Web-Scale Language-Independent Cataloging of Noisy Product Listings for E-Commerce

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 969–979, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Web-Scale Language-Independent Cataloging of Noisy Product Listings for E-Commerce Pradipto Das, Yandi Xia, Aaron Levine, Giuseppe Di Fabbrizio, and Ankur Datta Rakuten Institute of Technology, Boston, MA, 02110 - USA {pradipto.das, ts-yandi.xia, aaron.levine}@rakuten.com {giuseppe.difabbrizio, ankur.datta}@rakuten.com Abstract The cataloging of product listings through taxonomy categorization is a fundamental problem for any e-commerce marketplace, with applications ranging from personal- ized search recommendations to query un- derstanding. However, manual and rule based approaches to categorization are not scalable. In this paper, we compare sev- eral classifiers for categorizing listings in both English and Japanese product cata- logs. We show empirically that a combina- tion of words from product titles, naviga- tional breadcrumbs, and list prices, when available, improves results significantly. We outline a novel method using corre- spondence topic models and a lightweight manual process to reduce noise from mis- labeled data in the training set. We con- trast linear models, gradient boosted trees (GBTs) and convolutional neural networks (CNNs), and show that GBTs and CNNs yield the highest gains in error reduc- tion. Finally, we show GBTs applied in a language-agnostic way on a large- scale Japanese e-commerce dataset have improved taxonomy categorization perfor- mance over current state-of-the-art based on deep belief network models. 1 Introduction Web-scale e-commerce catalogs are typically ex- posed to potential buyers using a taxonomy cat- egorization approach where each product is cate- gorized by a label from the taxonomy tree. Most e-commerce search engines use taxonomy labels to optimize query results and match relevant list- ings to users’ preferences (Ganti et al., 2010). To illustrate the general concept, consider Fig. 1. A merchant pushes new men’s clothing listings to an online catalog infrastructure, which then orga- nizes the listings into a taxonomy tree. When a user searches for a denim brand, “DSquared2”, the search engine first has to understand that the user is searching for items in the “Jeans” category. Then, if the specific items cannot be found in the inventory, other relevant items in the “Jeans” cat- egory are returned in the search results to encour- age the user to browse further. However, achiev- ing good product categorization for e-commerce market-places is challenging. Commercial product taxonomies are organized in tree structures three to ten levels deep, with thousands of leaf nodes (Sun et al., 2014; Shen et al., 2012b; Pyo et al., 2016; McAuley et al., 2015). Unavoidable human errors creep in while upload- ing data using such large taxonomies, contributing to mis-labeled listing noise in the data set. Even EBay, where merchants have a unified taxonomy, reported a 15% error rate in categorization (Shen et al., 2012b). Furthermore, most e-commerce companies receive millions of new listings per month from hundreds of merchants composed of wildly different formats, descriptions, prices and meta-data for the same products. For instance, the two listings, “University of Alabama all-cotton …

Content truncated to 500 words