NLP Meetup at Yext: How to improve on BERT embeddings for long-form doc search

On April 25th, the Yext Developer team hosted the NLP Meetup group in NYC. This talk, presented by Yext's data science lead, Michael Misiewicz, and teammate, Allison Rossetto, covered improved BERT embeddings for long\-form document search as implemented in Yext's newest Document Search algorithm. At the event, we shared our results using Hidden Markov Models, document segmentation, BERT embeddings, and BM25 encoding to build an end\-to\-end document search system. Architecture, training procedures and performance considerations of the entire pipeline as well as our custom BERT embedding model (which uses a cross encoding architecture) were covered. The Data Science organization at Yext is responsible for developing algorithmic product features, a large number of which underpin Yext's search engine Answers. Before working at Yext, Michael worked at NYC ad\-tech company AppNexus, leading anti\-fraud efforts. Prior to that he attended McGill University in Montreal where he researched cellular biology of neurons using bioinformatics methods. Join our [developer community](https://www.yext.com/resources/subscribe\-developers) to stay tuned on future events and happenings at Yext. To learn more about Yext's latest Algorithm update check out the Spring '22 [Blog](https://hitchhikers.yext.com/blog/Spring\-22\-Release\-Is\-Now\-Live) or [Release Notes](https://hitchhikers.yext.com/releases/spring22/).