Meta-Learning Orthographic and Contextual Models for Language Independent Named Entity Recognition

Meta-Learning Orthographic and Contextual Models for Language Independent Named Entity Recognition Robert Munro and Daren Ler and Jon Patrick Language Technology Research Group Capital Markets Co-operative Research Centre University of Sydney {rmunro,ler,jonpat}@it.usyd.edu.au Abstract This paper presents a named entity classifica- tion system that utilises both orthographic and contextual information. The random subspace method was employed to generate and refine at- tribute models. Supervised and unsupervised learning techniques used in the recombination of models to produce the final results. 1 Introduction There are commonly considered to be two main tasks in named entity recognition, recognition (NER) and classi- fication (NEC). As the features that best classify words according to the two tasks are somewhat disparate, the two are often separated. Attribute sets may be further divided into subsets through sub-grouping of attributes, sub-grouping of instances and/or the use of multiple clas- sifying processes. While the use of multiple subsets can increase overall accuracy, the recombination of models has been shown to propagate errors (Carreras et al., 2002; Patrick et al., 2002). More importantly, the decision re- garding the separation of attributes into various subsets is often a manual task. As it is reasonable to assume that the same attributes will have different relative levels of sig- nificance in different languages, using the same division of attributes across languages will be less than optimal, while a manual redistribution across different languages is limited by the users knowledge of those languages. In this paper, the division and subsequent recombination of subgroups is treated as a meta-learning task. 2 Feature Representation It has been our intention to create linguistically driven model of named entity composition, and to search for the attribute representations of these linguistic phenom- ena that best suit inference by a machine learning algo- rithm. It is important to note that a named entity is a label that has been consciously granted by some person or per- sons, and as these names are chosen rather than assigned randomly or evolved gradually, there are generalisations that may be inferred about the words that may be used for naming certain entity types (Allan, 2001; Kripke, 1972). While generalisations relating to abstract connotations of a word may be difficult to infer, generalisations about the structure of the words are more emergent. As the use of a name stretches back in time, it stretches back to a different set of etymological constraints. It may also stem from another language, with a different ortho- graphic structure, possibly representing a different under- lying phonology. Foreign words are frequently named entities, especially in the domain of a newswire such as Reuters. In language in general it is also reasonable to as- sume that a foreign word is more likely to be an entity, as people are more likely to migrate between countries than prepositions. It is these generalisations of the structure of words that we have attempted to represent in the n-gram features. Another emergent structural generalisation is that of capitalisation, as named entities are commonly expressed in title-case …

Content truncated to 500 words