SimANS: Simple Ambiguous Negatives Sampling for Dense Text Retrieval
This work addresses a key bottleneck in dense retrieval training for information retrieval systems, offering an incremental improvement over existing negative sampling strategies.
The paper tackles the problem of uninformative or false negatives in dense text retrieval by proposing SimANS, a method that samples ambiguous negatives ranked around positives, leading to improved training effectiveness across four public and one industry datasets.
Sampling proper negatives from a large document pool is vital to effectively train a dense retrieval model. However, existing negative sampling strategies suffer from the uninformative or false negative problem. In this work, we empirically show that according to the measured relevance scores, the negatives ranked around the positives are generally more informative and less likely to be false negatives. Intuitively, these negatives are not too hard (\emph{may be false negatives}) or too easy (\emph{uninformative}). They are the ambiguous negatives and need more attention during training. Thus, we propose a simple ambiguous negatives sampling method, SimANS, which incorporates a new sampling probability distribution to sample more ambiguous negatives. Extensive experiments on four public and one industry datasets show the effectiveness of our approach. We made the code and models publicly available in \url{https://github.com/microsoft/SimXNS}.