This thesis is concerned with unsupervised learning of
syntactic structure from plain text corpora by aligning
sentences. Based on Harris (1951) linguistic notion of
substitutability, sentences in a plain text corpus can be
compared to each other and those parts that have similar 
context and in addition can be substituted for each other
without resulting in ungrammatical sentences are 
considered to be possible constituents called hypotheses. 

A system that uses such an approach is ABL 
(Alignment-Based Learning) [van Zaanen, 2000]. Currently,
the main method used to align sentences and produce 
hypotheses is of such algorithmic complexity that ABL is 
feasible for small corpora only. 

This thesis explores the main topics of unsupervised 
learning of syntactic structure of natural language and 
introduces new algorithms based on suffix trees. The al-
gorithmic complexity of the new algorithms allow ABL to 
learn from large corpora as well. Furthermore, it shows 
that the suffix tree data structure has important advan-
tages in finding regularities in corpora, which is a 
fundamental issue in many approaches to unsupervised 
and semisupervised grammatical inference. The perfor-
mance of the new algorithms with respect to a learning 
task are tested within the ABL framework on corpora of 
various sizes: the English ATIS, the Dutch OVIS, 
and the English Wall Street Journal treebank. 

In conclusion, we observe that the time needed for 
alignment learning using the new algorithms as a 
function of the corpus size is indeed reduced in such an
extent that learning on large sized corpora with sys-
tems like ABL is feasible. However, the recall of the 
hypothesis constituents introduced by the new algorithms
is lower than that of the established algorithm and 
justifies further research.