Shoe Size Resolution in Search Qeries and Product Listings using Knowledge Graphs Petar Ristoski Aritra Mandal Simon Becker pristoski@ebay.com arimandal@ebay.com sibecker@ebay.com eBay Inc. San Jose, USA Anu Mandalam Ethan Hart Sanjika Hewavitharana amandalam@ebay.com ejhart@ebay.com shewavitharana@ebay.com eBay Inc. San Jose, USA Zhe Wu Qunzhi Zhou zwu1@ebay.com qunzhou@ebay.com eBay Inc. San Jose, USA ABSTRACT The Fashion domain is one of the most proftable domains in most of the e-commerce shops, shoes being one of the top-selling categories within this domain. When shopping for shoes, one of the most im- portant aspects for the buyers is the shoe size. Shoe size charts difer between diferent brands, geographical regions, genders and age groups. Not providing some of these details, as a buyer or a seller, could lead to a query intent to inventory mismatch and reduced or wrong search results. Furthermore, buying the wrong shoe size is one of the top reasons for product returns, which causes shipping delays and loss in revenue. To address this issue, we propose an ap- proach for shoe size resolution and normalization in search queries and product listings using Knowledge Graphs. ACM Reference Format: Petar Ristoski, Aritra Mandal, Simon Becker, Anu Mandalam, Ethan Hart, Sanjika Hewavitharana, Zhe Wu, and Qunzhi Zhou. 2022. Shoe Size Res- olution in Search Queries and Product Listings using Knowledge Graphs. In Proceedings of the 31st ACM International Conference on Information and Knowledge Management (CIKM ’22), October 17–21, 2022, Atlanta, GA, USA. ACM, New York, NY, USA, 2 pages. https://doi.org/10.1145/3511808.3557519 1 INTRODUCTION Query understanding is a fundamental component of every suc- cessful e-commerce platform, bridging the semantic gap between buyers intent and available product inventory. Recent deep learn- ing and representation learning approaches [4, 5] have shown high performance in query understanding and information retrieval. However, such approaches often fall behind when working with input containing numeric values, such as shoe sizes, i.e., most of the embedding approaches consider numeric data as text tokens and usually lead to reduced or less relevant search result set. For example, the search query łNike UK 9 sneakersž could erroneously match a listing title łNike US 9 UK 9.5ž. In another example, the query łNike EU 44 sneakersž will not match product titles like łNike US 11.5 sneakersž, although łEU 44ž and łUS 11.5ž are the same shoe size. Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). CIKM ’22, October 17–21, 2022, Atlanta, GA, USA © 2022 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-9236-5/22/10. https://doi.org/10.1145/3511808.3557519 To address this issue, in this work we use a Knowledge Graph (KG) to model shoe sizes, unambiguously defned by their geograph- ical region, gender or age group, and the brand. Combined with machine learning-based entity linking approach, the shoe size KG allows us to identify shoe size mentions in search queries, product listing titles and product listing aspect-value pairs, which can then be represented with a unique canonical Uniform Resource Identi- fer (URI) in the KG. This allows us to cannibalize shoe size values across diferent shoe size conversion charts, and correctly match the user intent to the listings in the inventory. There have been several attempts to identify and normalize sizes in the Fashion domain [2, 3], mostly focusing on clothes and ftment recommendation. To the best of our knowledge this is the frst work to use a KG to precisely and unambiguously model all shoe size values across diferent shoe size charts. Extensive evaluation on human-labeled datasets, both in English and German, as well as A/B tests with live trafc, show signifcant increase in relevant recall and purchased items. 2 APPROACH Our approach for shoe size resolution and normalization follows a similar design we proposed for unit of measurement resolution in search [7], consisting of 3 main components: (i) knowledge graph, (ii) entity resolution, and (iii) query rewrite. To generate the KG, we consider 4 diferent generic gender- specifc conversion charts, i.e., mens, womens, kids, babies; and 4 diferent geo-specifc charts, i.e., US, EU, 1 UK, AU. Furthermore, some of the shoe brands have their own shoe size charts, which difer from the generic charts. We use 14 proprietary brand-specifc shoe size charts, including some of the most popular brands, like Nike, Gucci, Adidas, etc. We use this information to generate a KG of shoe size entities, where each entity has a canonical representation expressed as a unit of length. This allows us to perform shoe size conversion between geoographical/brand/gender-specifc conver- sion charts. For quick retrieval, we set a convertTo relation between each pair of shoe size nodes that can be converted to each other. The fnal KG contains more than 2,000 entities and more than 25,000 statements, including more than 5,000 convertTo relations. The entity linking component consists of 3 main parts: (i) men- tion detection, (ii) context detection, and (iii) entity retrieval. For mention detection, we use a hybrid neural network NER model [6], 1 EU includes the shoe sizes for the main European markets, which share the same conversion charts 5094