About this event

Using deep neural networks to map text to dense vector spaces (also known as semantic search), has brought tremendous progress to textual information retrieval. However, recent research showed that this widely adopted approach is extremely sensitive to data drift and performs poorly on out-of-domain and long tail queries.

In this talk, Nils will give an introduction to the next generation of semantic search architectures. First, he will talk about hybrid continuous-binary approaches that can lead up to a 100 times cost reduction for deploying semantic search. Then, he will present learned sparse representations that perform a lot better in terms of data drift, out-of-domain and long tail queries. These approaches are especially suited in domains with quickly evolving information needs like news retrieval.

Speaker

Nils Reimers is an expert on search relevance using pre-trained transformer networks. In 2018, he authored and open-sourced the popular sentence-transformers.