Exploring Efficient Truncation Strategies for Case Dismissal Prediction using Transformer Models

Recent advancements in Large Language Models (LLMs), Language Models (LMs) and the access to a wellstructured annotated dataset have improved the performance of legal case outcome prediction models. The proposed model utilizes the PredEx dataset, one of the highly annotated expert datasets and extends the prior work by applying the leveraged RoBERTa-Large model with a significant data preprocessing strategy. While implementing the transformer, instead of using the full case text with 512 token chunks as input, the proposed method uses $50 \%$ of the input and truncates it into 512 token chunks, still achieving a test accuracy of $78 \%$, thereby highlighting the effectiveness of the proposed methodology. By utilizing only $50 \%$ of the documents, the proposed system concludes that the full document context is not required for case dismissal classification. Beyond contributing a new perspective for efficient input handling, this method also reduces computational time and memory usage without compromising model performance. The proposed system also implements various traditional machine learning models and establishes a new benchmark baseline result.