text
stringlengths 0
203
|
|---|
employed memory-based learning (De Meulder and |
Daelemans, 2003; Hendrickx and Van den Bosch, |
2003). Transformation-based learning (Florian et |
al., 2003), Support Vector Machines (Mayfield et al., |
2003) and Conditional Random Fields (McCallum |
and Li, 2003) were applied by one system each. |
Combination of different learning systems has |
proven to be a good method for obtaining excellent |
results. Five participating groups have applied system combination. Florian et al. (2003) tested different methods for combining the results of four systems and found that robust risk minimization worked |
best. Klein et al. (2003) employed a stacked learning system which contains Hidden Markov Models, |
Maximum Entropy Models and Conditional Markov |
Models. Mayfield et al. (2003) stacked two learners |
and obtained better performance. Wu et al. (2003) |
applied both stacking and voting to three learners. |
Munro et al. (2003) employed both voting and bagging for combining classifiers. |
3.2 |
Features |
The choice of the learning approach is important for |
obtaining a good system for recognizing named entities. However, in the CoNLL-2002 shared task we |
found out that choice of features is at least as important. An overview of some of the types of features |
chosen by the shared task participants, can be found |
in Table 3. |
All participants used lexical features (words) except for Whitelaw and Patrick (2003) who implemented a character-based method. Most of the systems employed part-of-speech tags and two of them |
have recomputed the English tags with better taggers (Hendrickx and Van den Bosch, 2003; Wu et al., |
2003). Othographic information, affixes, gazetteers |
and chunk information were also incorporated in |
most systems although one group reports that the |
available chunking information did not help (Wu et |
al., 2003) Other features were used less frequently. |
Table 3 does not reveal a single feature that would |
be ideal for named entity recognition. |
3.3 |
External resources |
Eleven of the sixteen participating teams have attempted to use information other than the training |
data that was supplied for this shared task. All included gazetteers in their systems. Four groups examined the usability of unannotated data, either for |
extracting training instances (Bender et al., 2003; |
Hendrickx and Van den Bosch, 2003) or obtaining |
extra named entities for gazetteers (De Meulder and |
Daelemans, 2003; McCallum and Li, 2003). A reasonable number of groups have also employed unannotated data for obtaining capitalization features for |
words. One participating team has used externally |
trained named entity recognition systems for English |
as a part in a combined system (Florian et al., 2003). |
Table 4 shows the error reduction of the systems |
Zhang |
Florian |
Chieu |
Hammerton |
Carreras (a) |
Hendrickx |
De Meulder |
Bender |
Curran |
McCallum |
Wu |
G |
+ |
+ |
+ |
+ |
+ |
+ |
+ |
+ |
+ |
+ |
+ |
U |
+ |
+ |
+ |
+ |
- |
E |
+ |
- |
English |
19% |
27% |
17% |
22% |
12% |
7% |
8% |
3% |
1% |
? |
? |
German |
15% |
Subsets and Splits