Abstract
Structured data is often in the form of predefined structures or formats so that the incorporated information can be easily understood. However, most of the existing data is unstructured data, and its volume grows more rapidly than structured data. The challenge is how to effectively extract the information from unstructured text data. In this thesis, we explore extracting entity-oriented structured knowledge from unstructured text. First, we explore extracting mentioned entities in the raw text. The entity representations can be enhanced by incorporating the long-distance structured information captured by dependency trees. We believe both types of features – the contextual information captured by the linear sequences and the structured information captured by the dependency trees may complement each other. We propose a simple and robust model to better integrate the structured information leveraged from the long-range dependencies. Then, we target to improve the robustness of NER systems when there are no goldannotated datasets that are available for training. The distantly annotated datasets significantly reduce the annotation cost, but are often noisy and contain a considerable number of false negatives. Intuitively, the false negatives should have high similarities with positive samples that have the same gold entity type, and they should not have high similarities with other positive samples of different entity types. Therefore, we sample the negative samples that have high similarities with all the positive samples for training. Lastly, we focus on extracting entities and the sentiment relation between them from customer reviews text data. We focus on the recently developed Aspect Sentiment Triplet Extraction (ASTE) task that extracts the triplets of target entities, their associated sentiment, and opinion spans explaining the reason for the sentiment. We propose a novel position-aware tagging scheme by enriching label expressiveness to jointly extracting the aspect sentiment triplets. However, it still relies on the interactions between each target word and opinion word. Then, we propose a span-level approach that explicitly considers the interaction between the whole spans of targets and opinions when predicting their sentiment relation.