Scaling Context Space

Scaling Context Space James R. Curran and Marc Moens Institute for Communicating and Collaborative Systems University of Edinburgh 2 Buccleuch Place, Edinburgh EH8 9LW United Kingdom � jamesc,marc � @cogsci.ed.ac.uk Abstract Context is used in many NLP systems as an indicator of a term’s syntactic and se- mantic function. The accuracy of the sys- tem is dependent on the quality and quan- tity of contextual information available to describe each term. However, the quan- tity variable is no longer fixed by lim- ited corpus resources. Given fixed train- ing time and computational resources, it makes sense for systems to invest time in extracting high quality contextual in- formation from a fixed corpus. However, with an effectively limitless quantity of text available, extraction rate and repre- sentation size need to be considered. We use thesaurus extraction with a range of context extracting tools to demonstrate the interaction between context quantity, time and size on a corpus of 300 million words. 1 Introduction Context plays an important role in many natural lan- guage tasks. For example, the accuracy of part of speech taggers or word sense disambiguation sys- tems depends on the quality and quantity of con- textual information these systems can extract from the training data. When predicting the sense of a word, for instance, the immediately preceding word is likely to be more important than the tenth previ- ous word; similar observations can be made about POS taggers or chunkers. A crucial part of train- ing these systems lies in extracting from the data high-quality contextual information, in the sense of defining contexts that are both accurate and corre- lated with the information (the POS tags, the word senses, the chunks) the system is trying to extract. The quality of contextual information is often de- termined by the size of the training corpus: with less data available, extracting context information for any given phenomenon becomes less reliable. However, corpus size is no longer a limiting fac- tor: whereas up to now people have typically worked with corpora of around one million words, it has be- come feasible to build much larger document collec- tions; for example, Banko and Brill (2001) report on experiments with a one billion word corpus. When using a much larger corpus and scaling the context space, there are, however, other trade-offs to take into consideration: the size of the corpus may make it unfeasible to train some systems because of efficiency issues or hardware costs; it may also result in an unmanageable expansion of the extracted con- text information, reducing the performance of the systems that have to make use of this information. This paper reports on experiments that try to es- tablish some of the trade-offs between corpus size, processing time, hardware costs and the perfor- mance of the resulting systems. We report on ex- periments with a large corpus (around 300 mil- lion words). We trained a thesaurus extraction sys- tem with a range of context-extracting front-ends to demonstrate the interaction between context quality, …

Content truncated to 500 words