iIMoRaD31/Darija-Coherence-assessment-

This repository contains the implementation of the coherence assessment models discussed in the Morad Litime's paper "Improving Dialogue Management for DarijaGenie through Neural Coherence Modeling" # Improving Dialogue Management for DarijaGenie through Neural Coherence Modeling This code is an implementation of the Local Coherence Discriminator (LCD) and the Hierarchical Transformer (HT) described in: Morad Litime, "Improving Dialogue Management for DarijaGenie through Neural Coherence Modeling", unpublished manuscript, 2025. researchgate.net This implementation is based on PyTorch 2.7.1+cu126 ## Introduction Coherence is such an important outcome of language learning, it reflects the learner’s ability to produce meaningful text that follows a logical flow. Coherence is even more essential in dialogues, it ensures mutual understanding in human conversation. Historically, coherence has been an active field of research in natural language processing, going from the early rule-based and linguistic theories to the latest data-driven neural methods. Our focus is coherence checking for language learning. Previous work has been done to develop DarijaGenie, a tool that enables learners to practice the Darija language via scenario-based interaction with a bot. The primary objective of this work is to replace the existing dialogue management framework in DarijaGenie with a new system that leverages a coherence detection tool. Our general approach is to evaluate the performance of two architectures on the task of 2-way classification and integrate the best performer into DarijaGenie. ## Datasets You can extract the different Zip files containing the dialogues Each Zip file is specific to a training setting: -"final_datasets" are used for LCD in-domain -"final_datasets_cross" and "Test" are used in LCD cross-domain and HT cross-domain -"final_datsets_train" and "final_datasets_test" are used for HT in-domain ## Training process In case you wanted to run training from scratch you can go ahead and run the cells in the provided notebooks Each model has two notebooks each for a training setting (in-domain and cross-domain …