Mapping a search query to a set of relevant categories in the product\ntaxonomy is a significant challenge in e-commerce search for two reasons: 1)\nTraining data exhibits severe class imbalance problem due to biased click\nbehavior, and 2) queries with little customer feedback (e.g., tail queries) are\nnot well-represented in the training set, and cause difficulties for query\nunderstanding. To address these problems, we propose a deep learning model,\nDeepCAT, which learns joint word-category representations to enhance the query\nunderstanding process. We believe learning category interactions helps to\nimprove the performance of category mapping on minority classes, tail and torso\nqueries. DeepCAT contains a novel word-category representation model that\ntrains the category representations based on word-category co-occurrences in\nthe training set. The category representation is then leveraged to introduce a\nnew loss function to estimate the category-category co-occurrences for refining\njoint word-category embeddings. To demonstrate our model's effectiveness on\nminority categories and tail queries, we conduct two sets of experiments. The\nresults show that DeepCAT reaches a 10% improvement on minority classes and a\n7.1% improvement on tail queries over a state-of-the-art label embedding model.\nOur findings suggest a promising direction for improving e-commerce search by\nsemantic modeling of taxonomy hierarchies.\n