Alignment-based learning versus EMILE

In this paper we set out to compare two unsupervised grammar induction systems: Alignment-Based Learning (ABL) and EMILE. Both are motivated from a different background and with a different goal. ABL starts out from the linguistic notion of substitutability [14] aiming to learn a maximum number of correct constituents. On the other hand, EMILE stems from a mathematically sound theory of substitution classes which makes it possible to prove that EMILE’s learned string language converges to the string language from which samples are taken. Both systems will be described briefly, followed by a theoretical comparison between the two. In addition to this, the two systems are applied to two different corpora, the English ATIS corpus and the Dutch OVIS corpus. The properties of the systems as described in the theoretical comparison are reflected in the results on the two corpora.