Activate 2018

We present here a neural network model, which in contrast to the earlier dictionary based models, uses search logs to tag various parts of user query. Our current work is focused on brand, color, price, item type, gender, age, dimension.

Query tags help the downstream models for search relevancy to perform better when supplemented with these entity tags along with the query. These are helpful in efficiently applying filters and facets a priori while serving search results. Ex: white flower skirts for a 3 year old girl $50. If identified properly, we can switch on brand filters: (A New Day), gender: girl, ge: 3 years, price: around $50, color: white, item type: skirt, and pattern: flowered. This improves the precision of results and aids user experience by providing a query based filtering approach.

The Model Architecture:
Our training data corpus is click-through logs from Target's digital search. Training on textual data presents us with data sparsity challenge. We present here how we tackled the problem of using distributed word representations (w2v, glove), and present observed results on the appropriate datasets for generating the distributed word representations. We use bi-directional LSTMs to model word sequences in a query and add mix in signals from dictionary search results to train the model. We discuss the results obtained from various objective functions that were used as loss metric for back-propagation. We observed good accuracies and modeling of phrases, and disambiguation of words between various tags.