本篇之研究係基於多任務學習(Multi-task Learning)架構，而在上述每個工作中之子任務間亦是具有相依關係，換言之，一項子任務之預測結果也影響了其他子任務之預測結果。在本篇論文中，對於子任務之間應用了不同之拓譜架構如IMN(Iteracitve Message Passing Network)及TopJudge與本篇研究提出結合IMN與TopJudge形成之拓譜架構並搭配不同之語言模型，如Word2Vec、BERT、Lawformer 進行LJP任務，比較子任務形成不同之拓譜架構並應用不同之語言模型對於效能之影響。此外，由於大型語言模型中的參數數量巨大，對每個LJP任務進行Full-Tuning的成本將變得越來越昂貴。為解決這個問題，我們採用了LoRA(Low-Rank Adaptation)架構，這是屬於一種Parameter-Efficient Fine-Tuning(PEFT)的技術，以減少訓練參數數量並節省計算成本與模型訓練時間。實驗結果顯示，使用LoRA進行Fine-Tuning不僅降低了訓練時間（45\%），甚至對某些LJP任務帶來了性能提升效果（2.5\%的Macro F1）。;Legal Judgment Prediction (LJP) aims to predict the judgement results (such as article, charge, and penalty) based on the criminal facts of the case. Most previous research in this field was based on criminal statements from court verdicts. However, each verdict actually is based on the content from indictments. For prosecutors, will the case be dismissed or processed? If the case is accepted, is the penalty a jail sentence or a fine? What is the charge and article violated? In this study, we therefore define three novel LJP tasks for prosecutors, including prosecution outcome prediction (LJP\#1), fine prediction (LJP\#2) and imprison prediction (LJP\#3). Due to the huge number of parameters in a large language model, the cost of full-tuning for each LJP task will become increasingly expensive. To solve this problem, we adopt the LoRA (Low-Rank Adaptation) architecture, a technique for parameter-efficient fine-tuning (PEFT) to reduce the number of tuned parameters and save computational cost/time. The experiments show that using LoRA for fine-tuning not only improves not only reduce training time (45\%) but also brings performance improvement effect (2.5\% F1) for some LJP tasks.
