European Quant Marketing Workshop

May 5, 2025

Topic
Natural Affect DEtection (NADE): Using Emojis to Infer Emotions from Text

Speaker
Christian Hotz-Behofsits, Assistant Professor at WU Vienna

Date and Time
May 5, 2025 @ 15:00-16:00 CET / 9:00-10:00 ET

Abstract
Emotions are central to consumer communications, and extracting them from user-generated online content is crucial for marketers, as consumer opinions significantly shape brand perceptions, influence purchase decisions, and provide essential insights for marketing analytics. To leverage vast user-generated data, marketers and researchers require advanced text-to-emotion converters. However, existing tools for fine-grained emotion extraction face several limitations: lexica are constrained by their dictionaries, machine learning models by human-annotated training data, and large language models by insufficient validation. As a result, marketing research still often relies on basic sentiment detection instead of extracting more nuanced emotions from text.

We propose Nade (Natural Affect DEtection)—a novel, text-to-emoji-to-emotion converter that addresses these shortcomings through a two-stage architecture. In Stage I, an embedding-based multi-label classifier predicts 151 emojis, all selected from the Smileys & Emotion group defined by the Unicode standard. These emojis serve as freely available, user-labeled emotion proxies. The classifier was trained on 110 million generic English-language social media posts from Twitter/X. In Stage II, we use gradient boosting trained on 5,975 emotion-intensity-labeled words from the NRC lexicon to map emoji outputs to emotional intensity scores for eight core emotions based on Plutchik's theory. Plutchik's model is widely used in marketing and psychology and offers a structured yet flexible hybrid model that combines the benefits of basic emotions and dimensional emotion theories. While we adopt it as a default, Nade is not restricted to a specific emotional theory: we demonstrate its extensibility by adapting it for Valence-Arousal-Dominance (VAD). Using human raters and state-of-the-art converters as benchmarks, we validate Nade, establish the benefits of exploiting emojis, and showcase several marketing applications across diverse social media platforms. Nade scales efficiently, preserves user privacy through local inference, and is freely available as a web app and R/Python packages—offering a validated and accessible approach to emotion extraction in marketing and beyond.

Contact