Analysing the sample

Once a sample has been constructed it can be analysed. ICECUP includes a powerful statistical ‘knowledge discovery’ tool which can explore many combinations of variables quietly in the background. The tool generates independent hypotheses about the sample which are sent to a new ‘hypothesis panel’ in the sample viewer.

Analysing the sample

Every hypothesis is tested for statistical significance. If it is significant it is then scored according to a number of factors (below). The best hypotheses are then reported.

In this example, ICECUP has found two hypotheses which show

  1. that the transitivity of the verb phrase has an impact on the dependent variable, and
  2. that, in particular, copular and ‘trans’ cases are reliable predictors for the form being relative.

The following statistics are reported. Hypotheses are rated for utility, which is calculated as a combination of four factors: coverage, fitness, accuracy and swing. Once we have a measure of which hypotheses are “better” than others, the discovery algorithm can prioritise. More complicated hypotheses are considered only if they improve on a less complex one.

label summary explanation
+ve true positives number of cases correctly predicted by the hypothesis
-ve false positives number of cases incorrectly predicted by hypothesis
coverage proportion covered cases covered by hypothesis / total cases
fitness inverse accuracy positive examples / cases in target value
accuracy accuracy positive examples / cases covered by hypothesis
swing accuracy improvement accuracy - cases in target / total cases (scaled)
utility composite measure weighted product of coverage x fitness x… swing
What the hypothesis statistics mean

ICECUP can help you evaluate these hypotheses in terms of the cases they cover in the corpus.


This page last modified 12 June, 2013 by Survey Web Administrator.