Triangular machine translation is a special case of low-resource machine\ntranslation where the language pair of interest has limited parallel data, but\nboth languages have abundant parallel data with a pivot language. Naturally,\nthe key to triangular machine translation is the successful exploitation of\nsuch auxiliary data. In this work, we propose a transfer-learning-based\napproach that utilizes all types of auxiliary data. As we train auxiliary\nsource-pivot and pivot-target translation models, we initialize some parameters\nof the pivot side with a pre-trained language model and freeze them to\nencourage both translation models to work in the same pivot language space, so\nthat they can be smoothly transferred to the source-target translation model.\nExperiments show that our approach can outperform previous ones.\n