As high-frequency bands are considered for wireless communications, the in-phase and quadrature imbalance caused by the local oscillator and frequency mixer is becoming more severe. This study proposed an I/Q phase imbalance estimator based on the self-attention mechanism, which captures long-range dependencies to learn the relations among the input data symbols. Simulation results showed that the proposed self-attention-based method reduces the estimation error compared with the conventional methods. The proposed method can reduce the required data symbols to estimate the I/Q imbalance by half while achieving better performance at high signal-to-noise ratio regimes.