Combining Lexical and Dense Retrieval for Computationally Efficient Multi-hop Question Answering Georgios Sidiropoulos1, Nikos Voskarides2∗, Svitlana Vakulenko1, Evangelos Kanoulas1 1 University of Amsterdam, Amsterdam, The Netherlands 2 Amazon, Barcelona, Spain
[email protected],
[email protected],
[email protected],
[email protected] Abstract QA systems typically consist of (i) a retriever that identifies the passage/document in the underly- In simple open-domain question answering ing collection that contains the answer to the user’s (QA), dense retrieval has become one of the question, and (ii) a reader that extracts or generates standard approaches for retrieving the relevant the answer from the identified passage (Chen et al., passages to infer an answer. Recently, dense retrieval also achieved state-of-the-art results 2017). Given that often the answer cannot be found in multi-hop QA, where aggregating informa- in the top-ranked passage, inference follows a stan- tion from multiple pieces of information and dard beam-search procedure, where top-k passages reasoning over them is required. Despite their are retrieved and the reader scores are computed for success, dense retrieval methods are compu- all k passages (Lee et al., 2019). However, readers tationally intensive, requiring multiple GPUs are very sensitive to noise in the top-k passages, to train. In this work, we introduce a hy- thus making the performance of the retriever criti- brid (lexical and dense) retrieval approach that is highly competitive with the state-of-the-art cal for the performance of QA systems (Yang et al., dense retrieval models, while requiring sub- 2019).