AI Tools for Legal Entity Extraction in India

Legal texts from Indian courts are important for the justice system. They help keep the social and political order. However, with many court cases pending, there is a need for tools that can automate some legal processes using artificial intelligence. This article discusses a method to extract important names and terms from court case judgments. We test different models on a special dataset of legal texts.

#The Legal Entity Extraction Task

The goal of this task is to create a tool that can find Named Entities in Indian legal texts. Most Indian legal documents, like court judgments, are written in English but follow a unique format. This makes it hard to read them using simple techniques, like regular expressions. The specific names we need to find don't match with existing models, which often do not work well for this type of text.

#Advances in Natural Language Processing (NLP)

In the past decade, natural language processing has made great progress. Earlier models struggled to understand the meaning of sentences, but today's models can classify text and create sentences with minimal context. Many newer language models are trained on general text, but they can be fine-tuned for specific areas, like science. These approaches have provided impressive results in many tasks, including named entity recognition (NER), dependency parsing, and relation classification.

#Proposed Method

We propose a method using a deep learning model trained on a labeled legal dataset for named entity recognition. Our model consists of a Bi-LSTM layer for understanding the meaning of tokens and a CRf layer for labeling the sequence of words. To incorporate surrounding context, we utilize specific embeddings that have shown effective results in similar tasks. We have also put together the dataset in a standard format and made it available to others.

#Observations from Experiments

From our experiments, we make several observations:

• Contextual string embeddings help improve the accuracy of identifying custom named entities.
• The Bi-LSTM layer processes the context of words both forwards and backwards, creating a detailed representation for each token.
• The CRF layer uses probabilities from the tokens to find the best sequence of labels.

#Importance of Named Entity Recognition

Named entity recognition is a key part of natural language tasks. It is used in question answering, information retrieval, and co-reference resolution. Identifying names also helps with word sense disambiguation and summarization tasks.

#Legal NER in Research

Legal named entity recognition (NER) has gained attention in the research community. Some researchers focus on identifying entities in US legal texts and categorizing them into classes like judges, attorneys, and courts. In the Indian legal system, some studies aim to structure court judgments into clear parts to improve summarization and prediction tasks. Others have proposed models that use citation networks from legal documents to enhance learning.

#Our Approach

Building on previous research, we use both pre-trained models and contextual string embeddings to create our Bi-LSTM CRF model. We aim to match or exceed the best results that have been previously achieved in legal NER tasks.

#Model Setup

We introduce a deep learning architecture that uses contextual string embeddings for legal named entity recognition. Rather than using a traditional language model, we work with a character-based model. This allows us to learn different meanings for the same word based on context.

#Problem Statement

To begin, we need to identify named entities from a list of tokens. We have a set of predefined entity types, such as COURT, PETITIONER, and RESPONDENT. Using our dataset in a specific format, we are training a model to minimize the loss while identifying these names.

#Data Preparation

Our dataset comprises 11,970 samples from court judgments, with each sample labeled for named entities. We ensure that the classes are evenly distributed to avoid issues with imbalanced data. Each sample is converted to a specific format, and unnecessary words are removed from the sentences.

#Understanding the Model Architecture

Bi-LSTM networks are a type of recurrent neural network that can learn long-term patterns in data. The structure of the model uses two LSTMs stacked together to learn information from both directions in a text. The output is then sent to the CRF layer to determine the best labels for the tokens.

#Experimental Setup and Findings

We train our model on a computer equipped with 16 GB of RAM and a 4-core CPU. The training details include various parameters we have set intentionally to help the model learn the best way to identify named entities.

#Class Distribution for Training

We look at how the classes of named entities are distributed in both the training and validation datasets. Our aim is to maintain a similar distribution of classes in all sets to ensure the model learns effectively.

#Using Different Embeddings

In our research, we experiment with combining various types of embeddings. By adding classic embeddings, like GloVe, we hope to gather deeper meanings connected to word usage.

#Results and Performance Metrics

The dataset includes 9,896 labeled training samples. We split the data into training, validation, and test sets, monitoring the F1 Scores to measure performance. After fine-tuning the model on the training data, we obtain various F1 scores and accuracy metrics across the validation dataset.

#Evaluation of Results

Our experiments indicate that the model achieves a 72% F1 score in identifying named entities in legal texts. This shows that our approach of using contextual string embeddings is beneficial for solving these tasks.

#Conclusion

In summary, we developed a model to identify named entities in legal documents, specifically for Indian court judgments. We used a two-layered Bi-LSTM structure alongside a CRF layer to determine the most suitable label sequence for each token. By including contextual string embeddings, our model can understand words better, even when they have multiple meanings. The result is a system that achieves an F1 score of around 75% in legal NER tasks. This work is a crucial step in applying natural language techniques within legal fields, potentially leading to broader use of these advancements in various legal applications.