DML: Dynamic Multi-Granularity Learning for BERT-Based Document Reranking

Abstract

Recently, pre-trained language models have been successfully applied to the task of text retrieval and ranking. However, in real scenes, users' click behavior is usually affected by selection, position, or exposure bias, which may lead to insufficient positive annotations and introduce additional noise. And for different candidate documents of the same query, the previous optimization objectives usually use a single granularity and static loss weights. It makes the performance of ranking models more susceptible to the bias issue mentioned above. Thus, in this paper, we focus on BERT-based document reranking and propose Dynamic Multi-Granularity Learning (DML). By introducing Gaussian distribution into traditional loss functions, the weights of different documents can change dynamically according to the prediction probability to avoid the impact of unlabeled positive documents. Besides, both document-granularity and instance-granularity are considered to balance relative relations and absolute scores of candidate documents. Extensive experiments show that DML significantly outperforms previous state-of-the-art models on the MS MARCO document ranking dataset.