Lessons Learnt Building Domain Specific NLP Pipelines
summary
Describes building a large-scale NLP pipeline for a 3+ billion product catalog by crawling 5000+ retailer sites, detailing architecture, language models, word embeddings, a knowledge graph, and algorithms for classification, entity extraction, document similarity, and e-commerce query understanding. Then explains adapting and fine-tuning that e-commerce pipeline to tax-compliance text with limited data after Indix was acquired by Avalara.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:18About Me
- 2:31Indix - "Google Map" of Products
- 3:29NLP Stack - E-Commerce Domain @ Indix
- 6:11Embeddings Dense Representation of word vectors learned from large unlabelled corpus
- 7:23Embeddings - Useful Properties
- 8:04Embeddings - Training (Word2Vec - Skip Gram Model)
- 10:31Word Embeddings for Document Similarity Are these two products the same?
- 11:19Product Classification - E-Commerce Domain
- 11:48Product Classification - Using FastText
- 12:25Avalara - Part of "every" transaction
- 15:27Tax Rules in New York Taxes for specific type of Clothing
- 16:18Product Classification - Tax Domain
- 20:28Transfer Learning - History
- 21:51Transfer Learning - Using ULMFIT Step 1 - Pre-Training on Source Dataset
- 26:51Tax Domain - Tax Rate Changes
- 27:33Transfer Learning - Why does it work?
- 28:11Pre-Trained Language Models vs Word Embeddings
- 29:07Weak Supervision
- 30:00Data Augmentation & Synthesis
- 30:17NLP Stack - Tax Domain @ Avalara
related talks
55mDemystifying Natural Language Processing across several industry verticals
34mExposing Dark Data in the enterprise with custom NLP
24mRetail Transactions at scale for AI - From PoC to production
14mDriving Industrial Innovation with AI at Scale
06mAdvancements in Natural Language Processing (NLP) and Text Analytics
20mPolyglot AI: The Role of Natural Language Processing (NLP)
43mImplementing AI Document Processing at Scale: Syngenta
22mMachine Learning: The Engine Behind Modern E-commerce Search