This paper enhances cross-sectional currency momentum strategies by introducing a context-aware Learning to Rank (LTR) model that re-ranks initial asset rankings using a Transformer-based architecture. Traditional LTR models learn a global ranking function that performs well on average but fails to adapt to time-varying feature distributions, leading to suboptimal rankings during critical periods like risk-off episodes. The proposed method leverages the local ranking context—specifically the top and bottom ranked assets—to refine the initial rankings, analogous to re-ranking techniques in information retrieval. The model uses a Transformer encoder with self-attention and positional encodings to capture inter-item dependencies and order information. Backtesting on 31 currency pairs from 2000 to 2020 shows that the context-aware approach significantly improves risk-adjusted returns, increasing the Sharpe ratio by about 30% compared to baseline LTR models. The improvements are particularly pronounced during risk-off market states, where the model achieves substantially higher Sharpe ratios and better ranking accuracy (NDCG@3). The results demonstrate that incorporating context-awareness into LTR can mitigate the weaknesses of global rankers and enhance the robustness of cross-sectional currency strategies.