Models Combination in Multi-stage Information Retrieval Architectures
- vrijdag 20 januari 2023
Niels Bohrweg 1
2333 CA Leiden
Or attend via ZoomZoom link
Many state-of-the-art retrieval methods are limited by their efficiency as they are too slow to be applied on full retrieval test collections. Cascading the stages of simpler, less precise, but computationally efficient models followed by these slow but more effective models thus might be an option how such models can be applied. The uncertainty of the retrieval can be further reduced by using a combination of multiple information retrieval models, which is a well-known method for increasing the performance of retrieval as well as its robustness. Application of multiple-stages ranking and model combination raises the questions of how they might be best used together. For example, should a combination be performed after the first stage or after the second stage ranking? The talk will explore this design space using TREC Podcast and Deep Learning track test collections.