text
stringlengths
0
203
5%
7%
8%
5%
3%
6%
?
?
Table 4: Error reduction for the two development data sets when using extra information like
gazetteers (G), unannotated data (U) or externally
developed named entity recognizers (E). The lines
have been sorted by the sum of the reduction percentages for the two languages.
with extra information compared to while using only
the available training data. The inclusion of extra named entity recognition systems seems to have
worked well (Florian et al., 2003). Generally the systems that only used gazetteers seem to gain more
than systems that have used unannotated data for
other purposes than obtaining capitalization information. However, the gain differences between the
two approaches are most obvious for English for
which better gazetteers are available. With the exception of the result of Zhang and Johnson (2003),
there is not much difference in the German results
between the gains obtained by using gazetteers and
those obtained by using unannotated data.
3.4
Performances
A baseline rate was computed for the English and the
German test sets. It was produced by a system which
only identified entities which had a unique class in
the training data. If a phrase was part of more than
one entity, the system would select the longest one.
All systems that participated in the shared task have
outperformed the baseline system.
For all the Fβ=1 rates we have estimated significance boundaries by using bootstrap resampling
(Noreen, 1989). From each output file of a system,
250 random samples of sentences have been chosen
and the distribution of the Fβ=1 rates in these samples is assumed to be the distribution of the performance of the system. We assume that performance
A is significantly different from performance B if A
is not within the center 90% of the distribution of B.
The performances of the sixteen systems on the
two test data sets can be found in Table 5. For English, the combined classifier of Florian et al. (2003)
achieved the highest overall Fβ=1 rate. However,
the difference between their performance and that
of the Maximum Entropy approach of Chieu and Ng
(2003) is not significant. An important feature of the
best system that other participants did not use, was
the inclusion of the output of two externally trained
named entity recognizers in the combination process.
Florian et al. (2003) have also obtained the highest
Fβ=1 rate for the German data. Here there is no significant difference between them and the systems of
Klein et al. (2003) and Zhang and Johnson (2003).
We have combined the results of the sixteen system in order to see if there was room for improvement. We converted the output of the systems to
the same IOB tagging representation and searched
for the set of systems from which the best tags for
the development data could be obtained with majority voting. The optimal set of systems was determined by performing a bidirectional hill-climbing
search (Caruana and Freitag, 1994) with beam size 9,
starting from zero features. A majority vote of five
systems (Chieu and Ng, 2003; Florian et al., 2003;
Klein et al., 2003; McCallum and Li, 2003; Whitelaw
and Patrick, 2003) performed best on the English
development data. Another combination of five systems (Carreras et al., 2003b; Mayfield et al., 2003;
McCallum and Li, 2003; Munro et al., 2003; Zhang
and Johnson, 2003) obtained the best result for the
German development data. We have performed a
majority vote with these sets of systems on the related test sets and obtained Fβ=1 rates of 90.30 for
English (14% error reduction compared with the best
system) and 74.17 for German (6% error reduction).
4
Concluding Remarks
We have described the CoNLL-2003 shared task:
language-independent named entity recognition.
Sixteen systems have processed English and German
named entity data. The best performance for both
languages has been obtained by a combined learning system that used Maximum Entropy Models,
transformation-based learning, Hidden Markov Models as well as robust risk minimization (Florian et al.,
2003). Apart from the training data, this system also
employed gazetteers and the output of two externally
trained named entity recognizers. The performance
of the system of Chieu et al. (2003) was not significantly different from the best performance for English and the method of Klein et al. (2003) and the
approach of Zhang and Johnson (2003) were not significantly worse than the best result for German.
Eleven teams have incorporated information other
than the training data in their system. Four of them
have obtained error reductions of 15% or more for
English and one has managed this for German. The
resources used by these systems, gazetteers and externally trained named entity systems, still require a
lot of manual work. Systems that employed unannotated data, obtained performance gains around 5%.
The search for an excellent method for taking advantage of the fast amount of available raw text, remains
open.
Acknowledgements
Tjong Kim Sang is financed by IWT STWW as a
researcher in the ATraNoS project. De Meulder is
supported by a BOF grant supplied by the University
of Antwerp.