Statistical and Machine Learning Approaches for Network Analysis

Chapter 11

Hyponym Extraction Employing a Weighted Graph Kernel

Tim vor der Brück

Hyponyms are required for many applications in the area of natural language processing. Constructing a knowledge base with hyponyms manually requires a lot of work. Thus, many approaches were developed to harvest them automatically. However, most of them do not make use of deep semantic information but instead are based on surface representations. In this paper, we present a purely semantic approach, which is based on semantic networks. In the first step, hyponym hypotheses are extracted by application of deep semantic patterns. In the second step, the extracted hypotheses are validated employing a combined feature and graph kernel. Furthermore, a weighting scheme is described to weight the edges of the compared graphs. The graph kernel calculation is implemented in such a way that intermediate results are reused as much as possible to allow for a reasonable runtime. The evaluation shows that the graph kernel improves the evaluation results in contrast to a purely feature-based kernel. We expect that by optimizing the combination parameters for graph and feature kernels, a further improvement is possible.