Enhanced Hybrid Deep Learning Model With Improved Self-Attention Mechanism for Legal Judgment Prediction

The dawn of Natural Language Processing in the legal research field has achieved great heights. Legal Judgment Prediction (LJP) is one of the important tasks in legal intelligence that can assist attorneys and litigators in predicting judgments. Although current research techniques work effectively, they still have many drawbacks. Firstly, encoding lengthy case facts into vectors without losing information is one of the challenging tasks in LJP. Choosing an encoder that captures the syntax, semantics, and contexts of words can be extremely important for all downstream natural language tasks. Secondly, the features on which the deep learning model is trained also play an important role in testing the real-time cases. This research focuses on Indian cases that follow the common law, unlike civil law. Most of the existing LJP considers only civil law and does not emphasize the extraction of textual features. In this research a novel LJP approach has been suggested to overcome the issues raised above by improving the encoding part using the hybrid embedding method ELMo (Embeddings from Language Model) with Improved Principal Component Analysis (IPCA). Training a crucial set of features is done with a hybrid model, with a combination of Bi-GRU along with a modified attention mechanism and a deep-max-out network. The proposed Hybrid Deep Learning Model with Score Level Fusion (HDLMSF) is experimented with real-time Madras High Court criminal cases and compared with baseline classifier models. The results show that the proposed HDLMSF model has better prediction accuracy, 94.16% than other baseline classifiers.