A Comparison of Graph-Based and Statistical Metrics for Learning Domain Keywords

In this paper, we present a comparison of unsupervised and supervised methods for key-phrase extraction from a domain corpus. The experimented unsupervised methods employ individual statistical measures and graph-based measures while the supervised methods apply machine learning models that include combinations of these statistical and graph-based measures. Graph-based measures are applied on a graph that connects terms and compound expressions through conceptual relations and represents a whole corpus about a domain, rather than a single document. Using three datasets from different domains, we observed that supervised methods over-perform unsupervised ones. We also found that the graph-based measures Degree and Reachability generally over-perform (in the majority of the cases) the standard baseline TF-IDF and other graph-based measures while the co-occurrences based measure Pointwise Mutual Information over-performs all the other metrics, including the graph-based measures, when taken individually.