Research On Dependency Parsing With Partial Annotations

Supervised machine learning methods usually rely on large-scale and high-quality manually annotated data to train model parameters.However,it is time-consuming and laborious to construct large scale artificial data.And the quality of human annotation is difficult to control.This thesis focuses on dependency parsing with partial annotations.On one hand,it discusses how to reduce the cost of manual annotation by a large margin.On the other hand,it investigates how to make use of partial annotation data to effectively train model parameters.Main research contents of this thesis are as follows:(1)Dependency parsing based on neural network.In recent years,neural network has made great progress in dependency parsing.To further study dependency parsing with partial annotations,we use N3 LDG,a lightweight C++ neural network library,to implement two state-of-art dependency parsers,including a biaffine neural network graph-based parser and a globally normalized neural network transition-based parser.Based on our implemented parsers,we can expand them more flexibly to explore the training methods based on partial annotations.(2)Training methods of dependency parsing with partial annotations.We choose five mainstream dependency parsers to explore how different parsers learn from partial annotations.We use two different learning strategies:(a)directly-train,(b)complete-then-train.We generate simulated partial annotation data in three different ways and carry out large-scale simulation experiments to study the abilities of different dependency parsers to learn model parameters from partial annotations.Results show that the high-order probabilistic model based on conditional random field is the most effective in directly learning from PA.All other four parsers can achieve their best performance with the complete-then-train approach where partial annotations are completed into full annotations by the fine-trained parser.(3)Active learning for dependency parsing with partial annotations.We study how to use active learning to select partial annotations,so that the model can achieve better performance with least human efforts.We propose the more flexible dependency-wise partial annotation as a finer-grained unit for dependency parsing.We also propose and compare several uncertainty metrics for annotation tasks of different granularities through simulation experiments.Results show that active learning with partial annotations can greatly reduce manual annotation.We conduct human annotation experiments and find that partial annotation is less time-consuming than full annotation.In conclusion,this thesis studies how to build partially annotated data to reduce the cost of manual annotation and explores how to make use of partial annotations to improve the accuracy of dependency parsing.We have accomplished some primitive progress so far,and we hope that these conclusions can be applied to other natural language processing tasks.