Optimalisasi Chatbot Helpdesk TIK Berbasis Seq2Seq Encoder BiLSTM dan Retrieval

The Information and Communication Technology (ICT) helpdesk service at the Bone Bolango Regency Communication and Information Agency still faces problems such as limited service hours, dependence on staff availability, and inconsistent answers to recurring questions. These conditions result in inconsistent response times, potentially reducing service effectiveness and making it difficult to meet ICT information needs quickly and consistently within the local government environment. This study aims to evaluate and optimize the performance of ICT helpdesk chatbots in understanding the context of user questions, classifying service categories, and selecting relevant and consistent answers in the closed public service domain. The approach used is a quantitative method with an experimental design, in which the chatbot model was developed using a hybrid Sequence-to-Sequence (Seq2Seq) architecture with a Bidirectional Long Short-Term Memory (BiLSTM) encoder and attention mechanism, as well as a retrieval-based response approach in selecting answers. The dataset consisted of 7,192 question-answer pairs that were manually compiled based on real cases of ICT services and grouped into four service categories, while the analysis technique was carried out through model performance evaluation using accuracy, precision, recall, F1-score, top-3 accuracy metrics, and confusion matrix analysis to assess the stability and accuracy of the classification. The results show that the model is capable of achieving classification accuracy close to 99% with a high F1-score and stable top-3 accuracy on the test data, proving that the encoder-centric BiLSTM approach with a retrieval mechanism effectively produces relevant, consistent, and controlled responses for application in domain-specific public service helpdesk chatbots.