Passage Retrieval in question answering systems in Polish language

Abstract

This paper describes the submissions to Task 3 of PolEval 2022. Passage Retrieval is a problem of retrieving a passage relevant to the given query. It is an important problem with many practical use cases, especially in question answering. It is very beneficial if a model is generalizable, that is effective in various domains, even the ones it was not trained on. This is a challenge for many state-of-the-art models. In this paper I describe and test many different methods of approaching this problem -from standard techniques, such as BM25 and lemmatization to recently developed methods based on deep learning and transformers.