Legal Text Recognition Using LSTM-CRF Deep Learning Model.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Legal Informatics

Background:

• Named Entity Recognition (NER) is crucial for analyzing legal documents.
• Deep learning models offer advanced capabilities for NER in specialized domains.
• Existing NER models may face challenges with the complexity of legal text structures.

Purpose of the Study:

• To establish and evaluate a Bidirectional (Bi)-Long Short-Term Memory (LSTM)-Conditional Random Field (CRF) model for NER in legal texts.
• To compare the effectiveness of different annotation methods and segmentation techniques (word vs. character) on NER performance.
• To analyze the impact of various objective loss functions on the Bi-LSTM-CRF model's accuracy.

Main Methods:

• Development of a Bi-LSTM-CRF deep learning architecture tailored for legal NER.
• Comparative analysis of entity recognition using word sequence labeling versus character sequence labeling.
• Evaluation of different annotation strategies and objective loss functions (e.g., log-likelihood, maximum interval criterion).

Main Results:

• The Bi-LSTM-CRF model achieved an F1 score of 88.13% on named entities using word sequence labeling.
• Word segmentation outperformed character segmentation for recognizing place names (67.60% F1) and organization names (89.45% F1).
• Log-likelihood parameter learning yielded superior results compared to the maximum interval criterion.

Conclusions:

• The Bi-LSTM-CRF model, particularly with word segmentation, is highly effective for recognizing extended entities in legal texts.
• This research provides valuable insights and a robust methodology for advancing NER in the legal domain.
• The findings support the application of advanced deep learning techniques for automated legal text analysis.