... sections.¹

In case $c_{i-1}$ is part of a previous section or $c_{i+1}$ is in a next section, they are left empty.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... data.²

The noun phrase identification data is available from ftp://ftp.cis.upenn.edu/pub/chunker/

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... five³

The combination of open and close brackets, O+C, will be regarded as one data representation.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... data.⁴

The CoNLL-2000 shared task data is available from http://lcg-www.uia.ac.be/conll2000/chunking/

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...tksveenstra99eacl.⁵

The problem of using the predicted class of the current word was a result of an earlier study in which we did not use feature selection. The selection method used in this study would probably have disregarded this feature automatically. It would start out as the most informative feature but with the feature on its own we would get a worse performance than with combinations of other features (we perform feature selection while keeping the five best combinations).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... types⁶

The results of our arbitrary phrase identification work have earlier been presented by [Tjong Kim Sang(2000b)].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... way:⁷

This approach and the results achieved with it have earlier been discussed by [Tjong Kim Sang(2001a)].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...collins99.⁸

Available on http://www.research.att.com/~ mcollins/papers/heads

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... phrases.⁹

This performance was already reported by [Tjong Kim Sang(2000a)].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... task.¹⁰

An elaborate overview of most of the systems that have been applied to this task can be found on http://lcg-www.uia.ac.be/~ erikt/research/np-chunking.html

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...kudoh2001.¹¹

Although we do not wish to underestimate the power of Support Vector Machines, we should note that it seems that the optimal results presented by [Kudoh and Matsumoto(2001)] have been obtained by tuning the system to the test data.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... task.¹²

More results for the chunking task can be found on http://lcg-www.uia.ac.be/conll2000/chunking/

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... here.¹³

Our full parser, which was trained and tested on the same data as the Collins parser, obtained F$_{\beta =1}$=86.96 for recognizing NP phrases only.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Grammars.¹⁴

http://lcg-www.uia.ac.be/

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.