📄 A Grice-ful Examination of Offensive Language: Using NLP Methods to Assess the Co-operative Principle

Summary

The paper asks whether offensive online talk breaks Grice's Co-operative Principle more often than ordinary talk. The authors turn each of the four maxims (Quality, Quantity, Relevance, Manner) into an automatic, measurable proxy built from NLP tools. They apply these proxies to offensive and non-offensive instances from ToxiChat, a corpus of Reddit-style threads with titles and replies. They then compare the two classes statistically. Instances the metrics flag are hand-checked in a qualitative discourse analysis. The main finding is that the classes mostly behave alike. Clear differences appear only for Manner, driven largely by heavy profanity, and partly for Quality, since untruthful comments are more frequent among offensive items. The close reading also shows the metrics sometimes misfire, for example on a playful "No homo" / "yes homo" exchange. The work matters because it shows that pragmatic theory can be semi-automated. NLP narrows a large dataset to candidate cases, and human analysis then interprets them. The authors present this as a model for computationally assisted discourse analysis.

Key points

• Data: ToxiChat conversations labelled offensive vs. safe. The Quantity and Relevance metrics look at the whole thread (title plus replies), while Quality and Manner judge single utterances.
• Manner is operationalised through several signals: an ambiguity score, Flesch Readability, and profanity presence. Profanity is far more common in the offensive class (per-class splits are given in Figure 5).
• Statistically, offensive and non-offensive discourse do not differ significantly on most maxim metrics. The exceptions are Manner and some aspects of Quality.
• The paper separates violation (an unintended departure from a maxim) from flouting (a deliberate departure that triggers an implicature).
• Stated limitations: the metrics are imperfect and lack expert validation, and each qualitative example was checked against only one maxim, although several breaches can co-occur.

Interesting ideas

• I1. Each abstract pragmatic norm is split into its sub-maxims, and each sub-maxim gets its own cheap, measurable proxy, which makes a vague theory testable and comparable across classes.
• I2. The automatic metric is used only as a filter, and humans then read the flagged items; this exposed the metric's false positives and its errors on playful or broadened replies.
• I3. Some maxims are scored at conversation level and others at utterance level, because relevance and informativeness only make sense against prior turns.
• I4. Violation and flouting are kept apart, so the same surface breach is read either as an error or as a deliberate signal of implied meaning.
• I5. Scalar implicature over an ordered scale (excellent > good > OK) is used as the textbook way to infer unstated meaning.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): crate prag could score Quantity, Relevance and Manner per utterance with explicit, inspectable proxies, giving each implicature candidate a traceable trigger.
• I2 (learn): treat pragmatic detectors as candidate generators, and check a random sample of their flags against negative controls (for example, safe banter) before trusting them as silver labels.
• I3 (world): relevance checks should compare a reply's frames against entities and claims already in the world of the text, not against the reply alone.
• I4 (frames): when prag detects a maxim breach, it could branch explicitly into "noise" vs. "intended implicature/irony" and record which cue chose the branch.
• I5 (answer): readers of fables and tales could derive "merely satisfactory" style inferences from value scales stored in compiled knowledge, explaining the answer through the scale.