In this paper,we propose a novel approach learning bilingual representations to predict quality estimation of machine translation.We use two bi-directional Long Short-Term Memory(LSTM)based architectures map the source sentence and target sentence to two context vector of a fixed dimensionality,then we compute the weighted cosine distance of the two vectors to estimate the translation quality of the target sentence.Our experimental results show that our model improve the performance over a baseline system with 17 features in the English-to-Spanish sentence-level quality estimation task of WMT15.