Research On Addressing Data Sparseness In English Event Extraction

Event is defined as a specific occurrence involving participants. Event extraction is an important task in Automatic Content Extraction(ACE), which aims to identify and extract triggers of events and the corresponding arguments in raw texts. Current supervised event extraction methods generally suffer from sparse informative training data, which causes the difficulty in detecting and extracting the uncommon or undiscovered constitutes of events. We address this issue from the following points of view:Using Frame Semantic Knowledge to Improve Event ExtractionAs existing methods based on supervised learning often suffer from date sparseness, we use frame semantic knowledge to improve event extraction by regarding frame type as general feature and mapping frame to event. Finally we combine event recognition model with frame recognition model to make a joint decision. Experiments show that frame-based event recognition model is superior to traditional event recognition model.A Case Study on Active Learning for Event ExtractionThe currently available ground-truth data of event extraction are sparse and low qualified. Although favorable for enrichment of diverse ground-truth data, manual annotation is labor-intensive and time-consuming. To solve the problem we introduce active learning into the process of human-computer cooperated event annotation and extraction. In particular, we leverage the active learning methods in two state-of-the-art event extraction frameworks, including pipeline and joint models based extraction. Towards the pipeline model, we drive active learning to identify and annotate the most informative instances at each extraction stage(trigger and argument classification). It proceeds step-by-step and iteratively until the extraction at each stage reaches the optimal state. While for the joint model, we incorporate active learning with structured perceptron to identify and annotate the informative and interdependent event constituents. It proceeds over the integrated extraction results. Empirical studies show that active learning yields promising improvement as well as substantially reduces the annotation cost.Combining Statistical Model and Dictionary for Event ExtractionWe propose a method that combines results of multi-models and high-confidence dictionary for event extraction. This method introduces dictionary features into maximum entropy model and conditional random fields model respectively, then combines the results of two models in trigger classification. In addition, we consider the lexical length and the length of the dependency path between the trigger and negation or speculation word in event realis recognition.