The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Engineering Production NLP Systems at T-Mobile with Heather Nolis - #600
6 snips
Nov 21, 2022 In this conversation, Heather Nolis, a principal machine learning engineer at T-Mobile, shares her journey from neuroscience to machine learning. She discusses the challenges of deploying real-time deep learning models for customer support, focusing on supervised learning and the creation of customer intent models. Heather highlights the balance between small and large models, technical hurdles in speech recognition, and the importance of data quality. She also looks ahead to the future of NLP, exploring innovative applications in customer service and beyond.
AI Snips
Chapters
Transcript
Episode notes
Unsupervised Learning Falls Short
- Initial attempts with unsupervised learning revealed two major conversation topics: T-Mobile and phones, which lacked business actionability.
- This led to the decision to use supervised learning and create a custom taxonomy for real-time insights.
Building a Custom Taxonomy
- Building a custom taxonomy was challenging, requiring Nolis to advocate against existing enterprise taxonomies.
- One example highlighted how customer language about "network" lacked nuance, rendering a complex sub-taxonomy unnecessary.
Advocating for a Custom Taxonomy
- To advocate for a custom taxonomy, emphasize that modeling is the easier part of data science; defining the business case takes more time.
- Demonstrate the new taxonomy's value by showcasing information it captures that existing ones miss.