AffectVec is a new word emotion database with the following features:
Consider the word "prank". This word may be associated with joy, but also fear, suspense, etc. AffectVec provides scores that quantify the association of words with over 200 different emotion categories, while most other emotion databases only consider 4-8 emotions. Additionally, in our experiments, we show that AffectVec has a higher correlation with human judgments than other emotion resources and that it can be used for sentence or text-level emotion prediction.
AffectVec is available as a tab-separated values (TSV) file, suitable for importing into various applications, or as a word2vec-style word vector file, suitable for use into machine learning systems.
When using such resources, please first refer to Saif Mohammad's detailed discussion Practical and Ethical Considerations in the Effective use of Emotion and Sentiment Lexicons.
We have also induced lower-quality emotion lexicons for over 350 languages. These cover 8 emotions and typically describe a few thousand word forms per language. The quality is lower than for the main English AffectVec resource.
Due to the special nature of the data used to induce the lexicon, as well as the inherent risks of using emotion prediction in downstream applications, great care must be taken when using these sorts of resources. Please first consult our paper referenced below for details about the nature of the data, along with Saif Mohammad's detailed discussion in Practical and Ethical Considerations in the Effective use of Emotion and Sentiment Lexicons.
Cross-Lingual Emotion Lexicon Induction using Representation Alignment in Low-Resource Settings BibTeX
Arun Ramachandran, Gerard de Melo (2020)
In: Proc. COLING 2020.
We also provide the following additional resources:
For more information about the data and method, please consult our publication:
What Sparks Joy: The AffectVec Emotion Database BibTeX
Shahab Raji, Gerard de Melo (2020)
In: Proc. The Web Conference 2020. ACM.