TL;DR
This paper investigates the often-overlooked impact of "euphony" – the pleasantness of sounds in language – on persuasion. By analyzing various datasets and building predictive models, the authors demonstrate that euphonic features like rhyme and alliteration significantly contribute to a sentence's persuasiveness, even outperforming traditional lexical features in some contexts.
What Problem Does This Paper Address?
While computational linguistics has extensively explored how word choice, sentence structure, and semantic meaning contribute to persuasion, the role of phonetics – the sounds of language – has been largely ignored. This is a significant gap because, intuitively, the way something sounds can influence how we perceive it, making it more memorable, agreeable, or impactful. The authors aim to bridge this gap by systematically analyzing and quantifying the effect of phonetic features on various forms of persuasion.
Methods
The core of the paper involves developing and applying a "phonetic scorer" to analyze sentences. This scorer quantifies four phonetic devices:
- Plosives: Sounds created by blocking airflow (e.g., 'p', 't', 'k').
- Alliteration: Repetition of initial consonant sounds.
- Rhyme: Repetition of ending sounds.
- Homogeneity: The degree of phoneme repetition within a sentence.
These phonetic scores are then used as features in machine learning models (specifically Support Vector Machines) to predict whether a sentence is persuasive. The researchers compare the performance of models using only phonetic features, only lexical features (n-grams), and a combination of both. They also conduct within-dataset and cross-dataset experiments to assess the generalizability of their findings.
Datasets
The study utilizes four distinct datasets, each designed to capture different facets of persuasion:
- Twitter: Pairs of persuasive and non-persuasive tweets, controlled for author and URL, with persuasiveness determined by retweet counts.
- Movie: Pairs of memorable movie quotes and their non-memorable counterparts from the same movie, controlled for speaker and sentence length.
- CORPS: Pairs of political speech sentences, where one precedes an audience reaction (e.g., applause), controlled for speaker and sentence length.
- Slogan: Pairs of slogans and similar non-slogans from news articles, controlled for length and semantic similarity.
These datasets are carefully constructed to minimize confounding factors like author popularity or topic, focusing on the linguistic features themselves.
Key Insights and Ideas
- Euphony is a real factor in persuasion: The study provides strong empirical evidence that the pleasantness of sound contributes to a sentence's persuasive power.
- Rhyme and alliteration are key drivers: These devices consistently show higher scores in persuasive sentences across most datasets, suggesting they make messages more memorable and impactful.
- Phonetics can generalize across domains: Phonetic features, due to their inherent linguistic nature, appear to be more robust across different types of persuasive content (e.g., tweets, slogans, speeches) than purely lexical features.
- Homogeneity is surprisingly important: While less intuitive than rhyme or alliteration, phoneme homogeneity also plays a significant role in distinguishing persuasive from non-persuasive sentences.
- Plosives are a mixed bag: The impact of plosives on persuasion is less clear-cut, with some datasets showing higher scores in non-persuasive sentences, challenging initial intuitions.
- Political speeches are unique: The CORPS dataset stands out, with its "non-persuasive" sentences often exhibiting higher phonetic scores than persuasive ones in other datasets, suggesting a different underlying mechanism of persuasion in political discourse.
Key Results
- Within-dataset experiments: Models combining phonetic and n-gram features generally outperformed models using only n-grams, with phonetic features significantly improving accuracy in three out of four datasets. Phonetic features alone, while weaker than n-grams, still performed significantly better than random chance.
- Cross-dataset experiments: Phonetic features alone often outperformed lexical (n-gram) features when predicting across different datasets, highlighting their generality.
- Specific findings:
- Rhyme scores were significantly higher in persuasive sentences for Slogan and Twitter datasets (p < .001).
- Alliteration scores were highly significant across all datasets except CORPS.
- Homogeneity scores were significantly higher in persuasive sentences for most datasets, except CORPS where non-persuasive sentences scored higher.
- Plosive scores showed mixed results, sometimes favoring non-persuasive sentences.
Limitations and Future Work
- Limited phonetic features: The study focused on a specific set of phonetic devices. Future work could explore others like assonance, consonance, and rhythm.
- Dataset specificity: While efforts were made to control for confounding factors, the unique nature of the CORPS dataset suggests that the interaction between phonetic features and persuasion can be domain-dependent.
- Threshold definition: The authors acknowledge the lack of a theoretical basis for defining phonetic thresholds, relying on empirical methods (tongue twisters).
- Future directions: The authors suggest investigating sound symbolism, the connection between context/domain and phonetic impact, and exploring more advanced domain adaptation techniques for cross-dataset prediction.
Summary updated