Some features I like
(in a long line of LiWC-like lexicons) Chenhao Tan's list of hedging phrases
(some as regular expressions [README, list itself]
- Part-of-speech n-grams
- Language models on the most frequent words only
- Distinctiveness
- Language models on the content words
- Distributional similarity
... and one feature that I both like and drives me crazy: token length
What does this mean in the age of deep learning, where we don't need to worry about features anymore?
- BERT vs hand features, controversy paper
- Word embeddings
- BERT - word pieces!
- Language modeling
not hidden