PromotionGo at SemEval-2025 Task 11: A Feature-Centric Framework for Cross-Lingual Multi-Emotion Detection in Short Texts

Imagine you are trying to understand the mood of a room full of people speaking 28 different languages. Some are whispering, some are shouting, and many are feeling a mix of emotions at once—like being both happy and surprised. This is the challenge the PromotionGo team tackled for the SemEval-2025 Task 11 competition. Their goal was to build a computer system that can read short text messages in these many languages and correctly identify which emotions are present, even when multiple emotions are mixed together.

Here is how they did it, explained through simple analogies:

1. The Core Problem: One Label Isn't Enough

Traditional emotion detectors are like a strict teacher who forces a student to pick only one feeling for the whole day: "Are you happy, sad, or angry?" But real life is messier. You can be joyful about a promotion but scared about the new responsibility. The team's system is designed to handle this "mixed bag" of feelings, recognizing that a single sentence can hold several emotions at once.

2. The Three-Step Recipe

The team built a "feature-centric framework," which is just a fancy way of saying they created a modular kitchen where they could swap out ingredients to see what made the best dish for each language. Their recipe had three main steps:

Step A: Translating Words into Numbers (Document Representation)

Computers can't read words; they only understand numbers. The team tried three different ways to turn text into numbers:

• The "Word Count" Method (TF-IDF): Imagine you are counting how many times specific words appear in a text, but you ignore common words like "the" or "and" because they don't carry much emotion. This is like using a simple tally sheet. The team found this surprisingly effective, especially for languages where they didn't have a lot of training data (low-resource languages). It's a reliable, old-school tool that works well when you don't have a massive library of examples.
• The "Contextual Dictionary" Method (FastText & BPE): This is like giving the computer a dictionary that knows not just what a word means, but how its parts fit together. It can guess the meaning of a word it has never seen before by looking at its smaller pieces (like recognizing "unhappiness" even if it only knows "happy" and "un-").
• The "Deep Understanding" Method (Sentence-BERT): This is the "smartest" tool. It's like a polyglot who reads a sentence and understands the vibe and context, not just the individual words. It can tell the difference between "I'm so happy I could scream!" and "I'm so angry I could scream!" even if the words are similar. However, this tool is heavy and sometimes struggles with languages it wasn't specifically trained on.

The Surprise: For many of the languages in the study, the simple "Word Count" method (TF-IDF) actually worked better than the super-smart "Deep Understanding" method. It turns out that for some languages, counting words is more reliable than trying to guess the deep context.

Step B: Cleaning Up the Mess (Dimensionality Reduction)

Sometimes, the computer gets overwhelmed by too much information. Imagine trying to find a specific book in a library where every single book is stacked in a giant, chaotic pile.
The team used a technique called PCA (Principal Component Analysis). Think of this as a smart librarian who quickly sorts the books, throwing away the duplicates and the ones that don't matter, leaving you with a neat, organized shelf.

• The Result: This didn't necessarily make the computer smarter, but it made it much faster. It shaved off training time, especially for the complex models, without hurting their accuracy.

Step C: The Final Judge (Model Training)

Once the text was converted and cleaned, the team needed a "judge" to decide the emotions. They tried different judges:

• The "Solo Act" (Decision Trees): A simple judge who makes quick decisions based on a flowchart. Fast, but sometimes misses the nuance.
• The "Panel of Judges" (Voting Classifier): A group of different judges (like a Decision Tree, a Random Forest, and a K-Nearest Neighbor) who vote on the answer. This is more stable and reliable than a single judge.
• The "Deep Thinker" (MLP): A neural network that learns complex patterns. This was the champion. It was the best at spotting the subtle, mixed emotions, though it took the longest to "study" (train).

3. The Big Challenges They Found

• The "Imbalance" Problem: In their data, some emotions (like "not angry") appeared way more often than others (like "fear"). It's like a classroom where 90% of the students are quiet, and only 10% are shouting. The computer got really good at identifying the quiet students but struggled to spot the shouters. This led to high accuracy for common emotions but poor accuracy for rare ones.
• The "Unseen Language" Trick: For languages the computer had never seen before (like Oromo), they used a clever hack. They asked a large AI model to find a "cousin" language that the computer did know, and then used the dictionary from that cousin language to understand the new one. It's like using a Spanish dictionary to help you guess the meaning of a Portuguese word because the two languages are related.

4. The Bottom Line

The PromotionGo system proved that there is no "one-size-fits-all" solution for emotion detection.

• For some languages, the simple, fast tools (TF-IDF) are the winners.
• For others, the complex, deep tools (Sentence-BERT + MLP) are necessary.
• Speed vs. Smarts: You have to choose. If you need instant results, use the simple tools. If you need the highest possible accuracy and can wait a bit longer for the computer to "think," use the deep learning models.

The team's framework is like a Swiss Army knife: it doesn't rely on just one blade. Instead, it dynamically picks the right tool (representation), cleans the workspace (dimensionality reduction), and selects the best judge (model) for each specific language, ensuring the system works well across a diverse, global audience.