Abstract
Syntactic dependency structures leverage shallow semantic information, which could be helpful for downstream tasks in natural language processing (NLP). In this thesis, we present some of our recent works on improving named entity recognition (NER) and semantic parsing by making use of the rich structured information in the dependency trees. We statistically show the strong correlation between the dependency trees and named entities. Specifically, named entities often form subtrees under the dependency tree structures, and the dependency relations (e.g. nsubj) are strong indicators for the existence of entities. Motivated by the above observations, we proposed the dependency-guided structured models based on the conditional random fields (CRF) to capture the underlying correlations for the NER task. Our large-scale experiments on four languages demonstrate the effectiveness of the proposed model, especially for the Catalan and Spanish languages. Further analysis reveals that the improvements mainly result from the dependency relations and long-distance interactions provided by dependency trees. Our following research work attacks a more challenging and semantic-level task, semantic parsing without the given dependency trees. The semantic parsing task is to map the natural language sentences into (tree-structured) logical forms. We introduce a dependency-based hybrid tree representation to capture the interactions between the semantic units and the natural language words. However, as the dependency trees are not observed, we regard the dependency trees as the latent structures in our proposed latent-variable model. Our dynamic-programming inference procedure is similar to the Eisner’s algorithm for dependency parsing. The proposed model achieves state-ofthe-art performance on 7 out of 8 languages, especially for some datasets with flexible word order.