11.3 Drawbacks of Current Approaches

Pattern-based approaches are currently very popular for hyponymy extraction. But there is the problem that either recall or precision is poor. Consider a pattern collection containing among other things the pattern hyponym is a hypernym, which is very often applicable, but the precision of the extracted hypotheses is rather low. If this pattern is removed then the recall of the entire extraction process is seriously degraded. However, the precision probably increases. A second drawback is that the possibility of a pattern application is always a binary decision. A pattern is either applicable or not and therefore the pattern provides no quality estimate based on the semantic or syntactic sentence structure. Kernel-based approaches do not suffer these problems. However, for kernel approaches it is difficult to figure out, which sentence elements should be set as anchor points for the kernel. To compare all nouns with all other nouns would result in a very long runtime. In this work, the advantages of both approaches are combined. We extract hyponym hypotheses with quite general patterns and afterwards validate them with a graph kernel. In this way, the hypotheses extraction is very fast, and also the anchor points for the graph kernel are already determined and finally, we also get structure-based quality estimates.

Another drawback of current approaches is that they are usually based on words instead of concepts or word readings. Furthermore, a syntactic or surface representation is less suitable for semantic relation extraction. In MultiNet for instance, anthroponyms (person names) are already identified and represented as feature value structures. Moreover, the semantic network representation of several different syntactic representations often coincide if identical meaning is expressed. Thus, semantic patterns are more generally usable than patterns based on surface or syntactic representations.