Investigating Lstm For Punctuation Prediction

We present a neural network based punctuation prediction method using Long Short-Term Memory (LSTM) network. The proposed method uses bidirectional LSTM to encode both the past and future observation as its inputs. It models the dependency between input features and output labels through multiple layers. We also empirically study the impacts of modeling the dependency between output labels. Our results show that using a deep bi-directional LSTM is able to achieve state-of-the-art performance in punctuation prediction.