Mixed Feelings, Single Polarity

Key points

• We test where this works and where it fails, using SemEval-2018 Affect in Tweets (English, Arabic, Spanish), Affective Text aligned with EmoBank, raw GoEmotions annotations, naturalistic hashtag corpora, and a controlled annotation study.1
• Fixed projection works well when a text carries labels of one valence.1
• It fails on a subset identifiable from the labels alone: tweets whose emotion labels point in opposite directions.1
• Annotators produce such configurations routinely: 16.3% of multilabel GoEmotions judgments combine emotions from the dataset's own positive and negative parents.1

Sources (1)

• [1]Mixed Feelings, Single Polarity
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 11:25 PM

  Emotion datasets are routinely mapped onto positive and negative valence, by assigning each emotion category one polarity.

  We test where this works and where it fails, using SemEval-2018 Affect in Tweets (English, Arabic, Spanish), Affective Text aligned with EmoBank, raw GoEmotions annotations, naturalistic hashtag corpora, and a controlled annotation study.

Extractive summary: sentences quoted from the sources.