Target-Based Sentiment Analysis with T5
summary
Claims a 20% leaderboard gain using T5 repurposed as a question‑answering model, fine‑tuned with prefix prompts and techniques like layer freezing and cosine LR schedules. Works on domain factual articles with per‑entity sentiment labels (positive neutral negative), yields about 82–83% F1 with T5 small after heavy manual data cleaning and annotation, and the model is deployed in production with no complaint emails so far.
AI-written summary · about
chapters (10)
- 0:00T5 and target-based sentiment
- 4:12Data quality in ML
- 7:46Sentiment analysis types
- 11:39Targeted sentiment on facts
- 13:53Choosing T5
- 16:45T5 self-attention
- 17:31Pre-training and fine-tuning T5
- 18:33Results and performance
- 20:22Testing and production feedback
- 21:28Q&A
related talks
22mSentiment Analysis on Sustainability Reporting
39mSentiment Analysis, its techniques and applications
01h 31mAn Introduction to Sentiment Analysis of...
01h 05mSocial Text, Sentiment, and Tone Analysis
21mLabels to Inference: A Continuous Sentiment Pipeline
32mUnpaired Sentiment-to-Sentiment Translation: Using Reinforcement Learning
29mLies, damned lies and large language models
47mAI vs. Karen: Real-Time Customer Sentiment Detection at Scale