Rotated Canonical Correlation Analysis for Multilingual Corpora

This paper aims at proposing the joint use of Canonical Correlation Analysis and Procrustes Rotations (RCA), when we deal with a text and its translation into another language. The basic idea is representing words in the two different natural languages on a common reference space. The main characteristic of this space is to be language independent, although Procrustes Rotation is performed transforming the lexical table derived from translation by minimizing its distance from the lexical table belonging to the original corpus, while the subsequent Canonical Correlation Analysis treats symmetrically the two word sets. The most interesting RCA feature is building a unique reference space for representing the correlation structure in the data, inducing the two systems of canonical factors to lie on the same space. These graphical representations enables us to read distances between corresponding points in terms of different way of translating the same word in relation with the general context defined by the canonical variates. Trying to understand the distances between matched points could represent an useful tool for enriching lexical resources in a translation procedure. In this paper we propose the comparison of the most frequent content bearing words in the two languages, analyzing one year (2003) of Le Monde Diplomatique and its Italian edition.