Automated word puzzle generation using topic
models and semantic relatedness measures

Abstract. We propose a knowledge-lean method to generate word puzzles from unstructured and unannotated document collections. The presented method is capable of generating three types of puzzles: odd one out, choose the related word, and separate the topics. The difficulty of the puzzles can be adjusted. The algorithm is based on topic models, semantic similarity, and network capacity. Puzzles of two difficulty levels are generated: beginner and intermediate. Beginner puzzles could be suitable for, e.g., beginner language learners. Intermediate puzzles require more, often specific knowledge to solve. Domain-specific puzzles are generated from a corpus of NIPS proceedings. The presented method is capable of helping puzzle designers compile a collection of word puzzles in a semi-automated manner. In this setting, the method is utilized to produce a great number of puzzles. Puzzle designers can choose and maybe modify the ones they want to include in the collection.