2017年NLP技術在eBay的技術實踐.pdf

1、NLP 技術在 eBay 的技術實踐 Bachelor+Master ZJU Ph.D UC Riverside 10 Year Industrial ExperienceImage and video processingIntelligent video surveillance Video stabilization License plate detection,recognitionData scientist/tech lead at eBay About MeAbout eBay Founded by Pierre Omidyar in 1995 Has Marketplace,

2、Stubhub and eBay Classified Marketplace has sites in 25 countries 169 million active buyer$20B revenue in Q1 2017 One of the pioneer companies to use big data SearchClassifiersName Entity RecognitionMachine TranslationNLP at eBayNERApplications:-Enrich user filled in attributes for the item help MT,

3、search-Discover new values for attribute dictionaryConditionBrandModelStoragecapacityContractColorTypeClassifiers Classify items to categories category recommendation Classify items to product/accessory/bundle Classify items to product types Classify reviews to real product reviews/others Problem Cl

4、assify item to leaf category(16k categories)If doing randomly,accuracy 0.00625%Progress over the past 10 years 2006 2011:Histogram+Naive Bayesian 50%top1 accuracy 2011 2015H1:KNN based on search engine():73%top1 accuracy 2015H2:KNN+SLM:81%top1 accuracy 2017:DNN:90%top 1 accuracy Example:Item Category RecommendationExample:Product/Accessory/Bundle ClassifiervvvBundleAccessoryProductred leather case for iPhone 7A4P*S0=*S1=AS2=4S3=PT=4S=s0s1 sT-1P(Si+1=4|Si=A)P(Si+1=4|Si=A,Si-1=*)Main idea:build SLM on symbolized titlesSymbolized title:Q&A