Document retrieval is one of the most challenging tasks in Information Retrieval. It requires handling longer contexts, often resulting in higher query latency and increased computational overhead. Recently, Learned Sparse Retrieval (LSR) has emerged as a promising approach to address these challenges. Some have proposed adapting the LSR approach to longer documents by aggregating segmented document using different post-hoc methods, including n-grams and proximity scores, adjusting representations, and learning to ensemble all signals. In this study, we aim to reproduce and examine the mechanisms of adapting LSR for long documents. Our reproducibility experiments confirmed the importance of specific segments, with the first segment consistently dominating document retrieval performance. Furthermore, We re-evaluate recently proposed methods -- ExactSDM and SoftSDM -- across varying document lengths, from short (up to 2 segments) to longer (3+ segments). We also designed multiple analyses to probe the reproduced methods and shed light on the impact of global information on adapting LSR to longer contexts. The complete code and implementation for this project is available at:this https URL.
View on arXiv@article{lionis2025_2503.23824,
title={ On the Reproducibility of Learned Sparse Retrieval Adaptations for Long Documents },
author={ Emmanouil Georgios Lionis and Jia-Huei Ju },
journal={arXiv preprint arXiv:2503.23824},
year={ 2025 }
}