Discriminative Optimization of String Similarity and Its Application to Biomedical Abbreviation Clustering

Many string similarity measures have been developed to deal with the variety of expressions in natural language texts. With the abundance of such measures, we should consider the choice of measures and its parameters to maximize the performance for a given task. During our preliminary experiment to find the best measure and its parameters for the task of clustering terms to improve our abbreviation dictionary in life science, we found that chemical names had different characteristics in their character sequences compared to other terms. Based on the observation, we experimented with four string similarity measures to test the hypothesis, "chemical names has a different morphology, thus computation of their similarity should be differed from that of other terms." The experimental results show that the edit distance is the best for chemical names, and that the discriminative application of string similarity methods to chemical and non-chemical names may be a simple but effective way to improve the performance of term clustering.