Towards Open-World Product Attribute Mining: A Lightly-Supervised Approach

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12223–12239 July 9-14, 2023 ©2023 Association for Computational Linguistics Towards Open-World Product Attribute Mining: A Lightly-Supervised Approach Liyan Xu1,2∗ Chenwei Zhang3 Xian Li3 Jingbo Shang4 Jinho D. Choi2 1Pattern Recognition Center, WeChat AI 2Emory University 3Amazon.com liyan.xu@emory.edu4University of California, San Diego {cwzhang,xianlee}@amazon.com Abstract We present a new task setting for attribute mining on e-commerce products, serving as a practical solution to extract open-world at- tributes without extensive human intervention. Our supervision comes from a high-quality seed attribute set bootstrapped from existing resources, and we aim to expand the attribute vocabulary of existing seed types, and also to discover any new attribute types automat- ically. A new dataset is created to support our setting, and our approach Amacer is pro- posed specifically to tackle the limited supervi- sion. Especially, given that no direct super- vision is available for those unseen new at- tributes, our novel formulation exploits self- supervised heuristic and unsupervised latent attributes, which attains implicit semantic sig- nals as additional supervision by leveraging product context. Experiments suggest that our approach surpasses various baselines by 12 F1, expanding attributes of existing types signifi- cantly by up to 12 times, and discovering values from 39% new types. Our data and code can be found at https://github.com/lxucs/woam. 1 Introduction Attribute mining (or product attribute extraction) is to extract values of various attribute types (e.g. colors, flavors) from e-commerce product descrip- tion, which is a foundational piece for product un- derstanding in online shopping services, enabling better search and recommendation experience. Within this task regime, different settings have been studied. Most pioneer works deem it as a closed-world setting, where models are trained to identify a fixed set of pre-defined attribute types (Ghani et al., 2006; Putthividhya and Hu, 2011; Zheng et al., 2018), similar to the standard named entity recognition (NER). Recent works start to step up towards the open-world aspect that supports ∗Primary work done as an applied scientist intern at Amazon, before joining WeChat AI. Ingredient: lemon, chamomile, … Flavor Profile: sweet, sour, fruity, … Feature: organic, … Titles: Black Tea with Orange & Sweet Clove - 2.67 Ounces, 30 Sachets •Bullet Point Description: This loose leaf tea is selected from highest gardens of Himalayan mountains and Tibet at the peak of the growing season, giving a subtle muscatel flavor. • No artificial color. Stored Correctly will last one year or more. Sanitize Noisy Attributes (existing resources from retailers) Amacer Light Supervision: Seed Attribute Set Input: Product Context • Orange • Sweet Clove • subtle muscatel flavor • No artificial color • Tibet • Himalayan mountains • one year or more Seed Attribute Types (expanded up to 12 times) New Attribute Types (discovered 39% new types) Figure 1: Illustration of our task setting on one prod- uct: given light supervision from seed attributes, our approach Amacer aims to expand attribute vocabulary of seed types, and to also discover values of any new types …

Content truncated to 500 words