Research Summary: Stop Removing Stopwords: How an Inherited Preprocessing Default Distorts Legal Text-as-Data

Recent research from arXiv highlights a critical issue in empirical legal scholarship, where the inherited default practice of 'stopword removal' in text preprocessing significantly distorts analytical outcomes when treating judicial text as data. This inherited practice, stemming from mid-century information retrieval methods, has not been validated against modern classification accuracy, leading to flawed interpretations in data-driven legal analysis.

Why it matters

The findings are critical because reliance on flawed data preprocessing methods can lead to inaccurate insights and conclusions in data-driven decision-making, particularly in fields analyzing complex textual information. This can misdirect research, policy development, and strategic initiatives that depend on precise interpretation of textual data, undermining confidence in empirical methods.

Key insights

• Empirical legal scholarship increasingly utilizes judicial text as data, often relying on sparse, interpretable processing pipelines like TF-IDF features and linear classifiers.
• These processing pipelines inherit outdated preprocessing defaults, notably stopword removal, from mid-century information retrieval.
• The practice of stopword removal has not been validated against classification accuracy, despite its widespread adoption.
• The study introduces a single-word ablation method to directly measure the impact of preprocessing steps on downstream analytical objectives.
• Stopword removal is identified as an entrenched default that distorts analyses of legal text-as-data.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.19153