East Lombok Regency has diverse tourism destinations, but information is fragmented across various platforms. This study aims to
develop "GoLotim," a custom-built Natural Language Processing (NLP) chatbot integrated into a web-based tourism information system.
Unlike pre-built frameworks (e.g., Dialogflow, Rasa), this approach ensures data sovereignty, lightweight deployment, and full control
over localized domain terminology. The system was developed using the Waterfall SDLC with ReactJS, Flask, and MongoDB. The
knowledge base comprises a manually compiled JSON dataset consisting of 6 distinct intents (tags) and 42 training question patterns.
User queries undergo tokenization, case folding, and stemming, followed by text representation via a 128-dimensional Bag of Words
(BoW). The feature vector is classified using a Neural Network with a 64-neuron hidden layer (ReLU) and a Softmax output layer.
Evaluation was conducted via Black Box Testing and 30 predefined chatbot question scenarios (5 per intent). An "appropriate response"
is operationally defined as the system correctly predicting the intended tag and retrieving factually matching information from the
knowledge base. The chatbot achieved a scenario-based success rate of 93.3% (28 appropriate, 2 inappropriate). Technical classification
metrics yielded a Macro-Average Precision of 0.94, Recall of 0.93, and F1-Score of 0.93, supported by Confusion Matrix analysis. The
average response time was 1.20 seconds (Min: 0.85s, Max: 1.65s, Std. Dev: 0.22s). The two misclassifications were identified as False
Negatives caused by unseen vocabulary variations outside the 128-feature BoW limit. These findings demonstrate the technical feasibility
of the custom NLP implementation for regional tourism. However, the study is limited by the small dataset size and the absence of direct
user satisfaction evaluation, highlighting the need for future research on larger independent test sets and user-based assessments.
Baca artikel lengkap dalam format PDF.