Morphological Learning as Principled Argument
We develop a morphological learner that evaluates evidence supporting specific claims that a string of letters is a distributional meaningful unit. The distributional evidence is evaluated by selectional properties of morphs, while evidence towards meaning is modelled by looking at the relationship between stems and words. To assess a proposed affix, it gets a probability measure of meaning by comparing all the possible stems the affix occur with to the particular subset that also occur as words. Since for a stem to be a word counts as evidence towards its meaning, the ratio formed by taking stems that are words to the whole set of possible stems for an affix gives a predictive probability measure for the affix that measures the chance that it has combined with a meaningful stem. This measure, taken in conjunction with the selectional statistics of stems and affixes, provides a basis for deciding on the best morphological structure for a given word. The results for English show a combined precision and recall of 45.