Sachpazis Costas: Geotechnical Engineering: A student's Perspective Introduction
Affect Enriched Word Embeddings for News IR
1. Affect Enriched Word Embeddings for News
Information Retrieval
Tommaso Teofili, Niyati Chhaya
Adobe
{teofili,nchhaya}@adobe.com
July 25, 2019
2. Detecting affect in text
• Affect refers to the experience of feelings, emotions,
personality and moods
• An important aspect to capture for natural language
understanding
• Applications:
• Analysing consumer behaviour
• Opinion mining
• Sentiment analysis
3. Affect enriched word embedding models (S Khosla, N
Chhaya, K Chawla @ COLING 2018)
• Affect-enriched word distributions trained on Warriner’s
lexicon coupled with ”plain” word embeddings
• FPF Enron dataset
• Beat SotA in
• Intrinsic word similarity
• Sentiment analysis
• Personality detection
• Frustration detection in interpersonal communication
4. Aff2vec embeddings in News IR - why ?
• More resilient to synonym / antonym issue
• Analysis of affect score in news datasets
Dataset affect scoring
Dataset formality politeness frustration
NYT 0.7087 0.6291 0.6248
WP 0.7788 0.7456 0.6510
CACM 0.3619 0.1229 0.3511
ClueWeb09 0.4319 0.2708 0.6216
Table: Mean affect scores on some common IR datasets