Slides can be found here: https://docs.google.com/presentation/d/1uXOp2TnTqTaAY-RgCVMh7fLnwUfF0cNBDxpbxaVNPbM/edit?usp=sharing
Searching "flower beds" showed me beds even though we have garden beds on our site. This is a relevance algorithm issue because we don't
What it does
Uses NLP models to tag the same text with different meaning so we can calculate relevance using those
How we built it
We preprocessed all of wayfair UKs catalog, taking product names and descriptions and tagging them with part of speech and recognized entities. We then put that data into solr in another field
Challenges we ran into
The scale of data! We just hacked around it by making our own ad hoc scripts.
Accomplishments that we're proud of
We found a way to measure the impact of keyword search results without A/B testing! We identified how much further up the search results page that products (especially the ones people viewed or added to cart) moved. Although it's not a perfect measure, it was a great measure to get started
What we learned
Relevance and scoring is not a perfect science and there are more edge cases to finding the meaning of words that computers/machine learning models don't know how to handle
What's next for Using word sense in keyword search
Training a natural language model with wayfair data! We have tons of data, we can try to reshape it and capture where customers mean different things when they type the same word (eg. do most customers type "flower bed" and only click on garden products? is that different to when customers type "queen bed"? if so, "bed" probably means different things near "flower" and "queen"!