Compared to corpora-based machine translation methods, rule-based methods have
deficiencies, which make them unattractive for the researchers of this field. The first problem is
that these methods are language dependent. Rule-based methods require the syntactic information
about source and target languages. On the other hand, in many cases, especially for proverbs and
specific expressions, syntactic rules are not useful anymore. In such cases, the use of examplebased
approaches is inevitable. In this work, we propose and integrate a set of novel schemes to
introduce a new translation system, called BORNA. First a grammar induction method based on
the Expectation Maximization (EM) algorithm is proposed. After representing the extracted
knowledge in the form of a set of nested finite automata, a recursive model is proposed, which
uses a combination of rule and example based techniques. In the translation phase, through a
hierarchical chunking process, the input sentence is divided into a set of phrases. Each phrase is
searched in the corpus of examples. If the phrase is found, it will not be chunked anymore.
Otherwise, the phrase is divided into smaller sub-phrases. The simulation results show that
BORNA outperforms its counterparts, significantly. Compared to PARS, Frengly and Google
translators, BORNA receives the highest Bleu scores for its translations, while it results in the
minimum values for different error measures, including PER, TER and WER.