E-fashion Product Discovery via Deep Text Parsing
Uma Sawant, Vijay Gabale, Anand Subramanian · 2017
Transforming unstructured text into structured form is important for fashion e-commerce platforms that ingest tens of thousands of fashion products every day. While most of the e-commerce product extraction research focuses on extracting a single product from the product title using known keywords, little attention has been paid to discovering potentially multiple products present in the listing along with their respective relevant attributes, and leveraging the entire title and description text for this purpose. We fill this gap and propose a novel composition of sequence labeling and multi-task learning as an end-to-end trainable deep neural architecture. We systematically evaluate our approach on one of the largest tagged datasets in fashion e-commerce consisting of 25K listings labeled at word-level. Given 23 labels, we discover label-values with F1 score of 92.2%. When applied to 2M listings, we discovered 2.6M fashion items and 9.5M attribute values.