ๋ถ€ ๊ฐ•์ œํ‡ด๊ฑฐ ์‚ฌ๊ฑด์€ ๊ฑฐ๋Œ€ ์ธ์šฉ ๋„คํŠธ์›Œํฌ(giant citation network) ๋ฐ–์— ์œ„์น˜ํ•ด ์žˆ์—ˆ์Œ)
  • Relevant cases were distributed across different communities.(๊ด€๋ จ ์‚ฌ๊ฑด๋“ค์ด ์„œ๋กœ ๋‹ค๋ฅธ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ๋ถ„์‚ฐ๋˜์–ด ์žˆ์—ˆ์Œ)

Therefore, citation network analysis alone was insufficient for complete case retrieval.(๋”ฐ๋ผ์„œ ์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„๋งŒ์œผ๋กœ๋Š” ๊ด€๋ จ ํŒ๋ก€๋ฅผ ์™„์ „ํ•˜๊ฒŒ ๊ฒ€์ƒ‰ํ•˜๊ณ  ์‹๋ณ„ํ•˜๋Š” ๋ฐ ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚จ)

โ–  Finding 3(์—ฐ๊ตฌ ๊ฒฐ๊ณผ 3): The Combined Method Produced the Best Results(๊ฒฐํ•ฉ ๋ฐฉ๋ฒ•์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๊ณผ๋ฅผ ๋ณด์ž„)

The integration of topic modelling and citation network analysis generated smaller, more coherent, and topic-focused communities.(ํ† ํ”ฝ ๋ชจ๋ธ๋ง๊ณผ ์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„์„ ํ†ตํ•ฉํ•œ ์ ‘๊ทผ๋ฒ•์€ ๋” ์ž‘๊ณ , ์‘์ง‘๋ ฅ ์žˆ์œผ๋ฉฐ, ํŠน์ • ์ฃผ์ œ์— ์ง‘์ค‘๋œ ์ปค๋ฎค๋‹ˆํ‹ฐ๋ฅผ ์ƒ์„ฑ)

Compared with the individual methods, the hybrid approach(๊ฐœ๋ณ„ ๋ฐฉ๋ฒ•๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ ‘๊ทผ๋ฒ•์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์žฅ์ ์„ ๋ณด์ž„):

  • Improved thematic consistency within communities.(์ปค๋ฎค๋‹ˆํ‹ฐ ๋‚ด ์ฃผ์ œ์  ์ผ๊ด€์„ฑ(thematic consistency) ์„ ํ–ฅ์ƒ์‹œํ‚ด)
  • Reduced the influence of the resolution parameter.(ํ•ด์ƒ๋„ ๋งค๊ฐœ๋ณ€์ˆ˜(resolution parameter) ์˜ ์˜ํ–ฅ์„ ๊ฐ์†Œ์‹œํ‚ด)
  • Increased the concentration of eviction-related cases within identified communities.(์‹๋ณ„๋œ ์ปค๋ฎค๋‹ˆํ‹ฐ ๋‚ด์—์„œ ๊ฐ•์ œํ‡ด๊ฑฐ ๊ด€๋ จ ์‚ฌ๊ฑด์˜ ์ง‘์ค‘๋„๋ฅผ ๋†’์ž„)

โ–  Finding 4(์—ฐ๊ตฌ ๊ฒฐ๊ณผ 4): Discovery of Additional Eviction Cases(์ถ”๊ฐ€์ ์ธ ๊ฐ•์ œํ‡ด๊ฑฐ ๊ด€๋ จ ์‚ฌ๊ฑด์˜ ๋ฐœ๊ฒฌ)

The combined approach successfully identified 83 previously undiscovered eviction-related cases that had not been included in the manually collected dataset.(๊ฒฐํ•ฉ ์ ‘๊ทผ๋ฒ•(combined approach)์€ ์ˆ˜์ž‘์—…์œผ๋กœ ๊ตฌ์ถ•๋œ ๋ฐ์ดํ„ฐ์…‹์— ํฌํ•จ๋˜์ง€ ์•Š์•˜๋˜ 83๊ฑด์˜ ์ƒˆ๋กœ์šด ๊ฐ•์ œํ‡ด๊ฑฐ ๊ด€๋ จ ์‚ฌ๊ฑด์„ ์„ฑ๊ณต์ ์œผ๋กœ ์‹๋ณ„)

Researchers were able to identify over 211 eviction-related cases by examining only 361 cases, demonstrating a substantial improvement in legal information retrieval efficiency.(์—ฐ๊ตฌ์ง„์€ ๋‹จ 361๊ฑด์˜ ํŒ๋ก€๋งŒ ๊ฒ€ํ† ํ•˜์—ฌ 211๊ฑด ์ด์ƒ์˜ ๊ฐ•์ œํ‡ด๊ฑฐ ๊ด€๋ จ ์‚ฌ๊ฑด์„ ์ฐพ์•„๋‚ผ ์ˆ˜ ์žˆ์—ˆ์œผ๋ฉฐ, ์ด๋Š” ๋ฒ•๋ฅ  ์ •๋ณด ๊ฒ€์ƒ‰์˜ ํšจ์œจ์„ฑ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์—ˆ์Œ์„ ๋ณด์—ฌ์คŒ)

4. Conclusions and Implications(๊ฒฐ๋ก  ๋ฐ ์‹œ์‚ฌ์ )

โ–  Conclusion(๊ฒฐ๋ก )

The study concludes that combining topic modelling with citation network analysis is more effective than using either method independently for organizing and retrieving legal case law.(๋ณธ ์—ฐ๊ตฌ๋Š” ํ† ํ”ฝ ๋ชจ๋ธ๋ง๊ณผ ์ธ์šฉ ๋„คํŠธ์›Œํฌ ๋ถ„์„์„ ๊ฒฐํ•ฉํ•˜๋Š” ๊ฒƒ์ด, ๊ฐ ๋ฐฉ๋ฒ•์„ ๋…๋ฆฝ์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋ฒ•๋ฅ  ํŒ๋ก€๋ฅผ ์กฐ์งํ™”ํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ฐ ๋” ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒฐ๋ก ์„ ์ œ์‹œ)

The hybrid approach successfully utilizes both(ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ ‘๊ทผ๋ฒ•์€ ๋‹ค์Œ ๋‘ ๊ฐ€์ง€ ์š”์†Œ๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ํ™œ์šฉ):

  • The textual content of legal decisions(๋ฒ•์› ํŒ๊ฒฐ๋ฌธ์˜ ํ…์ŠคํŠธ ๋‚ด์šฉ),
  • The citation relationships among cases(ํŒ๋ก€ ๊ฐ„ ์ธ์šฉ ๊ด€๊ณ„).

As a result, it produces more meaningful and coherent legal communities and improves the identification of relevant cases.(๊ทธ ๊ฒฐ๊ณผ, ๋ณด๋‹ค ์˜๋ฏธ ์žˆ๊ณ  ์‘์ง‘๋ ฅ ์žˆ๋Š” ๋ฒ•๋ฅ  ์ปค๋ฎค๋‹ˆํ‹ฐ(ํŒ๋ก€ ์ง‘๋‹จ)๋ฅผ ํ˜•์„ฑํ•  ์ˆ˜ ์žˆ์—ˆ์œผ๋ฉฐ, ๊ด€๋ จ ํŒ๋ก€๋ฅผ ๋”์šฑ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚จ)

โ–  Academic Implications(ํ•™๋ฌธ์  ์‹œ์‚ฌ์ )

1. Advancement of Computational Legal Research(๊ณ„์‚ฐ๋ฒ•ํ•™์˜ ๋ฐœ์ „)

The study demonstrates how artificial intelligence and natural language processing can support large-scale legal analysis and legal information retrieval.(์ธ๊ณต์ง€๋Šฅ๊ณผ ์ž์—ฐ์–ด์ฒ˜๋ฆฌ ๊ธฐ์ˆ ์ด ๋Œ€๊ทœ๋ชจ ๋ฒ•๋ฅ  ๋ถ„์„๊ณผ ๋ฒ•๋ฅ  ์ •๋ณด ๊ฒ€์ƒ‰์— ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์คŒ)

2. Interdisciplinary Methodology(ํ•™์ œ๊ฐ„ ์—ฐ๊ตฌ ๋ฐฉ๋ฒ•๋ก  ์ œ์‹œ)

It provides a framework that combines:

  • Legal research(๋ฒ•ํ•™),
  • Natural language processing(์ž์—ฐ์–ด์ฒ˜๋ฆฌ),
  • Network science(๋„คํŠธ์›Œํฌ ๊ณผํ•™).

This interdisciplinary approach contributes to the growing field of computational legal studies.((๋ฒ•ํ•™, ์ž์—ฐ์–ด์ฒ˜๋ฆฌ, ๋„คํŠธ์›Œํฌ ๊ณผํ•™์„ ๊ฒฐํ•ฉํ•œ ์—ฐ๊ตฌ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•˜์—ฌ ๊ณ„์‚ฐ๋ฒ•ํ•™ ๋ถ„์•ผ ๋ฐœ์ „์— ๊ธฐ์—ฌํ•จ))

3. Improved Case Law Retrieval(ํŒ๋ก€ ๊ฒ€์ƒ‰ ๋ฐ ๋ถ„์„์˜ ํ–ฅ์ƒ)

The method enables researchers to(์ด ๋ฐฉ๋ฒ•์€ ์—ฐ๊ตฌ์ž๋“ค์ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ž‘์—…์„ ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›):

  • Identify relevant precedents more efficiently(๊ด€๋ จ ์„ ๋ก€(ํŒ๋ก€)๋ฅผ ๋”์šฑ ํšจ์œจ์ ์œผ๋กœ ์‹๋ณ„),
  • Discover hidden case clusters(๊ธฐ์กด์— ๋ฐœ๊ฒฌ๋˜์ง€ ์•Š์•˜๋˜ ์ˆจ๊ฒจ์ง„ ํŒ๋ก€ ์ง‘๋‹จ(case clusters)์„ ์ฐพ์•„๋‚ผ ์ˆ˜ ์žˆ์Œ),
  • Analyze the development of legal doctrines over time(์‹œ๊ฐ„์˜ ํ๋ฆ„์— ๋”ฐ๋ฅธ ๋ฒ•๋ฆฌ(legal doctrines)์˜ ๋ฐœ์ „ ๊ณผ์ •์„ ๋ถ„์„).

โ–  Practical Implications(ํŒ๋ก€ ๊ฒ€์ƒ‰ ํšจ์œจ์„ฑ ํ–ฅ์ƒ)

For judges, lawyers, and legal researchers, the proposed approach can(๊ด€๋ จ ์„ ๋ก€๋ฅผ ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ์ฐพ๊ณ , ์ˆจ๊ฒจ์ง„ ํŒ๋ก€ ์ง‘๋‹จ์„ ๋ฐœ๊ฒฌํ•˜๋ฉฐ, ๋ฒ•๋ฆฌ์˜ ๋ฐœ์ „ ๊ณผ์ •์„ ๋ถ„์„ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•จ):

  • Reduce the time required to locate relevant cases(๊ด€๋ จ ํŒ๋ก€๋ฅผ ์ฐพ๋Š” ๋ฐ ์†Œ์š”๋˜๋Š” ์‹œ๊ฐ„์„ ๋‹จ์ถ•ํ•  ์ˆ˜ ์žˆ์Œ),
  • Improve the completeness of legal research(๋ฒ•๋ฅ  ์—ฐ๊ตฌ์˜ ์™„์ „์„ฑ๊ณผ ํฌ๊ด„์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ),
  • Identify important cases that may be overlooked through traditional keyword-based searches(์ „ํ†ต์ ์ธ ํ‚ค์›Œ๋“œ ๊ธฐ๋ฐ˜ ๊ฒ€์ƒ‰ ๋ฐฉ์‹์œผ๋กœ๋Š” ๊ฐ„๊ณผ๋  ์ˆ˜ ์žˆ๋Š” ์ค‘์š”ํ•œ ํŒ๋ก€๋ฅผ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ์Œ).

โ–  Limitations and Future Directions(ํ•œ๊ณ„์  ๋ฐ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ)

The authors emphasize that computational methods should complement rather than replace legal expertise. While the hybrid approach significantly improves case retrieval, human judgment remains essential for interpreting legal significance and validating the results.

(์ €์ž๋“ค์€ ๊ณ„์‚ฐ์  ๋ถ„์„ ๋ฐฉ๋ฒ•์ด ๋ฒ•๋ฅ  ์ „๋ฌธ๊ฐ€์˜ ํŒ๋‹จ์„ ๋Œ€์ฒดํ•˜๊ธฐ๋ณด๋‹ค๋Š” ๋ณด์™„ํ•˜๋Š” ์—ญํ• ์„ ํ•ด์•ผ ํ•œ๋‹ค๊ณ  ๊ฐ•์กฐํ•จ, ๋น„๋ก ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ ‘๊ทผ๋ฒ•์ด ํŒ๋ก€ ๊ฒ€์ƒ‰ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์ง€๋งŒ, ๋ฒ•์  ์ค‘์š”์„ฑ์„ ํ•ด์„ํ•˜๊ณ  ๋ถ„์„ ๊ฒฐ๊ณผ์˜ ํƒ€๋‹น์„ฑ์„ ๊ฒ€์ฆํ•˜๋Š” ๊ณผ์ •์—์„œ๋Š” ์—ฌ์ „ํžˆ ์ธ๊ฐ„์˜ ์ „๋ฌธ์  ํŒ๋‹จ์ด ํ•„์ˆ˜์ ์ž„)

๋‹ต๊ธ€ ๋‚จ๊ธฐ๊ธฐ

์ด๋ฉ”์ผ ์ฃผ์†Œ๋Š” ๊ณต๊ฐœ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ•„์ˆ˜ ํ•„๋“œ๋Š” *๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค