Generation-Augmented Retrieval for Open-domain Question Answering

ACL-IJCNLP 2021 |

Conventional sparse retrieval methods such as TF-IDF and BM25 are simple and efficient, but solely rely on lexical overlap without semantic matching. Recent dense retrieval methods learn latent representations to tackle the lexical mismatch problem, while being more computationally expensive and insufficient for exact matching as they embed the text sequence into a single vector with limited capacity. In this paper, we present Generation-Augmented Retrieval (GAR), a query expansion method that augments a query with relevant contexts through text generation. We demonstrate on open-domain question answering that the generated contexts significantly enrich the semantics of the queries and thus GAR with sparse representations (BM25) achieves comparable or better performance than the state-of-the-art dense methods such as DPR \cite{karpukhin2020dense}. We show that generating various contexts of a query is beneficial as fusing their results consistently yields better retrieval accuracy. Moreover, as sparse and dense representations are often complementary, GAR can be easily combined with DPR to achieve even better performance. Furthermore, GAR achieves the state-of-the-art performance on the Natural Questions and TriviaQA datasets under the extractive setting when equipped with an extractive reader, and consistently outperforms other retrieval methods when the same generative reader is used.