Representing Text Chunks

Proceedings of EACL '99 Representing Text Chunks Erik F. Tjong Kim Sang Center for Dutch Language and Speech University of Antwerp Universiteitsplein 1 B-2610 Wilrijk, Belgium erikt@uia.ac.be Jorn Veenstra Computational Linguistics Tilburg University P.O. Box 90153 5000 LE Tilburg, The Netherlands veenstra@kub.nl Abstract Dividing sentences in chunks of words is a useful preprocessing step for parsing, information extraction and information retrieval. (l~mshaw and Marcus, 1995) have introduced a "convenient" data rep- resentation for chunking by converting it to a tagging task. In this paper we will examine seven different data repre- sentations for the problem of recogniz- ing noun phrase chunks. We will show that the the data representation choice has a minor influence on chunking per- formance. However, equipped with the most suitable data representation, our memory-based learning chunker was able to improve the best published chunking results for a standard data set. 1 Introduction The text corpus tasks parsing, information extrac- tion and information retrieval can benefit from di- viding sentences in chunks of words. (Ramshaw and Marcus, 1995) describe an error-driven transformation-based learning (TBL) method for finding NP chunks in texts. NP chunks (or baseNPs) are non-overlapping, non-recursive noun phrases. In their experiments they have modeled chunk recognition as a tagging task: words that are inside a baseNP were marked I, words outside a baseNP received an 0 tag and a special tag B was used for the first word inside a baseNP immedi- ately following another baseNP. A text example: original: In [N early trading N] in [N Hong Kong N] [N Monday N], [N gold N] was quoted at [N $ 366.50 N] [N an ounce g] • tagged: In/O early/I trading/I in/O Hong/I Kong/I Monday/B ,/O gold/I was/O quoted/O at/O $/I 366.50/I an/B ounce/I ./O Other representations for NP chunking can be used as well. An example is the representation used in (Ratnaparkhi, 1998) where all the chunk- initial words receive the same start tag (analo- gous to the B tag) while the remainder of the words in the chunk are paired with a different tag. This removes tagging ambiguities. In the Ratna- parkhi representation equal noun phrases receive the same tag sequence regardless of the context in which they appear. The data representation choice might influence the performance of chunking systems. In this pa- per we discuss how large this influence is. There- fore we will compare seven different data rep- resentation formats for the baseNP recognition task. We are particularly interested in finding out whether with one of the representation formats the best reported results for this task can be im- proved. The second section of this paper presents the general setup of the experiments. The results Can be found in the third section. In the fourth section we will describe some related work. 2 Methods and experiments In this section we present and explain the data representation formats and the machine learning algorithm that we have used. In the final part we describe the feature representation used in our experiments. 2.1 …

Content truncated to 500 words