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

Tue, 01 Oct 2024 00:00:00 +0000

Research setting

Pragmatics studies how people mean more than their words say. In this field, the philosopher H. P. Grice described the cooperative principle: in a conversation, each contribution should serve the shared purpose of the exchange. He broke the principle down into four maxims, listed below.

Maxim	What it asks of a speaker
Quality	Do not say what you believe to be false, or what you lack evidence for.
Quantity	Be as informative as required, and not more.
Relevance	Be relevant.
Manner	Avoid obscurity and ambiguity. Be brief and orderly.

The four Gricean Maxims. Source: Table 1 of the paper.

A speaker who breaks a maxim by accident violates it. A speaker who breaks it on purpose, for a rhetorical or humorous effect, flouts it. The paper asks whether offensive comments in online conversations break the maxims more often than safe ones. Its data is ToxiChat, a set of Reddit threads in which each turn (one message in the thread) is labeled as offensive or safe.

Title: [Question] Why do Libertarians get so much flack from the rest of reddit Like seriously I was downvoted when I said "Libertarian is a good one" on a post about third party voting.Safe Because the rest of reddit are unironically communists.Offensive Bullshit most are democratsOffensive

A three-turn thread from ToxiChat with its offensiveness labels (Table 4 of the paper).

Motivation

Natural Language Processing (NLP) usually treats offensive language as a detection task: a model labels each message as offensive or not. This view says little about how an offensive conversation works. Studies that apply the Gricean Maxims to hate speech read the texts by hand. Studies that turn the maxims into metrics do not look at offensive language.

Approach	Uses the Gricean Maxims	Measures them with NLP tools	Studies offensive language
Sarcasm in hate speech on Instagram (Pasa et al., 2021)	yes	no	yes
Dialog evaluation by user survey (Jwalapuram, 2017)	yes	no	no
Maxims as metrics (Freihat et al., 2018; Tewari et al., 2020; Ge et al., 2023)	yes	yes	no
Speech acts in hate speech (Hidayati and Arifuddin, 2021)	no	no	yes
Speech-act-guided toxic stance detection (Upadhyaya et al., 2023)	no	yes	yes
This paper This work	yes	yes	yes

How the paper relates to the prior work it discusses (Sections 1 and 3).

Objective. Translate each Gricean Maxim into an NLP metric, compare offensive and safe comments with these metrics, and use them to pick examples for a close qualitative reading.

Approach

Each maxim gets its own NLP tool, which scores every human turn of ToxiChat. The paper then compares the scores of offensive and safe turns. Lastly, it filters the offensive turns by these scores and reads selected examples by hand, as in traditional discourse analysis.

ToxiChat turns Reddit threads, each turn labeled offensive or safe.
One metric per maxim Each turn gets four maxim scores.
- Quality Deception classifier
- Quantity Syntactic cohesion
- Relevance Embedding similarity
- Manner Ambiguity, readability, profanity
Compare classes Offensive turns against safe turns, maxim by maxim.
Close reading Selected offensive examples analyzed by hand.

The analysis pipeline of the paper (Section 4). Only the human turns of the ToxiChat training set are used; the bot turns are left out.

Maxim	Tool	Signal of a broken maxim
Quality	A BERT text classifier trained to detect deceptive hotel reviews (1,600 reviews about 20 hotels), with a macro-F1 of 0.926 ± 0.021 on its test set	The turn is predicted as untrue
Quantity	Informativeness from syntactic cohesion, scored from 0 to 1	Below 0.25 (under-informative) or above 0.75 (over-informative)
Relevance	Cosine similarity of BERT embeddings between a reply, the thread title, and the previous reply	The reply is judged not relevant
Manner	Lexical ambiguity from WordNet word senses, Flesch readability (higher is easier), and a profanity word list	High ambiguity, low readability, or profanity

Maxim metrics. Source: Section 4.1 of the paper.

Results

The paper scores the human turns of the ToxiChat training set and compares, for each maxim, the share of offensive and safe turns that break it.

53.8%of offensive turns contain profanity, against 16.4% of safe turns
40%of offensive turns are predicted as untrue, against 30% of safe turns
95.37%of offensive turns are optimally informative, against 95.79% of safe turns

Measure	Offensive	Safe
Quality: predicted untrue	371 (40%)	673 (30%)
Quantity: optimally informative	886 (95.37%)	2186 (95.79%)
Quantity: over-informative	21 (2.26%)	50 (2.19%)
Quantity: under-informative	22 (2.37%)	46 (2.02%)
Manner: contains profanity	53.8%	16.4%

Main results per maxim. Source: Tables 5 and 6 and Figure 5 of the paper.

Most replies are relevant in both classes, and irrelevant replies are more frequent among safe turns (Figure 2 of the paper).

Offensive and safe turns score almost the same on Quantity and Relevance. Ambiguity and readability are also similar in the two classes. The clear difference is profanity, a break of the Maxim of Manner, which is much more frequent in offensive turns. Untrue statements, a break of the Maxim of Quality, are also more frequent in offensive turns. In the close reading, the authors disagree with the Relevance metric on both examples they chose. The metrics can therefore err on single turns.

Takeaways