Helping users to quickly find relevant products is of key importance to online marketplaces. The dominant search paradigms today are keyword search for descriptions and structured search for properties such as dates and prices.
At AutoScout24, Europe’s largest online car market, we want to enable users to search using natural language. However, understanding free text and mapping the relevant parts of a query to either keywords or filters is non-trivial.
Ideally, we want a system that can gracefully handle synonyms and typos, understand numeric comparisons and ignore irrelevant inputs. We tackle these challenges with sequence-to-sequence learning.
This talk describes the process of training and deploying a scalable solution that builds on state-of-the art research in neural machine translation. It also shows how the Tensor2Tensor library and TFX can be used to quickly go from idea to production.
Markus Ludwig is a Senior Data Scientist at Scout24 where he builds and deploys machine learning systems that power search and discovery. Before that he worked as a researcher, lecturer and consultant. Markus holds a Ph.D. in Computational Finance from the University of Zurich, Switzerland.