Abstract
This paper introduces the first Turkish crossword puzzle generator designed to leverage the capabilities of large language models (LLMs) for educational purposes. In this work, we introduced two specially created datasets: one with over 180,000 unique answer-clue pairs for generating relevant clues from the given answer, and another with over 35,000 samples containing text, answer, category, and clue data, aimed at producing clues for specific texts and keywords within certain categories. Beyond entertainment, this generator emerges as an interactive educational tool that enhances memory, vocabulary, and problem-solving skills. It's a notable step in AI-enhanced education, merging game-like engagement with learning for Turkish and setting new standards for interactive, intelligent learning tools in Turkish.
Introduction
Crossword puzzles are widely recognized as effective tools for language acquisition and cognitive development. However, the manual creation of high-quality puzzles is a labor-intensive task for educators. While automated generators exist for high-resource languages like English, Turkish—a language with unique morphological features—has remained largely underserved. This research addresses the gap by providing the first specialized tools and large-scale datasets to generate educationally-focused Turkish crossword clues, bridging the divide between standard text and interactive learning.
Methodology
The system employs a two-pronged approach to puzzle creation: clue generation and schema layout.
Clue Generation
The researchers fine-tuned several Large Language Models (LLMs), including GPT-3.5-Turbo, Llama-2-7b, and Llama-2-13b. Two distinct generation strategies were implemented:
- Keyword-based: Generating a clue based solely on a provided answer.
- Context-aware: Generating clues based on a keyword found within a specific educational paragraph, reinforcing active recall of recently read material.
To ensure educational quality, a self-instruct framework with GPT-4-Turbo was used to generate synthetic training data from Wikipedia, moving beyond simple dictionary definitions to more engaging wordplay.
Schema Generation
The "Schema Generator" algorithm constructs the puzzle grid using a multi-factor scoring function to optimize playability: $$\text{Score} = (FW + 0.5 \cdot LL) \times FR \times LR$$ where $FW$ is the total word count, $LL$ represents crossing letters, $FR$ is the filled space ratio, and $LR$ is the intersection density.
Datasets
The study introduces two landmark datasets for Turkish NLP:
- TAC (Turkish Answer-clue Pairs): A collection of over 252,000 pairs featuring nearly 55,000 unique answers, scraped from online sources and newspapers.
- T4TAC (Text for Turkish Answer-clue Pairs): A sophisticated dataset derived from 180,000 Wikipedia pages. After rigorous filtering for popularity and content length, it contains 35,000 samples across 29 themes (like History, Science, and Entertainment), linking specific text segments to answers and clues.
Results and Discussion
Experimental results highlighted the necessity of fine-tuning for Turkish. Base Llama-2 models performed poorly out of the box but saw massive performance gains after training on the specialized datasets. GPT-3.5-Turbo emerged as the top performer, achieving a ROUGE-L score of 23.48, significantly outperforming the Llama-2 variants.
Human validation by native Turkish speakers confirmed the quality of the output, with 51.8% of generated clues meeting high standards for "acceptability," meaning they were contextually relevant, innovative, and accurate.
Conclusion
This generator represents a significant advancement in AI-enhanced education for the Turkish language. By transforming Wikipedia content into interactive wordplay, it provides a scalable tool for educators to boost student engagement and vocabulary retention. Future work will focus on improving the "cleverness" of clues to increase the human-grade acceptability rate and expanding the framework to support multiple languages.
Summary updated