English Sequence Labeling Research Based On Neural Networks

With the rapid arrival of the Internet era,the network information starts to take on an explosive development trend.Among these information,text data occupies an important part.Understanding this kind of information will be of great significance to the state,society and individuals.How to use the machine to analyze these massive amounts of text information became a problem to be solved.Natural language processing is to explore how to utilize machine to understand and use natural language.Sequence labeling is a basic task in this field.It will be helpful for other related deeper-level research field if the existing methods of sequence labeling can be improved.Based on this background,two subtasks of sequence labeling are studied:Chunking(include NP-Chunking)task and NER(named-entity recognition)task.We find that the lexical information plays an important role in the sequence labeling task and the corresponding designed model achieves good performances on the two sequence labeling tasks experiments.The main works of this thesis are in the following aspects:1)We find that word hashing method has unique advantages in the acquisition of morphological characteristics,and the morphological features is important for the sequence labeling task.In this thesis,our experiments apply word hashing method to sequence labeling tasks.2)For the original sparse word hashing representation,we use auto-encoder to obtain hidden morphological representation features in a pre-stage.The learned character-level representation contains abandunt morphological information and pre-training stage makes the model architecture design easier and more convenient.3)We design a 2-layer bidirectional long short-term memory weight shared neural networks,this architecture conforms to the end-to-end pattern.The model is simple and does not rely on manually handcrafted features and data pre-processing.Our research studies the validity of the morphological features and discusse s the role of semantic features and the task features.The model finally performs solid results in Chunking and NP-Chunking experiments and obtain competitive performance in NER experiments.The model also can provide basic and important features for other tasks.