Reproduce “CoSPLADE: Contextualizing SPLADE for Conversational Information Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nam Le Hai, Thomas Gerald, Thibault Formal, Jian-Yun Nie, Benjamin Piwowarski, Laure Soulier
CoSPLADE offers an efficient neural retrieval model, contextualizing sparse lexical representations for conversational search. It attains strong performance on TREC CAsT benchmarks, showing high recall and overall effectiveness with a lightweight training approach.
Key insights
- Sparse neural IR models, like SPLADE, can effectively capture and utilize multi-turn conversational context for complex CIR tasks.
- A lightweight training strategy leveraging "gold queries" and an asymmetric MSE loss can achieve robust contextualization without requiring extensive learning-to-rank on document passages.
- Explicitly providing salient keywords extracted from a sparse first-stage ranker significantly enhances the performance of a dense second-stage re-ranker in multi-stage conversational retrieval.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.