Generative Retrieval for Conversational Question Answering

The Research paper by Microsoft proposes a novel information retrieval system called Generative retrieval for Conversational Question Answering (GCoQA) to overcome limitations in existing conversational retrieval methods, particularly those based on the dual-encoder architecture. GCoQA utilizes an encoder-decoder (autoregressive) architecture to retrieve relevant passages by generating unique passage identifiers token-by-token. This generative approach eliminates the “embedding bottleneck” of single-vector representations and enables token-level interaction with the full conversation context, leading to more accurate retrieval. Experiments showed that GCoQA achieved significant relative improvements of +13.6% in passage retrieval and +42.9% in document retrieval, while consuming only 1/10 of the memory and being significantly faster than baseline methods.