GWC 2008

Verification of Valency Frame Structures by Means of Automatic… 41
It is possible to assess the clustering quality so that to choose the optimal window
width (and other parameters). We consider clustering to be interpretable if verbs from
the same semantic tree are put together in one cluster, and inexplicable – if there are a
lot of representatives from different semantic groups mixed in one cluster. If there
were not morphological correlations among contexts of different verbs with similar or
relative meanings, we’d never received any explicable groupings.
Results of [4] show that a range of positions in the window has a cumulative
outcome. So we compared verb distributions per each position and per position range
and received expected result: there were no separate position in the window [-
10…+10], which was sufficient by itself to show reliable clustering. On Fig. 3, 4, 5
there are stemmas for clustering in the [-10] position, [+1] position and the better
range [-3,+5]. Grey background marks verbs from the same semantic group, numbers
at the top nodes show the step of clustering and an average similarity
5 An Optimal Tag Set for Distribution Capture
The tag set (TS) is another key point of distribution description. The first variant of
the tag set showed interpretable results, however, it was significant to what extent it
may deviate clustering. We tried 3 tag sets: the 1st TS was described above, the 2nd
TS was simple POS tagging without specification of grammar category values (e.g. N,
A, Adv, V, Pron, etc.); the 3rd TS was a kind of POS tag generalisation: all
substantives were united under one tag, however, case specification for substantives
were added (Nom, Gen, Dat, Acc, Abl, Loc, V, Adv, etc.). The comparison of
clustering parameters shows that clustering with the 2nd TS produces a flatter
structure without elaboration of the inner structure in groups, the clustering with the
3rd TS creates more detailed structure than the 1st TS.
Fig. 5. The stemma of verb clustering in the [-3,+5] position range
of the distribution window with the 1st TS