Emotion plays a big role in our daily life. When we try to perceive emotion, we do not only rely on one modality, but rely on several modalities. Psychology studies show that our human sensories perceives several signals from our environment and translate them to codes that are similar across people. In our work, we formulate the emotion recognition as emotion translation task using Sequence-to-sequence (Seq2seq) models which are widely used in neural machine translation task. Additionally, we add attention mechanism as this mechanism can help the model to remember long sequences. Motivated by Google Neural Machine Translation (GNMT), we also try to add residual connection to resolve the decreasing performance when the models have several stacks of hidden layers. We use CMU-MOSEI dataset to train and evaluate our models. Experiment shows that our proposed Seq2seq architecture outperforms the baseline model on emotion translation task. Moreover, the models that use several modalities achieve better performance than the models that only use one modality. This observation proves that multimodal representation escalates the performance of emotion translation or emotion recognition.