CEQE: Contextualized Embeddings for Query Expansion

In this work we leverage recent advances in context-sensitive language models to improve the task of query expansion. Contextualized word representation models, such as ELMo and BERT, are rapidly replacing static embedding models. We propose a new model, Embeddings for Query Expansion (CEQE), that utilizes query-focused contextualized vectors. study behavior contextual representations generated expansion ad-hoc document retrieval. conduct our experiments on probabilistic retrieval well combination with neural ranking evaluate CEQE two standard TREC collections: Robust Deep Learning. find outperforms embedding-based methods multiple collections (by up 18% 31% Learning average precision) also improves over proven pseudo-relevance feedback (PRF) further passes reranking result continued gains effectiveness CEQE-based approaches outperforming other approaches. The final model incorporating score achieves 5% P@20 2% AP state-of-the-art transformer-based re-ranking Birch.