#DEAM
This repository contains code for DEAM paper. If you use it please cite it as:
@inproceedings{ghazarian2022deam,
title={DEAM: Dialogue Coherence Evaluation using AMR-based Semantic Manipulations},
author={Ghazarian, Sarik and Wen, Nuan and Galstyan, Aram and Peng, Nanyun},
booktitle={Proceedings of the Conference of the 60th Annual Meeting of the Association for Computational Linguistics (ACL)},
year={2022}
}
For any comments/issues/ideas pleaae feel free to contact me.
Use any virtualenv manager to install all the packages mentioned in the requirements.txt file.
In order to train DEAM, first you need to create negative samples using AMRs.
In this step, first you need to have text-to-AMR and AMR-to-text models. Follow amrlib github page to download two finetuned T5 models 1) model_parse_t5-v0_1_0 as the text-to-AMR model and model_generate_t5-v0_1_0 as the AMR-to-text model. For your convinience, these models can be downloaded from amr_models. You need to place this folder in th root. We use the model_parse_t5-v0_1_0 model to parse the conversations and retrieve AMRs. You can download the input data from Topical_chat and Persona_chat datasets from here. You needd to save this folder in the data directory.
python utils/amr_parse_convs.py --path data/topical_persona/ --fname train
In this step, we apply DEAM's manipulations to the extracted AMRs of the conversations and generate conversations from the manipulated AMRs leveraging the AMR-to-text model.
python code/manipulate_amr_incoherent_convs.py --data_path data/topical_persona/ --o_data_path data/topical_persona/ --fname train
Next, we finetune the Roberta-Large model on the set of data which includes both coherent/incoherent conversations resulted from AMRs/Manipulated_AMRs.
python code/ft_coh_model.py
--mode train
--train_data_path data/topical_persona/train_amr_manamr_cont_coref_pirel_eng.txt
--valid_data_path data/topical_persona/valid_amr_manamr_cont_coref_pirel_eng.txt
--model_path coh_models/
--model_name roberta-large --model_type roberta \
You can skip this step and download the trained DEAM model from here.
We use DEAM to predict scores for FED and DSTC9 test sets.
python code/ft_coh_model.py
--mode predict
--train_data_path data/topical_persona/train_amr_manamr_cont_coref_pirel_eng.txt
--valid_data_path data/topical_persona/valid_amr_manamr_cont_coref_pirel_eng.txt
--model_path coh_models/
--model_name roberta-large --model_type roberta \
At the end, we compute the correlation of scores predicted by DEAM for the test sets with human judgments.
python utils/compute_correlation.py --model_path coh_models/
We ignore each of the four manipulations in the negative samples generation step to examine their effectiveness. The results of the training data following different ablations can be found in the downloaded topical_persona/ folder. To train DEAM using ablations, predict scores and compute the correlations between DEAM's scores with human judgments, you can follow next steps (here we only show the scripts for the Contradiction ablation as an example):
python code/ft_coh_model.py
--mode train
--train_data_path data/topical_persona/train_amr_manamr_ablated_cont.txt
--valid_data_path data/topical_persona/valid_amr_manamr_ablated_cont.txt
--model_path coh_models/ablation_cont/
--model_name roberta-large --model_type roberta
--learning_rate=2e-6
--max_length=512
--train_batch_size=2 --valid_batch_size=2
--logging_steps=600 --save_steps=500
python code/ft_coh_model.py
--mode predict
--train_data_path data/topical_persona/train_amr_manamr_ablated_cont.txt
--valid_data_path data/topical_persona/valid_amr_manamr_ablated_cont.txt
--model_path coh_models/ablation_cont/
--model_name roberta-large --model_type roberta \
python utils/compute_correlation.py --model_path coh_models/ablation_cont/