This repo contains a collection of Kaggle notebooks that were used to train the models in our paper.
trainandtestcontains skeleton code for training and testing the NLI models, which were adapted to suit different models and datasets for our experimentsbm25-crossencodercontains the code used to generate the top-100 snippets for each claim, and the top-5 reranking using a cross encoder. The dataset is available at https://www.kaggle.com/datasets/askuiper/quantemp, which also includes the QuanTemp dataset itself, and the corpus.claimdecompcontains code used to train the BART model to decompose the claims in QuanTemp, and the reranking code to extract the top-5 snippets from the bm25 top-100. The dataset is available at https://www.kaggle.com/datasets/askuiper/quantemp-decomp-data.temporal-rerankingcontains the code used to rerank the top-5 snippets from the bm25 top-100 by taking into account temporal information. The dataset is available at https://www.kaggle.com/datasets/lucasvm/quantemp-temporal-rerank.strategyqacontains the code used to invistigate the use of a model trained on StrategyQA in order to decompose questions and answer them iteratively using a Deberta model trained on SQuAD. The dataset is available at https://www.kaggle.com/datasets/lucasvm/strategyqa-decomp-quantemp.