Bilingual Continuous Space Language Model

We attempt using simple feedforward neural networks to learn a continuous space language model (CSLM) and evaluate its performance on French and English. The training of the model is performed bilingually, in which we try to capture not only different features from one language, but also the multilingual relation demonstrated in the parallel training data. We achieved the perplexity of 191.2 on French and 193.5 on English, and visualized some interesting monolingual and bilingual relations.