On Learning Context-Free Grammars Using Skeletons

Abstract

In 1992, Sakakibara introduced a well-known approach for learning context-free grammars from positive samples of structural descriptions (skeletons). In particular, Sakakibarapsilas approach uses reversible tree automata construction algorithm RT. Here, we introduce a modification of the learning algorithm RT for reversible tree automata. With respect to n, where n is the sum of the sizes of the input skeletons, our modification for RT, called e_RT, needs O(n3) operations and achieves the storage space saving by a factor of O(n) over RT. Using our e_RT, we give an algorithm e_RC to learn reversible context-free grammars from positive samples of their structural descriptions. Furthermore, we modify e_RC to learn extended reversible context-free grammars from positive-only examples. Finally, we present summary of our experiments carried out to see how our results compare with those of Sakakibara, which also confirms our approach as efficient and useful.