The increasing number of cases processed in Indonesian courts has led to a rapid growth in court decision documents, which contain crucial legal information. However, due to their unstructured textual nature, diverse classifications, linguistic variations, and inconsistent document structures, extracting meaningful information from these documents remains a significant challenge. This study presents a comparative analysis of machine learning approaches for information extraction (IE) from Indonesian court decisions in criminal tribunal, employing Conditional Random Fields (CRF), Support Vector Machines (SVM), Bidirectional Long Short-Term Memory (Bi-LSTM), and Bidirectional Encoder Representations from Transformers (BERT). Experimental results demonstrate that CRF outperforms SVM in terms of F1-score (0.65 vs. 0.19), indicating its relative robustness for structured prediction tasks. Meanwhile, Bi-LSTM achieves an accuracy of 0.37, reflecting limitations in handling the linguistic complexity of legal texts. Notably, BERT significantly surpasses all other methods, achieving an outstanding accuracy of 0.96. The superior performance of BERT is attributed to its deep contextualized representation and ability to leverage pre-trained knowledge, making it highly effective for handling domain-specific variability in legal documents. These findings highlight the potential of utilizing BERT-based models for automated legal information extraction to support the development of intelligent legal systems and the independence of judiciary in Indonesia. © (2025), (International Information and Engineering Technology Association). All rights reserved.