Multilingual sentiment analysis & emotion detection based on social media data
summary
Linkfluence converts emoji signals into a language-agnostic pipeline: a large English-trained emoji predictor outputs emoji-to-sentence-probabilities, then a small per-task decoder maps those 64-d distributions to sentiment or emotions. The production model is an attentive CNN + two-layer decoder that trains on ~512M English tweets, drops epoch time from ~12 days to 44 hours, and yields ~5 ms per inference (vs 50 ms before). Deployment uses TensorFlow Java API with MKL, frozen graphs and graph transforms. Results: ~6–10 point accuracy gain for EN/FR/DE sentiment, ~80% emotion accuracy, but Chinese usage fails (~30pt drop).
AI-written summary · about
chapters (13)
- 0:00Introduction
- 4:20Overview
- 7:21Initial goal
- 8:58Leveraging the lack of manually annotated data
- 11:27What did the team do
- 14:35The idea
- 17:06Benchmark
- 18:46Results
- 21:10Verification
- 26:34Summary
- 27:49Planning and production
- 31:17Optimizations
- 33:41Key takeaways
related talks
39mThe art and science of making computers understand sentiment and emotion
27mSentiment analysis of tweets in Polish language using deep...
32mDecoding Emotions from Art to Memes
43mSentiment Analysis in Action: Building Your Real-time Pipeline
32mUnpaired Sentiment-to-Sentiment Translation: Using Reinforcement Learning
21mEmotion Estimation for Social Robotics
01h 23mBeyond Sentiment Emotion Mining with Python and machine learning
47mAI vs. Karen: Real-Time Customer Sentiment Detection at Scale