📄 FunnyBorg at SemEval-2026 Task 1: Humor Generation

Summary

This system description covers the FunnyBorg team's entry to SemEval-2026 Task 1 (MWAHAHA), a shared task on generating computational humor. Most LLM research targets STEM rather than humor, and LLMs struggle with phonetics and pragmatic inference. The task had two parts. Subtask A asked for jokes in English, Spanish and Chinese, built either from a news headline or around two required words. Subtasks B1 and B2 involved GIFs: B1 required a funny caption, and B2 required filling the blanks in a given caption.

The team relied on prompt engineering with a voting step. Their first design was a set of REST microservices: a Runner that orchestrates calls, a Validator, a Translator, an image-recognition service, several prompt-building services, Prompters that call LLMs, and an Evaluator that picks the funniest candidate. They also built a simpler pipeline of TSV-processing steps run in Google Colab. A Gemma model generated the jokes. The entry ranked first in one subtask and second in three others, which suggests careful prompting of a modest model can compete well.

Key points

• Subtasks: A (headline-based or two-word jokes in English, Spanish and Chinese), B1 (GIF captioning) and B2 (fill-in-the-blank GIF captions). The team placed 1st in one subtask and 2nd in three.
• The microservice design had these components:
  • Runner, which orchestrates the other services.
  • Validator, which checks joke length and mandatory words.
  • Translator, which handles EN/ES/ZH.
  • Image recognition, which extracts hints from GIFs.
  • Several Prompt engineerers, each optionally using its own joke-class database.
  • Prompters, which call the LLMs.
  • Humor Evaluator, which selects the best joke.
• The authors discuss security trade-offs of the microservice design, especially prompt injection, along with mitigations such as authentication, sanitizing prompts and restricting LLM capabilities.
• The alternative pipeline passes TSV files between steps in Google Colab. It handles failures with retry loops and append-mode output files with deduplication.
• The model could not accept GIFs, so a single PNG frame was extracted from each one. Its description was generated once and cached (memoized) in a hints TSV file. Gemma performed well despite being smaller than Gemini.