Bootstrapping Syntax and Recursion Using Alginment-Based Learning

Abstract

This paper introduces a new type of unsupervised learning algorithm, based on the alignment of sentences and Harris’s (1951) notion of interchangeability. The algorithm is applied to an untagged, unstructured corpus of natural language sentences, resulting in a labelled, bracketed version of the corpus. Firstly, the algorithm aligns all sentences in the corpus in pairs, resulting in a partition of the sentences consisting of parts of the sentences that are similar in both sentences and parts that are dissimilar. This information is used to find (possibly overlapping) constituents. Next, the algorithm selects (nonoverlapping) constituents. Several instances of the algorithm are applied to the ATIS corpus (Marcus et al., 1993) and the OVIS corpus (Bonnema et al., 1997). Apart from the promising numerical results, the most striking result is that even the simplest algorithm based on alignment learns recursion.