In my talk, I will present the **chunk tagger**, i.e., a shallow parser based
on Hidden Markov Models (HMMs). The parser assigns tree segments of limited
depth to sequences of part-of-speech tags in the same way a POS tagger finds
the optimal tag sequence for a sequence of words, e.g.
Though ambiguity is much higher than with standard POS tagging, the chunker
achieves fairly good accuracy as far as recognition of simple as well as
complex NPs, PPs and APs is concerned (89.5% unlabelled total match). The
representation format for structures enables us to take advantage of
disambiguation hints provided by strictly local lexical contexts (trigrams of
POS tags).
The chunker is available in two versions which differ w.r.t. the techniques
used for parameter estimation (linear interpolation vs. maximum entropy). I
will contrast these two method and compare their results.