<![CDATA[This research proposes an Indonesian-language mental health chatbot model based on the LSTM Sequence-to-Sequence (Seq2Seq) architecture as an adaptive initial support solution. Unlike static classification models, this generative approach aims to capture emotional dependencies and conversational context through context vectors. The research methodology utilizes the public PSYCHIKA dataset, which includes 5,667 conversation pairs a significant volume for a low-resource language. Evaluation was conducted by comparing 80:20 and 70:30 data split schemes. Experimental results showed the best performance with the 80:20 split, achieving a BLEU-1 score of 0.137, compared to the 70:30 split, which only reached 0.043. The model achieved stable convergence at 15–16 epochs via an early-stopping mechanism without any signs of overfitting. Although training stability was maintained, the low BLEU score confirms that the use of a pure Seq2Seq LSTM without an attention mechanism is not yet sufficient to generate highly fluent responses. These findings provide a reproducible technical baseline for the development of mental health dialogue systems in Indonesia, while also emphasizing the urgency of more advanced architectures to improve the quality of empathy in the future.]]>
Copyrights © 2026
