Legal Documents Categorization by Compression: Beyond Feature Engineering

TL;DR

Categorizing legal texts is a "hard" problem in AI due to linguistic complexity and the lack of large labeled datasets. This paper proposes a "feature-free" approach using data compression algorithms (like Gzip) to measure document similarity based on Kolmogorov complexity. In a study of Italian normative texts, the Best-Compression Neighbor (BCN) method significantly outperformed standard Support Vector Machines (SMO) and Decision Trees, proving that information theory can succeed where traditional NLP struggles.

Problem & Motivation: The "Residual Power" Grey Zone

In the Italian legal system, "residual legislative power" refers to all subjects not explicitly reserved for the State or concurrent jurisdiction. This "negative definition" creates a classification nightmare: identifying which subjects belong to the Regions requires analyzing hundreds of subjective, complex judgments from the Constitutional Court.

Traditional Text Categorization (TC) faces two major walls here:

1. Semantic Overlap: Laws for "Industry" and "Business Incentives" are linguistically almost identical.
2. The Small Data Problem: High-quality training sets for specific legal niches are rare. Traditional feature-based models (TF-IDF + SVM) require large amounts of data to overcome the noise of convoluted legal syntax.

Methodology: The Logic of "Zipping"

Instead of converting text into vectors of words (Bag-of-Words), the authors treat text as a string of information. They utilize the concept of Kolmogorov Complexity ($K(x)$)—the length of the shortest program that can produce a string $x$.

Core Intuition

If document $A$ and document $B$ are similar, they contain similar patterns. If you concatenate them and compress the result ($A+B$), the compressor will find those shared patterns and use fewer bits to represent the second document.

The authors investigated three specific protocols:

• AMDL (Approximate Minimum Description Length): Compressing a test file against a massive archive of all training documents in a class.
• BCN (Best-Compression Neighbor): A k-NN variant that finds the single training document that provides the best compression "boost" to the test document.
• NCD (Normalized Compression Distance): A mathematically rigorous metric that normalizes the distance between 0 (identical) and 1 (completely different).

Figure 1: Complex structure of legal judgments including "Object provisions" and "Parameters" which the algorithm must process.

Experiments & Results

The researchers tested these methods on a dataset of Italian object provisions spanning 7 categories (Agriculture, Tourism, Education, etc.).

Key Findings:

• BCN Scalability: The BCN method was the clear winner. Interestingly, when the model was allowed to pick a "top 3" list of candidates, it was correct 97.14% of the time.
• Parameter-Free Advantage: Unlike SVMs, which require careful tuning of kernels and feature selection, compression techniques require zero preprocessing (no stemming, no stop-word removal).
• Robustness to Data Size: In a follow-up experiment on Italian literature (authorship attribution), the compression models scaled exceptionally well as document size increased, reaching 80.41% accuracy.

Figure 2: Accuracy comparison highlighting the superiority of compression modes (BCN/NCD) over standard ML (J48/SMO).

Critical Analysis & Conclusion

This paper serves as a vital reminder that in the era of high-dimensional feature engineering, the fundamental principles of Information Theory remain potent.

The Takeaway: Compression-based classification is an ideal "first-pass" filter. It handles the morphological variants of languages like Italian naturally and captures phrasal structures that simple word-counts miss.

Limitations: The current approach is computationally expensive for massive datasets because $k$-NN requires comparing each test file to every training file. However, for specialized legal domains where the "training set" consists of a few hundred landmark cases, this approach is both more accurate and easier to deploy than a deep learning model.

Future Outlook: The authors envision an "extended architecture" where compression handles the initial candidate selection (shortlisting 2-3 classes), followed by a semantically-rich ontological tool to make the final legal distinction.