Technical reports
Investigating classification for natural language processing tasks
June 2008, 138 pages
This technical report is based on a dissertation submitted September 2007 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Fitzwilliam College.
DOI: 10.48456/tr-721
Abstract
This report investigates the application of classification techniques to four natural language processing (NLP) tasks. The classification paradigm falls within the family of statistical and machine learning (ML) methods and consists of a framework within which a mechanical ‘learner’ induces a functional mapping between elements drawn from a particular sample space and a set of designated target classes. It is applicable to a wide range of NLP problems and has met with a great deal of success due to its flexibility and firm theoretical foundations.
The first task we investigate, topic classification, is firmly established within the NLP/ML communities as a benchmark application for classification research. Our aim is to arrive at a deeper understanding of how class granularity affects classification accuracy and to assess the impact of representational issues on different classification models. Our second task, content-based spam filtering, is a highly topical application for classification techniques due to the ever-worsening problem of unsolicited email. We assemble a new corpus and formulate a state-of-the-art classifier based on structured language model components. Thirdly, we introduce the problem of anonymisation, which has received little attention to date within the NLP community. We define the task in terms of obfuscating potentially sensitive references to real world entities and present a new publicly-available benchmark corpus. We explore the implications of the subjective nature of the problem and present an interactive model for anonymising large quantities of data based on syntactic analysis and active learning. Finally, we investigate the task of hedge classification, a relatively new application which is currently of growing interest due to the expansion of research into the application of NLP techniques to scientific literature for information extraction. A high level of annotation agreement is obtained using new guidelines and a new benchmark corpus is made publicly available. As part of our investigation, we develop a probabilistic model for training data acquisition within a semi-supervised learning framework which is explored both theoretically and experimentally.
Throughout the report, many common themes of fundamental importance to classification for NLP are addressed, including sample representation, performance evaluation, learning model selection, linguistically-motivated feature engineering, corpus construction and real-world application.
Full text
PDF (1.4 MB)
BibTeX record
@TechReport{UCAM-CL-TR-721,
author = {Medlock, Ben W.},
title = {{Investigating classification for natural language
processing tasks}},
year = 2008,
month = jun,
url = {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-721.pdf},
institution = {University of Cambridge, Computer Laboratory},
doi = {10.48456/tr-721},
number = {UCAM-CL-TR-721}
}