AdityaKane2001
commented
2 years ago - WASSA 2017 shared task: http://saifmohammad.com/WebPages/EmotionIntensity-SharedTask.html
- DailyDialog (an emotion classification dataset): https://aclanthology.org/I17-1099.pdf
- SemEval-2017, SemEval-2016, SemEval-2015, SemEval-2014, SemEval-2013 Shared Tasks on Sentiment Analysis in Twitter. 11 emotions + neutral class
Architectures for emotion classification (from Papers with Code):
- Improving Multi-label Emotion Classification by Integrating both General and Domain Knowledge
- They use a CNN on the embeddings to find domain-specific patterns in text (tweets in this case). Then they apply BERT on the features extracted from the CNN. They hypothesize the CNN helps in capturing domain specific patterns whereas BERT helps to gain a more general understanding of the text. See Section 3.3 in the paper.
- EmoGraph: Capturing Emotion Correlations using Graph Networks
- This paper aims to mitigate the fuzziness in emotions. They use BERT to generate representations of input text and "apply a graph" on the same. From the paper:
-
The GCN takes the features of each emotion node as inputs and applies a convolution over neighboring nodes to generate the classifier.
- See paper for more details.
Point 4 from Google Doc.