Evolutionary Language Acquisition

W. Cyre (USA)


Genetic Algorithms, Natural Language Processing, Knowledge Acquisition


This paper reports on a genetic algorithm for evolving a competent grammar from a small partial grammar using an unlabelled corpus of sentences in a language. Experiments were performed on three context-free grammars having up to 67 rules. Competence, recall, precision and ambiguity of the grammars were measured. Competence of the grammars evolved from a few percent to between 95% and 100%. Recall and precision for the evolved grammars were on the order of 70%. In the experiments, example grammars were used to generate training corpori of sentences. The initial grammars for evolution consisted of 33% to 50% of these rules. The evolved grammars rediscovered about half the lost rules. Somewhat specialized rules were constructed by the algorithm to replace the other lost rules.

Important Links:

Go Back