Debunking the Myth of Judicial Ghost-Writing 施中右 李韶曼 周炫谷 張凱雄 The Case of the Constitutional Court of Taiwan

Justices of the Constitutional Court perform their work in a collective deliberation process. However….. 小法官看大法官的提名，林臻嫺，2015年05月09日，蘋果日報 這幾年大法官做出的解釋文，愈來愈少人要看......過去偶能 見擊掌叫好佳作，現在縱不同意見書也讓人呵欠連連，既 然解釋文沒有什麼功用，我們又何必在乎是誰做的？...... 大法官們，在關起門的會議室裡，會做出如何理論有 餘而開創不足的釋憲解釋，應都可預見了。

Problems of the Constitutional Court’s Products The Court has its own writing style, probably as a result of the unique collective decision-making process. obscure diction and convoluted syntax overflow too short for an articulate reasoning not transparent enough to be discussed

Problems May Be Worse... Accountability Highly controversial issues v. Unattributed decisions Significant disagreement within the Court Information for future decision-making

The Myth of Judicial Authorship How the interpretations are collectively made? Is it true that the Court has its own writing style? How to identify possible authors of the collective decisions?

What is the deliberation process? Sentence-by Sentence Collective Writing Major Ghost-writers 大法官A 大法官B 大法官C 人多 聲音大 大法官F 大法官E 大法官A+B+C+D

Solving Attributional Questions 社會秩序維護法第八十條第一項第一款就意圖得利與人姦、宿者，處三日以下拘留或新臺幣三萬元以下罰鍰之規定，與憲法第七條之平等原則有違，應自本解釋公布之日起至遲於二年屆滿時，失其效力。 蘇永欽? 葉百修? 陳新民? 大法官會議主席  大法官  賴英照 大法官  謝在全  徐璧湖  許宗力  許玉秀 林錫堯  池啟明  蔡清遊  黃茂榮 陳    敏  葉百修  陳春生  陳新民

Solving Attributional Questions Algorithmically Using Python to retrieve signed separate opinions (training data), unattributed holdings and reasonings of interpretations (test data) on the Judicial Yuan website Preprocessing: word segmentations Creating a list of common function words (adapted from Bei Yu’s list & Judicial Yuan’s set expressions list) Analyzing frequencies and locations of those words in training data Creating virtual examples for text classification from training data Feeding dissected texts (in form of vectors and numeric data, etc.) into machine learning codes to train computers to identify the authorship

Solving Attributional Questions Algorithmically Training set: we take 432 signed opinions from 26 judges serving from 2003-2011, the last two terms of the Court Virtual example: bootstrapping Cross Validation: randomly pick 20% of the data from training for validation Test set: we take 118 Interpretations (including holding and reasoning) ranging from No. 573 to No. 690, made by the last two terms of the Court

Bootstrapping: resampling with replace 100 virtual examples per text 3000 words

Results Insights on Writing Style

Random Forest *Cross Validation: randomly pick 20% for validation

的 0.107907 之 0.087986 其 0.050723 也 0.049300 此 0.039223 不 0.038115 這 0.037961 只 0.037886 在 0.037703 系爭 0.036384 是 0.033684 於 0.026813 了 0.025063 與 0.025012 Feature Importance

Results Authorship Prediction Comparison to Predictions by Domain Experts