Tagger evaluation

 

As a base line, we tested two different taggers (the TreeTagger at STR, and the HMM-Tagger at RXRC, see section 4.3.1) on the standard STTS tagset as documented in [&make_named_href('', "node40.html#Teufel:95a","[Teufel 1995a]")] and [&make_named_href('', "node40.html#Schiller+al:95","[Schiller et al 1995]")]. This test run serves as a comparative value for all subsequent tagset evaluation tests in section 6.2.

For every test, the following types of tables are given:

1. Corpus statistics (cf. section 5.1.1)

   The corpus statistics table shows numbers of tokens, tags and ambiguity classes in the training and test corpus, which depend on the lexicon and on the guesser or suffix lexicon.

2. Error statistics (cf. section 5.1.2)
   • Accuracy by ambiguity type
     The first table contains the number of errors in the test corpus, classified by the number of ambiguous tags and the type of errors (LE = lexical error; DE = disambiguation error).
   • Most frequent errors
     There are two tables for the most frequent errors: errors classified by wrongly tagged word forms, and errors classified by tags (across word forms). In both cases, we compare the tags asigned manually with the result of automatic taggers..

 
• TreeTagger: Standard test
• Xerox HMM-tagger: Standard test
• Interpretation