Parameterless Information Extraction Using (k,l)-Contextual Tree Languages

Recently, several wrapper induction algorithms for structured documents have been introduced. They are based on contextual tree languages and learn from positive examples only but have the disadvantage that they need parameters. To obtain the optimal parameter setting, they use precision and recall. This goes in fact beyond learning from positive examples only. In this paper, a parameter estimation method for a wrapper based on (k; l)- contextual tree languages is introduced that is solely based on a few positive examples. Experiments show that the quality of the wrappers is very close to that of wrappers with the optimal parameter setting.