Semantic Structure from Correspondence Analysis

Friday Seminar by Barbara McGillivray (U. Pisa)

A common problem for clustering techniques is that clusters overlap,
which makes graphing the statistical structure in the data difficult.
A related problem is that we often want to see the distribution of
factors (variables) as well as classes (objects). Correspondence
Analysis offers a solution to both these problems.
We performed a correspondence analysis to study the semantic behaviour
of verbs using the co-occurrences of verbs and nouns extracted from
corpora. This distributional method was tested on Italian and English
and was able to discover an underlying structure in our data.
We will report the results of our analysis, showing a comparison
between these two languages. We will also discuss work in progress for
translations viewed as correspondences (cf. the Mirrors Method (Dyvik,
2002)).