CAO: A Fully Automatic Emoticon Analysis System

Michal Ptaszynski, Jacek Maciejewski, Pawel Dybala, Rafal Rzepka, Kenji ArakiAAAI 2010 · DOI

Abstract

In digital text communication, emoticons serve as vital non-verbal cues, yet their structural complexity—particularly in Eastern-style kaomoji—poses significant challenges for automated analysis. This paper presents CAO (emotiCon Analysis and decOding of affective information), a system that decomposes emoticons into semantic areas based on kinesics. By treating emoticons as generative structures rather than static strings, CAO achieves a coverage of over 3 million combinations. Evaluation on a massive blog corpus demonstrates a 97.1% F-score in extraction and up to 97% accuracy in mapping emotions to a 2D affect space, significantly outperforming existing n-gram and kernel-based methods.

Introduction

Text-based communication often lacks the nuance of face-to-face interaction, such as facial expressions and gestures. To bridge this gap, users employ emoticons. While Western emoticons (e.g., :)) are relatively simple, Eastern emoticons (e.g., (^_^)) are highly complex, utilizing a vast array of characters to represent specific facial features. Traditional analysis methods rely on fixed databases, which fail to keep pace with user creativity. The CAO system addresses this by analyzing the internal structure of emoticons, treating them as a language with its own "grammar."

The CAO Methodology

The core innovation of CAO is its use of "kinemes"—the smallest meaningful units of facial expression. Instead of matching whole strings, the system divides an emoticon into semantic areas: the Mouth [M], Left/Right Eyes [$E_L, E_R$], Borders [$B_1, B_2$], and "Additional Areas" [$S_1$–$S_4$] representing features like blush marks or hands.

The multi-stage extraction process used by CAO to identify and decode emoticons.

The system employs a multi-stage extraction logic. It first attempts to match an input against a raw database. If no match is found, it looks for "eye-mouth-eye" triplets ($E_L M E_R$). If that also fails, it uses regular expressions to identify individual semantic components. This tiered approach allows CAO to recognize over 3 million unique emoticon combinations from a base dictionary of only ~10,000 entries.

Emotion Mapping and Scoring

CAO maps identified components to ten specific emotion types (e.g., joy, anger, sorrow, fear). These emotions are further categorized within a 2D affect space defined by valence (positive/negative) and activation (active/inactive).

The 2D affect space used to categorize emotions based on valence and activation.

To determine the final emotion of an emoticon, the researchers tested three scoring methods:

  1. Occurrence: Simple presence of an emotion-linked component.
  2. Frequency: Normalized by the total number of elements.
  3. Unique Frequency: Normalized by the number of unique elements.

Performance and Results

The system was evaluated using the Yacis Blog Corpus, containing 350 million Japanese sentences. CAO demonstrated exceptional robustness, achieving a 99.75% F-score on training data and maintaining a 97.1% F-score in "wild" blog text.

In terms of emotion estimation, the "Unique Frequency" method proved superior. On isolated emoticons, CAO reached 93.5% accuracy for specific emotion types. When evaluated on the broader 2D affect space, accuracy rose to nearly 97%. Even at the sentence level—where emoticons were the sole source of affective data—the system achieved 80.2% accuracy for specific emotions and 94.6% for general affect groups.

Discussion and Future Work

CAO's generative approach represents a significant leap over previous lookup-table methods. By understanding the "kinemes" of an emoticon, the system gains a human-like ability to interpret novel expressions. However, limitations remain, particularly regarding contextual blindness; an emoticon's meaning can shift based on the surrounding text (e.g., sarcasm). Future iterations of CAO could integrate full-sentence sentiment analysis or be applied to real-world problems like cyber-bullying detection and cross-generational communication tools.

Summary updated

BibTeX
@inproceedings{ptaszynski2010cao,
  author = {Michal Ptaszynski and Jacek Maciejewski and Pawel Dybala and Rafal Rzepka and Kenji Araki},
  title = {{CAO: A Fully Automatic Emoticon Analysis System}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7715},
}
Figure