11.10 Features for Hyponymy Extraction

Beside the graph kernel approach we also estimated the hypernymy hypothesis correctness by a set of features. The following feature are used:

• Pattern Application. A set of binary features. A pattern feature is set to one, if the hypothesis was extracted by this pattern, to zero otherwise.
• Correctness. In many cases the correctness can be estimated by looking on the hyponymy and hypernymy candidate alone. An automatic approach was devised that calculates a correctness estimation based on this assumption [25].
• Lexicon. The lexicon features determines a score based on the fact that if both hypernym and hyponym candidates (or the concepts associated to their base words) were contained in the lexicon or only one of them. This procedure is based on the fact that a lexicon-based hyponymy hypotheses validation is only fully possible, if both concepts are contained in the deep lexicon.
• Context. The context features investigates if both hyponym and hypernym candidates are connected in the semantic network to similar properties.
• Deep/Shallow. This binary feature is set to one if a hypotheses is only extracted by either deep or shallow extraction rules (0) or by both together (1).