devesh-002 / nlp_project

Transformer Models for Text Coherence Assessment

In this repository, we are working on Text Coherence Assessment of paper.

Install Preprocessed dataset from here and add in folder processed_data, folder.

To train the model, you have to use the run.sh file and change the parameters in it as required. Then simply do the following:

The metrics are as follows:

• corpus can take one of 'gcdc' or 'wsj'.
• sub_corpus can take anyone value from 'Clinton', 'Enron', 'Yelp' or 'Yahoo' if corpus is gcdc
• arch can take one of vanilla, hierarchical
• task can take one of 3-way-classification, minority-classification,sentence-ordering or sentence-score-prediction for GCDC dataset and only sentence-ordering for WSJ dataset
• model_name defines transformer model to use. (by-default its's roberta-base) For training custom model

bash try.sh

To make changes to try.sh file

python3 main.py --arch <arch_name> --corpus <corpus_name>   --task <task_name>

For evaluating on datasets, do the following:

bash infer.sh

To make changes in inferences:

python3 main.py --sub_corpus <name if gcdc> --inference --arch <arch_name>  --corpus <dataset_name>  --freeze_emb_layer  --task <task_name>  --checkpoint_path <saved_checkpoint_path>

We also have submitted the models here