\begin{abstract}
In this study, a method for discourse segmentation is introduced 
and applied to the Turkish language 
domain. This method is based on the idea that new segments start 
when related words do not occur frequently and discourse markers 
appear at sentence initial position.
A high correlation is observed  by means of statistical methods 
among human subjects in segmenting 
discourse. 
Decision trees for various combinations of 
criteria have been constructed by a Machine Learning technique.
The features that seem to be important are first mention of a word, 
tense change, time
shift, location shift, discourse markers and Centering Theory transitions.
Among these, it is observed that discourse segmentation 
is consistent with the predictions
of  both the Centering Theory  and the hypothesis that discourse markers
segment the text. 
Other than these, related words are observed to appear more
frequently inside a discourse segment than across a discourse segment boundary.
This feature is used in the construction of  the Automatic 
Turkish Discourse
Segmentation (ATDS) tool together with one of the most important computable features: 
discourse markers.
An important by-product of the project is the semantically 
related Turkish word 
network for nouns and verbs. The network is not only important
for ATDS but also for other studies since there is no thesaurus or semantically related words database
in Turkish. 
Relations are synonymy, hypernymy, antonymy, meronymy and coordinated words for nouns and synonymy, antonymy, hypernymy, 
entailment, cause and coordinated words for verbs.
Implementation of ATDS consists of two phases:
\begin{itemize}
\item gathering information about the most important parameters with the
help of C4.5.
\item developing a Neural Network for automatic segmentation that uses the most successful $4$ 
combinations of the first phase.
\end{itemize}
We benchmark our algorithm and compare it with the English discourse 
segmentation algorithms.
\end{abstract}