The following tagger programs were available at STR and RXRC for a further extension of the tests described. The reasons why they were not used for the actual experiments are given below.
The bigram/trigram tagger by André KEMPE ([&make_named_href('', "node40.html#Kempe:94","[Kempe 1994]")]) is available for English, French and German. It uses a (modified) version of the Viterbi algorithm and is based on HMM modelling, i.e. the method for tagging is basically the same as the one used by the Xerox HMM tagger. Thus, we don't expect important differences for the results.
The NetTagger ([&make_named_href('',
"node40.html#Schmid:94a","[Schmid 1994a]")]) uses a recursive single-layer
feed-forward neural network for part-of-speech disambiguation. The
network input consists of vectors of lexical tag probabilities for
the words in the context of the current word (about three preceding
and two following words). The output of the network is a also a
probability vector
which
is fed back into the network as the probability vector of the
preceding word. The tagger is trained on tagged data using a
modification of the Backpropagation algorithm.. Like the
TreeTagger, the NetTagger uses a suffix lexicon to guess the tag
probabilities of an unknown word form.
The rule based tagger by Eric BRILL ([&make_named_href('',
"node40.html#Brill:92","[Brill 1992]")],
[&make_named_href('',
"node40.html#Brill:94","[Brill 1994]")]) is trained in two steps. In a first phase the
tagger assigns each word its most likely tag, estimated by examining
a large tagged corpus. The second consists of improving the
results using contextual information. Errors in tagging are
recorded as triples 
and for each triple the
contextual rule is determined which results in the greatest error
reduction. For each rule the net improvement is calculated by
substracting the number of errors introduced by this rule
from the number of corrected errors. The rule which results
in the greates improvement is added to the rule set. The rule
acquisition procedure iteratively continues.
The Brill tagger is available for English and the part-of-speech
categories of the Penn TreeBank.
At RXRC, Jean-Pierre CHANOD has worked on a rule-based tagger for French ([&make_named_href('', "node40.html#Chanod+Tapanainen:95","[Chanod, Tapanainen 1995]")]) which uses finite-state techniques for disambiguation.
A set of finite-state rules describe permissible patterns of tags over a sentence. Composing these constraints over a morphologically analyzed input sentence eliminates most spurious tags. In the tagger, the rules are represented as finite-state transducers that are composed one after another with an ambiguous sentence. After all the transducers have been applied, the sentence has only one analysis.
The RXRC finite-state tagger is only available for French. As we restricted our experiments to German (and a tagset conforming to the EAGLES standards), we excluded this tagger from the tests. A detailed comparison of the Xerox HMM tagger and the finite state tagger for French can be found in [&make_named_href('', "node40.html#Chanod+Tapanainen:94","[Chanod, Tapanainen 1994]")].