AffectEventKB
AffectEventKB is a knowledge base of events developed by the
NLP group at University of Utah.
The events in the knowledge base were extracted from personal stories which were identified
from web blogs.
Each event is represented using a frame-like tuple
which consists 4 fields:
Agent,
Predicate,
Theme,
PrepPhrase(PP).
Each event is associated with
a polarity label (highest score), and
a distribution of scores over three polarities: positive,
negative, and neutral.
You could download the knowledge base using the following link:
- Method-1:
Use the polarity label with highest probability score. This method gives around 70% precision
based on a random set of events with manual annotations.
- Method-2:
Only use affective events whose polarity probability scores are larger than some thresholds.
Based on our estimation, using 0.5 threshold can select a set of affective events
with over 90% precision for positive, and over 80% precision for negative.
If you use this resource in your project,
please cite this paper;