On training bi-directional neural network language model with noise contrastive estimation
Tianxing He, Yu Zhang, Jasha Droppo, Kai Yu · 2016
Although uni-directional recurrent neural network language model(RNNLM) has been very successful, it's hard to train a bi-directional RNNLM properly due to the generative nature of language model. In this work, we propose to train bi-directional RNNLM with noise contrastive estimation(NCE), since the properities of NCE training will help the model to acheieve sentence-level normalization. Experiments are conducted on two hand-crafted tasks on the PTB data set: a rescore task and a sanity test. Although(regretfully), the model trained by NCE did not out-perform the baseline uni-directional NNLM, it is shown that NCE-trained bi-directional NNLM behaves well in the sanity test and outperformed the one trained by conventional maximum likelihood training on the rescore task.