Detection of Grice’s Maxims Violations: A Deep Learning Approach Using the CMV Dataset

Our research aims to create a computational model that can identify breaches of Grice's maxims in conversational discourse, focusing specifically on relevance and manner.

Fine-tuning FLAN-T5 Model

This code is training the Flan-T5-Base model on Manner Violation Detection without augmentations.

In order to finetune on Relevance Violation Detection please change line 71 to get_relevance_class.

To use augmentations please uncomment lines 100-123.

Results

Detection	Model Name	Class 0 - Precision	Class 0 - Recall	Class 0 - F1-Score	Class 1 - Precision	Class 1 - Recall	Class 1 - F1-Score	Accuracy
Relevance	GPT API	0.923	0.431	0.593	0.071	0.523	0.12	0.441
Relevance	Over Sampling - Linear SVM	0.949	0.428	0.59	0.078	0.656	0.133	0.441
Relevance	Over Sampling - Flan-T5	0.96	0.95	0.95	0.28	0.32	0.29	0.91
Manner	GPT API	0.8	0.5	0.64	0.16	0.59	0.25	0.51
Manner	Under Sampling - Flan-T5	0.92	0.66	0.77	0.21	0.62	0.31	0.66

Read More