Generative Retrieval (GR) directly generates document identifiers (DocIDs) without external indices. However, existing GR methods suffer from performance degradation due to the gap between queries seen during training and those encountered at inferenc...
Generative Retrieval (GR) directly generates document identifiers (DocIDs) without external indices. However, existing GR methods suffer from performance degradation due to the gap between queries seen during training and those encountered at inference time. This thesis analyzes and addresses this training-inference gap from two perspectives.
First, we identify granularity collapse in GR: during training, GR models develop a preference for generating DocIDs with a specific granularity, causing retrieval failures when it differs from that of the target document's DocID. We propose to quantify this mismatch using "isotropy" and confirm that retrieval accuracy degrades as such granularity mismatch increases. To address this, we propose Granularity Routing for Generative Retrieval (GR^2), which trains multiple expert models specialized for different granularity levels and routes each query to the appropriate expert. By manipulating the token embedding space through principal component analysis, we generate DocIDs with target granularity levels and select optimal DocIDs through a retrieval-aware optimization process.
Second, we propose D3 (Dynamic DocID Decoding) to resolve the misalignment between statically indexed DocIDs and query intents. D3 consists of a verification module that detects intent misalignment at inference time and a dynamic decoding module that extends DocIDs dynamically. It measures the probability difference between constrained and unconstrained decoding to detect misalignment, and extends DocIDs with tokens extracted from queries to better reflect query intent.
Experiments demonstrate that our methods achieve consistent performance improvements over existing approaches. Notably, we observe greater improvements on queries with unseen intents during training, indicating that our approach effectively mitigates the training-inference gap in generative retrieval.