Combination of estimation algorithms and grammatical inference techniques to learn Stochastic Context-Free Grammars

Some of the most widely-known methods to obtain Stochastic Context-Free Grammars (SCFGs) are based on estimation algorithms. All of these algorithms maximize a certain criterion function from a training sample by using gradient descendent techniques. In this optimization process, the obtaining of the initial SCFGs is an important factor, given that it affects the convergence process and the maximum which can be achieved. Here, we show experimentally how the results can be improved in cases when structural information about the task is inductively incorporated into the initial SCFGs. In this work, we present a stochastic version of the well-known Sakakibara algorithm in order to learn these initial SCFGs. Finally, an experimental study on part of the Wall Street Journal corpus was carried out.