You are here
Extracting the Lowest-Frequency Words: Pitfalls and Possibilities
In a medical information extraction system, we use common word association techniques to extract side-effect-related terms. Many of these terms have a frequency of less than five. Standard word-association-based applications disregard the lowest-frequency words, and hence disregard useful information. We therefore devised an extraction system for the full word frequency range. This system computes the significance of association by the log-likelihood ratio and Fisher's exact test. The output of the system shows a recurrent, corpus-independent pattern in both recall and the number of significant words. We will explain these patterns by the statistical behavior of the lowest-frequency words. We used Dutch verb-particle combinations as a second and independent collocation extraction application to illustrate the generality of the observed phenomena. We will conclude that a) word-association-based extraction systems can be enhanced by also considering the lowest-frequency words, b) significance levels should not be fixed but adjusted for the optimal window size, c) hapax legomena, words occurring only once, should be disregarded a priori in the statistical analysis, and d) the distribution of the targets to extract should be considered in combination with the extraction method.