Abstract
In this paper, we aim to learn associations between visual attributes of fonts and the verbal context of the texts they are typically applied to. Compared to related work leveraging the surrounding visual context, we choose to focus only on the input text as this can enable new applications for which the text is the only visual element in the document. We introduce a new dataset, containing examples of different topics in social media posts and ads, labeled through crowd-sourcing. Due to the subjective nature of the task, multiple fonts might be perceived as acceptable for an input text, which makes this problem challenging. To this end, we investigate different end-to-end models to learn label distributions on crowd-sourced data and capture inter-subjectivity across all annotations.
Introduction
Choosing the right font is a cornerstone of graphic design, yet it remains a notoriously subjective and time-consuming task. While professional designers often rely on intuition to match a typeface to a message, casual users are often left scrolling through endless menus. Existing font recommendation tools typically focus on visual similarity—helping users find a font that looks like one they already like. However, they often ignore the verbal context: the actual meaning and emotional weight of the words. There is a significant gap in automated systems that can look at a sentence like "Big Sale Today!" versus "Deepest Sympathies" and understand that they require fundamentally different typographic "personalities."
Methodology: Label Distribution Learning
The authors frame font selection not as a simple classification task with a single "right" answer, but as a ranking and distribution problem. Because font choice is inherently subjective, the research utilizes Label Distribution Learning (LDL) to capture the degree of "congruency" for multiple fonts across different users.
Four deep learning architectures were evaluated:
- GloVe-BiLSTM: A standard NLP approach using pre-trained word embeddings.
- NRC Model: An emotion-focused model using lexicons that map words to specific affect intensities (e.g., joy, anticipation).
- BERT: A transformer-based model providing deep contextual representations of the text.
- Emoji Model (DeepMoji): A novel approach using a model pre-trained on 1.2 billion tweets to predict emojis. The rationale is that the emotional nuance required to select the right emoji for a tweet is analogous to the nuance needed to select the right font for a sentence.
The Short Text Font Dataset
The researchers introduced the Short Text Font Dataset, consisting of 1,309 short text instances sourced from Adobe Spark (now Adobe Express), including ads, social media posts, and motivational quotes. Ten representative fonts (F0-F9) were selected by an expert to cover a wide range of styles, from formal serifs to casual scripts and bold display fonts. Each text was labeled by 6–8 crowd-workers who ranked their top three font choices.
Key Results
The BERT and Emoji models emerged as the top performers, significantly outperforming the majority baseline. As shown in the results table below, BERT achieved the highest accuracy, with a Font Recall (Top 3) of 33.54% and an F-score (Top 1) of 26.97%.
Notably, the Emoji Model performed nearly as well as BERT. This suggests that models focused specifically on emotional "vibe" can be highly effective for typographic tasks, even without the massive scale of general-purpose language models.
Discussion: Subjectivity and Emotional Personas
The study confirmed that users consistently associate specific fonts with specific "tones." For example, "formal" fonts (F0, F2) are preferred for business contexts, while "script" fonts (F4, F9) are favored for emotional or personal messages.
One of the most interesting findings was the high correlation between fonts and emojis. Script fonts were highly correlated with "Heart" and "Love" emojis, while playful fonts showed negative correlations with emojis representing discomfort. By using label distributions instead of "hard" labels, the models learn that multiple fonts can be highly acceptable for a given text, while others are completely inappropriate.
Limitations and Future Work
The current study is limited to a palette of 10 representative fonts. In real-world design, the selection is vast. Future work could leverage font similarity to map these categories to larger libraries. Additionally, while this paper intentionally focuses on verbal context, integrating visual elements like background images and colors remains a logical next step for holistic design automation.
Summary updated