This thesis explored grammatically correct but semantically deviant syntagmas. Syntagmas are pairs of nouns and its attributes. For the pairs to be grammatically correct, they have to match in gender, number and case. For them to be semantically very deviant, their usage should be rare in common conversations. Pairs are even more semantically deviant if they have never been used in conversations and writings. Those syntagmas are a basis of a word guessing board game which are to be guessed during all 3 rounds of the game. The first round is a game similar to Alias. For the rest of the rounds all of the syntagmas are being repeated and explained using charades in the second round and then using only one word for the whole syntagm. The combination of these three games create another background game which last during the whole process. A game of memory. It’s the key element of the whole game because it eliminates idling of other players while it’s not their turn to play. It keeps everyone engaged all the time because they have to listen to others to remember the syntagmas for the next rounds. The first step is gathering all nouns and attributes from the Croatian language cor- pus. The corpus contains lexically and semantically modeled words which contain data needed to summarize them into a smaller set. In the end that set contains around 7500 nouns and 7000 attributes which gives out over 50 million combinations of syntagmas for the game. Also, around 14500 commonly used syntagmas have been gathered from the corpus and are used as a blacklist for the game. Common syntagmas are a lot easier to explain, even using only one word in the third round. Syntagmas were considered common if there is at least 100 of them showing up in the corpus. The application is completely built using Haskell. The Scotty library was used for the server side and it has only three tasks: deliver the web page, generate random syntagmas, and gather data about how much time it takes to guess those syntagmas. The client side of the application was built as a single-page app using the Miso fra- mework which is based on the Elm architecture and functional reactive programming. The application is divided into three main parts: model, view of the model, and update of the model.