MultiContrievers: Analysis of Dense Retrieval Representations
This work addresses the problem of understanding information retention in dense retrievers for researchers in information retrieval and NLP, but it is incremental as it builds on existing models and probing methods.
The paper analyzed what information dense retrievers preserve or lose compared to their base language models, finding that contriever models have significantly increased extractability of specific information like gender and occupation, but this extractability correlates poorly with benchmark performance, and results are highly sensitive to random initializations and data shuffles.
Dense retrievers compress source documents into (possibly lossy) vector representations, yet there is little analysis of what information is lost versus preserved, and how it affects downstream tasks. We conduct the first analysis of the information captured by dense retrievers compared to the language models they are based on (e.g., BERT versus Contriever). We use 25 MultiBert checkpoints as randomized initialisations to train MultiContrievers, a set of 25 contriever models. We test whether specific pieces of information -- such as gender and occupation -- can be extracted from contriever vectors of wikipedia-like documents. We measure this extractability via information theoretic probing. We then examine the relationship of extractability to performance and gender bias, as well as the sensitivity of these results to many random initialisations and data shuffles. We find that (1) contriever models have significantly increased extractability, but extractability usually correlates poorly with benchmark performance 2) gender bias is present, but is not caused by the contriever representations 3) there is high sensitivity to both random initialisation and to data shuffle, suggesting that future retrieval research should test across a wider spread of both.