📄 Affect, Body, Cognition, Demographics, and Emotion: The ABCDE of Text Features for Computational Affective Science

Summary

Research in computational affective and social science often requires text that has already been tagged with features such as emotion-word usage or the author's age. The tools for producing these labels are scattered and difficult to use, especially for scholars without a computer science background. To lower this barrier, the authors release ABCDE, a dataset of more than 400 million text instances. It combines social media posts, blogs, books, and AI-generated material, including human–LLM chats, reasoning traces, preference data, and detection corpora. Each instance comes with 136 pre-computed lexical features grouped into five themes: Affect, Body, Cognition, Demographics, and Emotion. Most features are lexicon-based, for example valence/arousal/dominance, warmth/competence, body part mentions, and thinking verbs. Demographic attributes are inferred from explicit self-disclosure statements using regex patterns. Unlike toolkits such as LIWC or VADER, which only extract features, ABCDE ships the data and the annotations together. This gives researchers a standardized, reproducible starting point for population-level statistical analysis across fields such as psychology, sociology, the digital humanities, and NLP.

Key points

• Data sources:
  • Twitter (TUSC, 2015–2021): 45.2M geolocated tweets.
  • Reddit (Pushshift, 2010–2022): about 78.6M posts.
  • Google Books Ngram English Fiction (1800–2012): 177.1M 5-gram occurrences, covering 1.74M unique 5-grams.
  • Blogs: 34.2M instances.
  • AI-generated text (2022–2025): 68.9M instances.
• Reddit filtering: posts flagged over_18, promoted posts, posts with images or videos, and posts shorter than 5 or longer than 1,000 words were removed.
• Feature set:
  • Emotion and VAD scores, warmth/competence, and anxiety/calmness.
  • Body part mentions from a released anatomical inventory.
  • Cognition terms in categories such as memory, decision making, and creativity, plus tense and pronoun use.
• Demographic extraction: self-disclosure regex patterns label age, occupation, gender, country, city, and religion, for example from statements of the form "I work as" a given occupation.
• Availability: the dataset is at hf.co/datasets/jpwahle/abcde and the code at github.com/jpwahle/abcde.