The reusability of induced knowledge for the automatic semantic markup of taxonomic descriptions

To automatically convert legacy data of taxonomic descriptions into extensible markup language (XML) format, the authors designed a machine-learning-based approach. In this project three corpora of taxonomic descriptions were selected to prove the hypothesis that domain knowledge and conventions automatically induced from some semistructured corpora (i.e., base corpora) are useful to improve the markup performance of other less-structured, quite different corpora (i.e., evaluation corpora). The “structuredness” of the three corpora was carefully measured. Basing on the structuredness measures, two of the corpora were used as the base corpora and one as the evaluation corpus. Three series of experiments were carried out with the MARTT (markuper of taxonomic treatments) system the authors developed to evaluate the effectiveness of different methods of using the n-gram semantic class association rules, the element relative position probabilities, and a combination of the two types of knowledge mined from the automatically marked-up base corpora. The experimental results showed that the induced knowledge from the base corpora was more reliable than that learned from the training examples alone, and that the n-gram semantic class association rules were effective in improving the markup performance, especially on the elements with sparse training examples. The authors also identify a number of challenges for any automatic markup system using taxonomic descriptions.