Research On English Noun Phrase Event Anaphora Resolution

Anaphora Resolution is focus on the Anaphora phenomenon between entity or event. Anaphora resolution is one of the key tasks of natural language processing, this task's research had widely used in machine translation,information extraction et al. The previous researches of anaphora resolution are all focus on entity and had made leaps and bounds. But the research of event anaphora resolution has only just begun. Because the difference of entity between event : event has some characters that entity doesn't have. So, the research of event resolution is necessary and have important research value.According to the different anaphor, Event Anaphora Resolution can be divide into Pronoun Event Anaphora Resolution and Noun Phrase Event Anaphora Resolution. Noun Phrase include some grammar and semantic information. This information can help promote the performance of Anaphora Resolution System.so, we focus on the noun phrase and do research on Noun Phrase Event Anaphora Resolution.First, we focus on the instance imbalance problem in noun phrase anaphora resolution. It can generate a large number of negative instances in instancegeneration. So,we present a method with Centering Theory and Named Entity Recognition information to balance the positive and negative instances form Onto Notes 4.0.Secondly, due to event has some special characteristic that different from entity. We make a research on it. When we describe a event, we need know some keywords like when, who, where, what, how. We find it can match the semantic role form SRL like Arg M-TMP, Arg0. Arg M-LOC, Arg1, Verb. So we use a verb and a set of the verb's semantic roles to show an event and use this way to calculate the similarity between two events.Finally,we add the Arg M-LOC and Arg M-TMP into the SRET, and compare the performance with MET, FET and SET.We explore this feature's influence on event anaphora resolution framework with two candidate model. Form the experiment result, we can find that these features can contribution to the event anaphora resolution framework. It outperforms the baseline system by 0.49% in Precision and 0.2% in F-measure.Then we use the twin-candidate model optimize our framework,It outperforms the baseline system by 2.24% in Precision and 2.23% in F-measure.