Latent Intent Forecasting for Low-Resource: A JEPA Approach to Amharic Conversational Intelligence Using Dialogue Dataset

Intent forecasting in dialogue models remains a challenge for low-resource languages such as Amharic. Amharic is the official language of Ethiopia. More than 57 million people speak it. Amharic lacks large annotated datasets and high-performance computing, which limits model accuracy and slows progress in conversational intelligence models. We present Joint-Embedding Predictive Architecture (JEPA), a lightweight adaptation of the Joint-Embedding Predictive Architecture to text-based dialogue. JEPA operates entirely in latent space: a frozen multilingual encoder extracts 512-dimensional representations of each dialogue turn, and a compact three-layer Transformer predictor learns to forecast the latent embedding of the next turn without generating text. We introduce the Amharic Dialogue Benchmark (ADB-1K), a curated corpus of 1000 context-response pairs spanning five intent categories, augmented with orphological and noisy variants. Trained with an Exponential-Moving-Average target branch and mean-squared-error loss, JEPA reaches a validation Latent Cosine Similarity of 0.9228 at epoch 10 and achieves 60% intent-probe accuracy on the held-out test set, outperforming a fine-tuned GPT-2 baseline (15%) and a random control (20%) while using only 2.1% of GPT-2's parameter count (2.6M versus 124.4M). Morphological robustness degradation is zero (∆mr = 0.000), confirming tolerance to Amharic inflectional variation.