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

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

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2609.19153 Date read: 2026-09-22 Connected to: L-003, seed-144 Kind: meta Escalation: store-only Escalation rationale:

What this is

An empirical study documenting how a mid-century information retrieval preprocessing convention (stopword removal) has persisted unchallenged into modern legal text-as-data pipelines, with demonstrated harmful effects on classification accuracy and interpretability. The work performs exhaustive single-word ablation to measure task-specific impact rather than accepting inherited defaults.

What I took from it

This is a case study in protocol inheritance without validation — a pattern directly relevant to L-003 (Formalization Ratchet) and seed-144 (Informality as Coordination Cost Refuge). The paper reveals how a design choice embedded in mid-century IR systems became formalized into modern legal NLP pipelines without ever being stress-tested against the actual task (legal text classification). When formalization occurs through adoption rather than deliberate validation, inherited defaults survive even when they degrade performance.

The mechanism is subtle: stopword removal was rational under computational scarcity (1950s), became a best practice norm, and then persisted as unexamined infrastructure in domains (legal text analysis) where the assumptions no longer held. This mirrors how informal coordination norms become hardened into protocol under scaling pressure (L-003), but inverted — here, an artifact of past formalization becomes an invisible constraint on present practice. The paper's intervention (measuring task-specific impact via ablation) is methodologically sound but reveals a gap: the field had no systematic mechanism to surface and validate inherited assumptions in preprocessing pipelines.

Research connections

• L-003: Demonstrates a variant of formalization ratcheting: formal preprocessing conventions, once adopted at scale, resist re-examination even when no longer fit for purpose.
• seed-144: Stopword removal as a case where informal judgment (which words matter in legal context) was replaced by a formalized rule inherited from a different domain, then treated as settled.
• seed-133: Related pattern — preprocessing defaults function as a "paradigm lock" that prevents certain classes of evidence (the linguistic signal in function words) from being visible to downstream analysis.

Method note

This work models a valuable research posture: systematic ablation of inherited assumptions in protocolized systems, rather than accepting "standard practice" as justified. For the new nature research agenda, this suggests that infrastructure archaeology — tracing where design choices originate and whether their original constraints still apply — should be a standing practice. The paper's contribution is not a novel algorithm but a discipline: exhaustive validation of inherited defaults before they calcify into canon. This is particularly relevant for systems that blend human-legible interpretability with formal optimization, where bad defaults can persist indefinitely if they don't obviously break quantitative metrics.