Labels to Inference: A Continuous Sentiment Pipeline
summary
Company rebuilds sentiment as a continuous pipeline that labels, trains and deploys daily using Mechanical Turk, Elasticsearch, Kafka/Storm, and Amazon SageMaker with Lambda and CodePipeline, training Keras/TensorFlow LSTM models from five-worker labels aggregated by expectation‑maximization. Equal sampling across streams and worker feedback produce about 25–30% accuracy improvement, ~50% of crowd workers return 10+ days, and customers cut manual sentiment overrides by 60–90%. Deployment uses canary rollouts and model artifacts in S3 so models update reliably without manual intervention.
AI-written summary · about
chapters (10)
- 0:00Introduction to Signal Labs
- 1:16What Signal Labs does
- 2:59Why rebuild the sentiment pipeline
- 4:44Redefining sentiment analysis
- 7:51Labeling data at scale
- 12:08Building the ML model
- 13:52Deploying with SageMaker
- 16:21Results and improvements
- 18:27Customer feedback and trials
- 19:29Q&A session
related talks
43mSentiment Analysis in Action: Building Your Real-time Pipeline
22mFrom experiment to production - a journey of a machine learning model
47mAI vs. Karen: Real-Time Customer Sentiment Detection at Scale
30mOperationalizing Machine Learning
43mMachine Learning Ops: DevOps for Data Science
38mBuilding robust machine learning systems
29mMachine learning in production
18mBeyond Inference: Bringing ML into Production