ex_ex_gen_multi-hop_QA

In this study, we present a pipeline for an explainable, extractive-generative multi-hop OpenQA model trained on the HotPotQA dataset. Our model features a Retriever Reader architecture, which consists of an Interaction-based Retriever and Extractor and a Generative module. We show that our model outperforms the baseline model of the dataset. Results show that the Retriever and Extractor submodules perform better on comparison question types rather than bridge question types, while the opposite is true for the Generator submodule. Performance decreases for the Retriever with increasing reasoning difficulty, but not for the Extractor and Generator. Additionally, fine-tuning on the SQuAD v2 dataset was found to enhance multi-hop reasoning ability in the Generator module.