Unpaired Sentiment-to-Sentiment Translation: Using Reinforcement Learning
summary
Produces semi‑supervised parallel corpora from unpaired reviews by running a cycle reinforcement learning loop of a neutralization and emotionalization modules (self‑attention classifier + LSTM seq2seq) and training with a policy gradient reward that combines sentiment confidence (derived F1) and BLEU. Evaluated on Yelp (≈400k items, 10k train / 3k val / rest test) and Amazon (≈230k train), it raises the geometric G‑score from ~1.17 (cross‑alignment) / 1.64 (MDAL) to 22.46 G‑score on Yelp, preserving content well though intensity estimation and some edge cases (multiple emotions, negations) remain imperfect.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:23Unpaired SentimenttoSentiment Translation
- 9:04Preserving Context
- 10:37Contributions
- 12:58Experimental Results
- 16:24Rewards
- 17:57Previous work
- 19:09Evaluation
- 21:07Applications
- 22:03Conclusion
- 22:43Improvements
- 23:30Questions
related talks
01h 23mBeyond Sentiment Emotion Mining with Python and machine learning
39mSentiment Analysis, its techniques and applications
01h 31mAn Introduction to Sentiment Analysis of...
43mGetting Sentimental: exploration of iOS sentiment analysis
22mSentiment Analysis on Sustainability Reporting
37mMultilingual sentiment analysis & emotion detection based on social media data
39mDeep Learning vs. Conventional Machine Learning
21mLabels to Inference: A Continuous Sentiment Pipeline