<![CDATA[The increasing volume of incoming correspondence at the Kupang Military Court III-15 Office has made the conventional letter classification process complex and time-consuming. Purpose: this study aims to develop an incoming letter classification system that categorizes letters into four classes (regular letters, circulars, decrees, and orders) using the Long Short-Term Memory (LSTM) method; the main contribution of this study is the application of LSTM to a local military-court correspondence dataset that has not been widely studied, together with a replicable preprocessing pipeline and K-Fold evaluation protocol, providing practical implications for accelerating correspondence administration in military judicial offices. Methods: the dataset consists of 500 incoming letter records in Excel format that underwent a preprocessing stage including cleaning, case folding, normalization, stopword removal, stemming, tokenizing, encoding, and padding, and was then divided into 80% training data and 20% testing data, evaluated using 5-Fold Cross Validation with accuracy, precision, recall, and F1-score as performance metrics. Results: the average model performance results were Accuracy 42.04%, Precision 42.39%, Recall 42.04%, and F1-Score 39.93%, with the highest accuracy obtained in Fold 2 (56.44%) and the lowest in Fold 5 (32.00%), while the model's main difficulty lay in distinguishing between the Circular and Decree categories, which share similar text patterns. Conclusion: the LSTM method is capable of recognizing textual patterns in letters and performing classification; however, its performance remains variable and relatively low on this dataset, so the developed system has the potential to improve the efficiency of incoming letter management at the Kupang Military Court III-15 Office, although further optimization is still needed before independent deployment.]]>
