Product discovery from E-commerce listings via deep text parsing
Uma Sawant, Vijay Gabale · 2018
Understanding unstructured text in e-commerce catalogs is important for product search and recommendations. In this paper, we tackle the product discovery problem for fashion e-commerce catalogs where each input listing text consists of descriptions of one or more products; each with its own set of attributes. For instance, [this RED printed short top paired with blue jeans makes you go green] contains two products: item top with attributes {pattern=printed, length=short, brand=RED} and item jeans with attributes {color=blue}. The task of product discovery is rendered quite challenging due to the complexity of fashion dictionary (e.g. RED is a brand or green is a metaphor) added to the difficulty of associating attributes to appropriate items (e.g. associating RED brand with item top). Beyond classical attribute extraction task, product discovery entails parsing multi-sentence listings to tag new items and attributes unknown to the underlying schema; at the same time, associating attributes to relevant items to form meaningful products. Towards solving this problem, we 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 e-commerce consisting of 25K listings labeled at word-level. Given 23 labels, we discover label-values with F1 score of 92.2%. To our knowledge, this is the first work to tackle product discovery and show effectiveness of neural architectures on a complex dataset that goes beyond popular datasets for POS tagging and NER.