Enhancing Product Safety in E-Commerce with NLP Kishaloy HalderyJosip Krapac Dmitry GoryunovyAnthony Brewy Matti Lyra Alsida DizdariyWilliam Gillett Adrien Renahy Sinan Tang
2025-05-06
0
0
277.43KB
7 页
10玖币
侵权投诉
Enhancing Product Safety in E-Commerce with NLP
Kishaloy Halder†∗ Josip Krapac Dmitry Goryunov†Anthony Brew†
Matti Lyra Alsida Dizdari†William Gillett Adrien Renahy Sinan Tang
†Work done while at Zalando
kishaloh@amazon.com josip.krapac@zalando.de d.f.goryunov@gmail.com
atbrew@gmail.com matti.lyra@zalando.de alsida.dizdari@gmail.com
william.gillett@zalando.de adrien@zalando.fr sinan.tang@zalando.de
Abstract
Ensuring safety of the products offered to the customers is of paramount importance to any e-
commerce platform. Despite stringent quality and safety checking of products listed on these
platforms, occasionally customers might receive a product that can pose a safety issue arising out
of its use. In this paper, we present an innovative mechanism of how a large scale multinational
e-commerce platform, Zalando, uses Natural Language Processing techniques to assist timely
investigation of the potentially unsafe products mined directly from customer written claims
in unstructured plain text. We systematically describe the types of safety issues that concern
Zalando customers. We demonstrate how we map this core business problem into a supervised
text classification problem with highly imbalanced, noisy, multilingual data in a AI-in-the-loop
setup with a focus on Key Performance Indicator (KPI) driven evaluation. Finally, we present
detailed ablation studies to show a comprehensive comparison between different classification
techniques. We conclude the work with how this NLP model was deployed.
1 Introduction
Keeping the platform safe for all customers is one of the top priorities for many (if not all) large e-
commerce businesses (Ullrich, 2019; Satheeshkumar et al., 2021). As a popular e-commerce platform for
fashion (clothes, shoes, accessories) and beauty products (cosmetics), Zalando1observes a large number
of customer returns for size and fit, or quality related reasons. While such returns are not surprising as
fashion is often manifested from personal preferences, occasionally customers report safety issues with
products such as broken heel or sharp protruding edges. We consider such occurrences as Product Safety
(PS) cases in this work. To investigate the gravity of individual cases with due diligence, the customers
are required to submit a description of their experience with the product in plain text, along with an
optional image of the product while reporting such a case.
We break down the overall PS investigation process in three stages as depicted in Figure 1. The process
starts from the customer submitted comments. In the first stage, all the customer claims are examined
by the Front-line Customer Care (FCC) agents. They escalate some of these cases as potential PS issues
which are usually rare and thus much smaller in volume compared to that of all the customer comments.
In the next stage of investigation, a team of PS agents (PS team hereafter) who have gone through
specialized training to spot PS related issues from the customer comments, examine the escalated cases.
They further filter escalated cases after close investigation. A tiny fraction of the escalated cases, which
they perceive to be truly related to safety, are forwarded to the next stage of the process i.e., laboratory
investigation. Based on the findings the PS team takes necessary actions against the unsafe products e.g.,
take down product listing and escalate the case to the manufacturer. Finally, they also record the outcome
of the investigations and whether each case is indeed about safety.
As the escalations traverse through the stages, the depth of scrutiny of individual cases increases
substantially, and so does the operational cost. To make the optimal use of the investigative resources, it
is imperative that at each stage, only those cases with high likelihood to be PS related, are escalated.
∗Work done prior to joining Amazon
1zalando.de
arXiv:2210.14363v1 [cs.CL] 25 Oct 2022
Figure 1: Product Safety investigation workflow in Zalando. A Legacy, end-to-end manual Workflow is
shown on the LHS. The proposed and newly implemented AI-in-the-loop workflow is on the RHS.
The rarity of true PS issues creates a unique challenge in keeping the platform safe. A diverse set of
issues can be misplaced as being generic quality issues by the regular FCC agents, risking thus missing
true PS cases.
In this work, we present a Natural Language Processing (NLP) based framework to address this core
problem of flagging all potential Product Safety cases without overwhelming the limited capacity of PS
experts with irrelevant cases. To summarize our contributions are the following:
• We present a systematic overview for the problem of PS investigation in a large-scale ecommerce
company.
• We identify key signals in the investigation workflow to map this business problem into a data-driven
modeling task.
• We demonstrate how modern NLP techniques can be used in conjunction with strategic training
procedures to overcome the challenges in a production such as class imbalance, noise, and multilin-
guality.
2 Problem Description
PS cases are a subset of all customer claims where the comment indicates a broader safety issue from
using the product. The PS team (second stage in Figure 1) categorizes all such claims into the four
categories presented in Table 1.
In this work we regard the classifications obtained from the PS team as the gold-standard, ground-truth
data. The first three case types i.e., “Allergic Reaction”, “Chemical Smell”, and “Injury” are formally
termed as PS cases, and are further investigated following lab-protocols. A significant section of all the
comments is regarding generic quality-related issues, and is marked as “Not Product Safety” by the PS
specialist team.
2.1 Practical Challenges
This core business problem of identifying PS cases from the customer claims early comes with a unique
set of challenges, especially when applying the process to the scale of Zalando’s large customer base
(46.3 million active customers in as of Q3 2021).
摘要：
展开>>
收起<<
EnhancingProductSafetyinE-CommercewithNLPKishaloyHalderyJosipKrapacDmitryGoryunovyAnthonyBrewyMattiLyraAlsidaDizdariyWilliamGillettAdrienRenahySinanTangyWorkdonewhileatZalandokishaloh@amazon.comjosip.krapac@zalando.ded.f.goryunov@gmail.comatbrew@gmail.commatti.lyra@zalando.dealsida.dizdari@gmail.co...
声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！
分类：图书资源
价格：10玖币
属性：7 页
大小：277.43KB
格式：PDF
时间：2025-05-06
作者详情
-
IMU2CLIP MULTIMODAL CONTRASTIVE LEARNING FOR IMU MOTION SENSORS FROM EGOCENTRIC VIDEOS AND TEXT NARRATIONS Seungwhan Moon Andrea Madotto Zhaojiang Lin Alireza Dirafzoon Aparajita Saraf10 玖币0人下载
-
Improving Visual-Semantic Embedding with Adaptive Pooling and Optimization Objective Zijian Zhang1 Chang Shu23 Ya Xiao1 Yuan Shen1 Di Zhu1 Jing Xiao210 玖币0人下载
渝公网安备50010702506394
