Fine-grained Named Entity Classi cation in Machine Reading

Fine-grained named entity classi cation or FG-NEC refers to the process of classifying a set of named entities from naturally occurring texts to the maximum granularity. It is essentially di erent from the traditional coarse-grained NEC (PER, LOC, ORG) in that it requires deep semantic analysis and the FG semantic classes are highly ambiguous. While research has been conducted in an application-oriented manner, few works have addressed this problem per se. This thesis addressed this problem, with a special focus on the person category. Our methodology is to extract the key property of each candidate instance rst and automatically classify them according to a reference taxonomy. The classi cation takes into account the non-uniformity and insu ciency of context clues and uses a cascade framework such that named entities with di erent kinds of context clues are resolved at di erent stages. The cascade framework is highly e cient since the simple instances can be ltered out at early stages thereby the system can focus on the more di cult ones. We also developed a joint-inference based property extraction algorithm for entities whose target properties are explicitly speci ed in the texts. Evaluated on the Wall Street Journal corpus, the extractor achieves an F1 score of 91.91, which is quite competitive. Trained on newswire texts, this framework can be easily tuned to apply to texts in other styles.