Automatic Digitization of Chess Scoresheets with RNNs

During chess games, players write each move on a chess scoresheet. Afterwards, they manually insert these moves onto a computer, a process that can take up to an hour. To this end, this paper presents a solution to instantaneously digitize scoresheets using a Convolutional Bi-LSTM. This paper proposes 4 key contributions: 1) A general model that achieves a character accuracy rate (CRA) of $\mathbf{9 4. 3 6 \%}$. 2) A personalized model that is fine-tuned on dataset of one specific handwriting that achieves a higher CRA of $\mathbf{9 5. 2 2 \%}$. 3) A data extractor for data collection. And 4) A Valid Move Checker that flags invalid p redictions for result analysis. The proposed model outperforms baseline models and decreases the time of digitizing a game from 5-10 minutes to a few seconds.