XFLT - Exploring Techniques for Generating Cross Lingual Factually Grounded Long Text

This repository contains the code and the processed dataset for cross lingual fact to long text generation. The paper describing the methods has been accepted at the European Conference on Artificial Intelligence (ECAI 2023).

Dataset

The processed XLAlign dataset is present in the XLAlign-Dataset directory. The directory contains the subdirectories for each of the languages.

Code

The code is present within the XFLT-code directory and is organised as follows.

1. clustering - This contains the code for training the fact organisation model
   • mT5-baseline - End-to-end clustering
   • statistical_clustering - Statistical spectral clustering
2. dataset_prep - This contains code for data preprocessing
   • coverage_classifier - Code and data for training coverage prompt classifier
3. eval_module - This contains the code for running evaluation using NLG metrics and the defined X-PARENT metrics.
4. generation - This contains code for training models using different methods
   • mT5-baseline - Training baseline mT5 method
   • prompt_uni - Training with coverage prompt
   • grounded_decoding - Inference with grounded decoding. Requires installing the modified transformers package included in the directory
5. rl_msme - This contains code for training with RL rewards

The default hyperparameter settings can be found in the run bash files in the respective directories. The requirements for all methods in the generation directory can be found in generation_reqs.txt. The same for RL can be found in rl_reqs.txt.

Contributors

Bhavyajeet Singh

Aditya Hari

Rahul Mehta

Tushar Abhishek

Manish Gupta

Vasudeva Varma