💬 NLP

Towards Intelligent Legal Document Analysis: CNN-Driven Classification of Case Law Texts

This paper proposes a lightweight, high-accuracy CNN-based framework that combines lemmatization, FastText embeddings, and multi-kernel convolution to achieve superior citation-treatment classification in legal documents, outperforming heavyweight transformer models like BERT in both accuracy and inference speed.

Original authors: Moinul Hossain, Sourav Rabi Das, Zikrul Shariar Ayon, Sadia Afrin Promi, Ahnaf Atef Choudhury, Shakila Rahman, Jia Uddin

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Moinul Hossain, Sourav Rabi Das, Zikrul Shariar Ayon, Sadia Afrin Promi, Ahnaf Atef Choudhury, Shakila Rahman, Jia Uddin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a librarian in a massive, ancient library. Every day, thousands of new books (legal court cases) arrive. Each book is written in a very specific, old-fashioned language, full of long, confusing sentences and strange Latin phrases. Your job is to sort these books into different bins based on how they treat other books they mention. Do they praise a previous book? Do they say it's wrong? Do they just mention it casually?

Doing this by hand is exhausting, slow, and prone to mistakes. That's where this paper comes in. The authors built a smart robot librarian that can read these legal books and sort them almost instantly, and they did it without needing a supercomputer the size of a house.

Here is the story of how they built this robot, explained simply:

1. The Problem: The "Heavy" Robot

In the past, people tried to solve this using "Heavy Robots" (like BERT). Think of these as giant, powerful AI models that have read the entire internet. They are incredibly smart and accurate, but they are also:

Heavy: They require massive amounts of electricity and expensive computer chips.
Slow: They take a long time to read a single page.
Clumsy: Sometimes they get confused by the weird, specific legal jargon because they try to understand everything at once.

The authors asked: "Can we build a 'Lightweight Robot' that is just as smart but runs on a regular laptop and works at the speed of light?"

2. The Solution: The "Smart Trio"

The authors created a new system that combines three simple but powerful tools. Think of it as a team of three specialists working together:

A. The Translator (Lemmatization)

Legal texts are messy. A lawyer might write "citing," "cited," "cites," or "citation." To a computer, these look like four different words.

The Analogy: Imagine a translator who instantly converts all these variations into their "root" form. So, "citing," "cited," and "cites" all become just "cite."
Why it helps: It stops the robot from getting confused by grammar tricks and lets it focus on the meaning.

B. The Dictionary (FastText Embeddings)

Legal documents are full of rare words, Latin phrases, and terms that don't exist in normal dictionaries.

The Analogy: Instead of just looking up a whole word, this tool looks at the letters inside the word. If the robot sees a weird legal word it's never seen before, it breaks it down into small chunks (like "un-", "just-", "-able") to guess what it means based on its parts.
Why it helps: It's like having a detective who can figure out a stranger's identity even if they are wearing a disguise, just by looking at their shoes and hat.

C. The Scanner (Multi-Kernel CNN)

This is the brain of the operation. The robot doesn't just read one word at a time; it scans the text with three different "magnifying glasses" at the same time.

The Analogy: Imagine looking at a painting through three different lenses:
- Lens 1 (Small): Looks at tiny details (2-word phrases).
- Lens 2 (Medium): Looks at standard sentences (3-word phrases).
- Lens 3 (Wide): Looks at the whole scene (5-word phrases).
Why it helps: By using all three lenses at once, the robot catches short legal cues, medium-sized arguments, and long, complex sentences all in one go.

3. The Results: The Underdog Wins

The team tested their "Lightweight Robot" against the "Heavy Robots" (like BERT) and some older, simpler methods. Here is what happened:

Accuracy: Their robot got 97.26% of the answers right. This is actually better than the giant, heavy BERT model (which got 97.12%).
Speed: This is the real magic. The heavy BERT model took 4.25 milliseconds to read a document. The new robot did it in 0.31 milliseconds.
- The Metaphor: If BERT is a Ferrari that takes 10 seconds to start its engine, the new robot is a Formula 1 car that starts instantly. It is 13 times faster.
Efficiency: The heavy robot needs a massive computer with 110 million "brain cells" (parameters). The new robot only needs 5.1 million. It's like comparing a supercomputer to a high-end smartphone.

4. Why Does This Matter?

Imagine a law firm with thousands of cases.

Before: They have to hire a team of tired lawyers to manually sort documents, or they have to pay for expensive, slow computers to do it.
Now: They can use this lightweight system. It's cheap, it runs on normal computers, and it sorts documents in the blink of an eye with almost perfect accuracy.

The Bottom Line

The paper proves that you don't always need the biggest, most expensive AI to solve a hard problem. By being clever about how you clean the data (the translator), how you understand words (the dictionary), and how you scan for patterns (the scanner), you can build a system that is faster, cheaper, and just as smart as the giants.

It's a reminder that sometimes, the best solution isn't to build a bigger hammer, but to build a sharper, lighter one.

1. Problem Statement

The legal domain faces a critical challenge due to the exponential growth of digital case law archives. Manual triage and classification of these documents are time-consuming and error-prone due to:

Complex Linguistics: Formalized registers, lengthy sentence structures, and specialized Latin terminology.
Morphological Variability: Inflectional variations and out-of-vocabulary (OOV) terms that standard NLP models struggle to normalize.
Computational Costs: State-of-the-art solutions often rely on heavy Transformer models (e.g., BERT, Legal-BERT) which require massive computational resources, large training datasets, and suffer from high inference latency, making them unsuitable for real-time or resource-constrained legal analytics.

The specific task addressed is Citation-Treatment Classification: determining how a current judgment treats a cited authority (e.g., cited, referred to, distinguished, overruled, positive, neutral, negative).

2. Methodology

The authors propose a lightweight, high-accuracy pipeline that combines traditional linguistic preprocessing with deep learning, avoiding the computational overhead of large Transformers.

A. Data Preprocessing Pipeline

Cleaning & Normalization: Removal of URLs, digits, punctuation, and boilerplate legal text. Conversion to lowercase.
Tokenization & Stopword Removal: Splitting text into tokens and removing standard English stopwords and short tokens (<3 chars).
Lemmatization: A critical step where tokens are reduced to their canonical dictionary form (e.g., citing, cited $\to$ cite) using WordNet. This mitigates lexical sparsity caused by inflectional variations common in legal texts.
Embedding: The lemmatized corpus is used to train FastText embeddings (dimension $d=500$ ). FastText is chosen for its ability to handle OOV words and rare legal terms via subword (character n-gram) decomposition.

B. Proposed Architecture: Multi-Kernel 1D-CNN

The core model is a Convolutional Neural Network (CNN) designed to capture legal patterns at multiple granularities simultaneously:

Input: Sequence of FastText vectors ( $X \in \mathbb{R}^{L \times d}$ ).
Parallel Convolutional Branches: Three parallel 1D convolutional layers with different kernel sizes:
- $k=2$ : Captures short citation bigrams.
- $k=3$ : Captures mid-range legal phrases.
- $k=5$ : Captures longer clausal patterns.
Filtering: Each kernel size uses 128 filters with ReLU activation.
Pooling: Global Max Pooling is applied to each branch to extract the most salient features.
Classification: Pooled features are concatenated, passed through a Dropout layer (rate 0.4), and fed into a fully connected Softmax layer for multi-class prediction.
Training: End-to-end training using cross-entropy loss with the Adam optimizer. FastText embeddings are fine-tuned jointly with the CNN parameters.

3. Key Contributions

Lightweight High-Performance Pipeline: The proposed CNN + FastText + Lemmatization architecture achieves near-Transformer accuracy (97.26%) while being significantly faster and smaller.
Effective Handling of Legal Morphology: The study empirically validates that combining lemmatization with subword-aware FastText embeddings outperforms standard deep learning baselines in handling morphological variants and OOV legal terms.
Efficiency vs. Performance Trade-off: The model operates with only 5.1 million parameters (compared to BERT's 110M) and achieves an inference latency of 0.31 ms (13x faster than BERT), making it viable for real-time deployment.
Domain-Specific Preprocessing: A reusable pipeline tailored to case-law semantics that generalizes without manual feature engineering.

4. Experimental Results

The system was evaluated on a public corpus of 25,000 annotated legal documents (75% training / 25% test).

Performance Metrics

Model	Accuracy (%)	Macro F1-Score (%)	Parameters (M)	Inference Latency (ms)
Proposed (CNN+FT+Lem)	97.26	96.82	5.1	0.31
BERT (Fine-tuned)	97.12	96.88	110.0	4.25
LSTM + FastText	95.68	95.10	7.2	0.78
CNN (Random Embed)	93.51	92.88	3.6	0.45
KNN (TF-IDF)	89.42	88.60	0.02	1.87

Accuracy: The proposed model achieved 97.26%, surpassing fine-tuned BERT (97.12%) and all other baselines.
AUC-ROC: It attained the highest Area Under the Curve at 97.83%, indicating superior sensitivity and specificity across all citation classes.
Efficiency: The model is 13x faster in inference and requires ~95% fewer parameters than BERT.

Error Analysis

Confusion Matrix: Errors were primarily confined to semantically adjacent categories (e.g., Positive vs. Neutral, Distinguished vs. Overruled), reflecting the inherent linguistic ambiguity in legal phrasing rather than model failure.
Ablation Study: The multi-kernel configuration $\{2, 3, 5\}$ was proven essential. Using a single kernel ( $k=3$ ) dropped accuracy to 95.8%, while the triple-kernel approach maximized performance by capturing multi-scale patterns.
Robustness: Under Gaussian noise injection ( $\sigma=0.05$ ), accuracy degraded by only 1.3% (to 95.9%), demonstrating high resilience.

5. Significance and Conclusion

This paper demonstrates that carefully designed, lightweight convolutional architectures can rival heavyweight Transformer models in specialized domains like legal text analysis.

Practical Impact: The proposed system offers a scalable, resource-efficient alternative for legal institutions, enabling real-time document triage without the need for expensive GPU clusters or massive pre-training data.
Scientific Contribution: It challenges the assumption that large language models are strictly necessary for high-accuracy legal NLP, highlighting the value of domain-specific preprocessing (lemmatization) and subword embeddings (FastText) when combined with efficient architectures.
Future Directions: The authors suggest extending the pipeline to multilingual/multijurisdictional corpora, integrating lightweight Transformer layers for long-range context, and incorporating Explainable AI (XAI) to aid legal practitioners in understanding model predictions.