Abstract
We investigate deep neural models for lexical analysis and entity detection tasks, which include word segmentation, part-of-speech (POS) tagging and named entity recognition (NER). We build two deep sequence models for these tasks: the neural sequence labeling model and the neural transition-based model. For the neural sequence labeling model, long short-term memory (LSTM) network structure is utilized to encode the informa-tion of input sentence, followed by a conditional random fields (CRF) to give optimal global decoding. The neural transition-based model is studied for word segmentation by formalizing the segmentation task as a state transition process. In this thesis, we study several approaches to improve the above two deep sequence models by incorpo-rating a wider range of external information. We propose a neural sequence labeling model by integrating the discrete statisti-cal features and show that the combination of discrete features and neural features improves the system significantly. We also develop a state-of-the-art Chinese word segmentor with a neural transition-based model, which utilizes external knowledge through rich pretraining. To integrate the word or subword information for Chinese word segmentation and NER, we propose a lattice LSTM structure which encodes word/subword embeddings into a character LSTM sequence. The lattice LSTM model gives comparable results with the best models on both segmentation and NER tasks.