Fine-Grained Classification of Named Entities Exploiting Latent Semantic Kernels

Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL), pages 201–209, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Fine-Grained Classification of Named Entities Exploiting Latent Semantic Kernels Claudio Giuliano FBK-irst I-38100, Trento, Italy giuliano@fbk.eu Abstract We present a kernel-based approach for fine- grained classification of named entities. The only training data for our algorithm is a few manually annotated entities for each class. We defined kernel functions that implicitly map entities, represented by aggregating all con- texts in which they occur, into a latent seman- tic space derived from Wikipedia. Our method achieves a significant improvement over the state of the art for the task of populating an ontology of people, although requiring con- siderably less training instances than previous approaches. 1 Introduction Populating an ontology with relevant entities ex- tracted from unstructured textual documents is a crucial step in Semantic Web and knowledge man- agement systems. As the concepts in an ontology are generally arranged in deep class/subclass hierar- chies, the problem of populating ontologies is typi- cally solved top-down, firstly identifying and classi- fying entities in the most general concepts, and then refining the classification process. Recent advances have made supervised ap- proaches very successful in entity identification and classification. However, to achieve satisfactory per- formance, supervised systems must be supplied with a sufficiently large amount of training data, usually consisting of hand tagged texts. As domain specific ontologies generally contains hundreds of subcate- gories, such approaches are not directly applicable for a more fine-grained categorization because the number of documents required to find sufficient pos- itive examples for all subclasses becomes too large, making the manual annotation very expensive. Consequently, in the literature, supervised ap- proaches are confined to classify entities into broad categories, such as persons, locations, and or- ganizations, while the fine-grained classification has been approached with minimally supervised (e.g., Tanev and Magnini (2006) and Giuliano and Gliozzo (2008)) and unsupervised learning algo- rithms (e.g., Cimiano and V¨olker (2005) and Giu- liano and Gliozzo (2007)). Following this trend, we present a minimally su- pervised approach to fine-grained categorization of named entities previously recognized into coarse- grained categories, e.g., by a named-entity recog- nizer. The only training data for our algorithm is a few manually annotated entities for each class. For example, Niels Bohr, Albert Einstein, and Enrico Fermi might be used as examples for the class physi- cists. In some cases, training entities can be acquired (semi-) automatically from existing ontologies al- lowing us to automatically derive training entities for use with our machine learning algorithm. For instance, we may easily obtain tens of training en- tities for very specific classes, such as astronomers, materials scientists, nuclear physicists, by querying the Yago ontology (Suchanek et al., 2008). We represent the entities using features extracted from the textual contexts in which they occur. Specifically, we use a search engine to collect such contexts from the Web. Throughout this paper, we will refer to such a representation as multi-context representation, in contrast to …

Content truncated to 500 words