“splink” is happy and “phrouth” is scary: Emotion Intensity Analysis for Nonsense Words

Valentino Sabbatino, Enrica Troiano, Antje Schweitzer, Roman KlingerACL 2022

TL;DR

This paper explores whether humans associate emotions with made-up words based on their sound and structure, not just their meaning. They created a dataset of nonsense words with emotion intensity ratings and found that certain phonetic patterns do indeed influence emotional perception, though real-world word meaning still plays a larger role in human judgment.

What Problem Does This Paper Address?

We often associate emotions with words – "joy" feels happy, "fear" feels scary. But how much of this is due to the word's actual meaning, and how much is due to its sound or structure? This is especially interesting when we consider words that don't have a meaning, like nonsense words. Can we predict how people will feel about a made-up word just by how it sounds? This paper investigates this by analyzing how humans assign emotion intensities to nonsense words and whether computational models can learn these associations. This has implications for understanding language acquisition and for applications like branding and marketing where new, meaningless words are coined.

Methods

The researchers used a Best-Worst Scaling (BWS) annotation method to collect human judgments. Participants were presented with quadruplets of words and asked to pick the one they associated most and least with a given emotion (joy, sadness, anger, fear, disgust, surprise). This method is known for eliciting more consistent judgments than simple rating scales.

They then used these annotations to train character-level and phonology-based intensity regressors. These models aim to predict emotion intensity scores based on the input word's structure. They compared models trained on their new nonsense word dataset versus models trained on a large existing lexicon of real words (NRC Emotion Lexicon).

Datasets

  • Newly created dataset: 272 nonsense words and 68 real words, annotated with emotion intensity scores for six basic emotions (joy, sadness, anger, fear, disgust, surprise) using a crowdsourcing study.
  • NRC Emotion Lexicon (EmoLex): A large lexicon of real words with emotion intensity scores, used for comparison and for training models.

Key Insights and Ideas

  • Sound Symbolism is Real (Even for Nonsense): Humans do associate emotions with nonsense words, suggesting that phonetic and orthographic patterns play a role beyond semantic meaning.
  • Front-Loading Matters: The initial sounds of a word seem to have a stronger influence on perceived emotion than later sounds, aligning with previous research on sound symbolism.
  • Phonemes and Emotions: Specific phonemes appear to be linked to certain emotions. For example, /p/ and /s/ at the beginning of a word might increase perceptions of joy, while /sh/ is linked to surprise.
  • The Power of Meaning: While sound matters, the meaning of real words tends to "overrule" phonetic patterns when humans make judgments. This is reflected in models trained on real words performing better on real words, and vice-versa.
  • Model Generalization is Limited: Models trained on nonsense words don't generalize well to real words, and models trained on real words only achieve limited success on nonsense words.

Key Results

  • Phonetic Patterns Identified: The study found specific phonetic patterns associated with different emotions. For instance, 's' as a first phoneme contributes to joy, 'sh' to surprise, and 'p' as a last phoneme leans towards disgust.
  • Model Performance: A regressor trained on real words from the NRC lexicon achieved a Pearson correlation of r=0.17 when tested on nonsense words. Models trained solely on nonsense words showed even lower performance on real words.
  • Reliability: Split-half reliability scores for the nonsense word annotations were consistently lower than for real words, but still indicated strong correlations, suggesting the annotations are reliable.

Limitations and Future Work

  • Corpus Size: The authors acknowledge that their compiled corpus of nonsense words is relatively small. They suggest future work should involve larger datasets and more diverse annotators.
  • Phoneme Interactions: The current analysis focused on single phonemes. Future research could explore interactions between multiple phonemes and the role of less frequent phonemes.
  • Model Complexity: The study used a specific CNN-BiLSTM architecture. Exploring other model architectures, especially those incorporating more sophisticated linguistic features or pre-trained embeddings, could yield better results.
  • Cross-Lingual Differences: The study focused on English. Investigating whether these phonetic-emotion associations hold across different languages would be a valuable extension.

Summary updated

BibTeX
@inproceedings{sabbatino2022splink,
  author = {Valentino Sabbatino and Enrica Troiano and Antje Schweitzer and Roman Klinger},
  title = {{“splink” is happy and “phrouth” is scary: Emotion Intensity Analysis for Nonsense Words}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure