Learning Morphology: Algorithms for the Identification of the Stem Changes

The aim of the current work is to create tools' for the automatic recognition of the Estonian stem changing rules'. The main problem consists in bringing together the ,fi)rmal classification,features available to the computer and classification based on human knowledge. This paper introduces two algorithms. First, in STLearn the supervised inductive learning technique is used to find out the suitable jeatures Jor automatic recognising of the stem changes. Two stem variants" can be bounded by more than one stem change. The second algorithm is created Jor the identifjdng the whole set of rules Jor stem pairs'. Current work is a part of a project based on the open model of language [Viks94] according to which all regular and productive phenomena of the natural language are represented by different types of rules and irregular phenomena are listed in small dictionaries exception lists. This approach gives opportunity to process the regular words not listed in dictionaries new derivatives, loan-words etc. Subsystem of morphology plays the central role in processing of the morphologically complex languages as the Estonian language is. The number of possible stem variants can strongly vary in Estonian: in some inflection types there are no stem variants at all, in some of them a word can have even five different regular stem variants. Current work presents tools ~br creating a formal description of the Estonian stem changing rules, starting from the pair of the stem variants. The Concise Morphological Dictionary of the Estonian (CMD) [Viks92] serves as a bases for current work and contains over 36 000 headwords, each of them has two stem variants on the averages. The principle types of changes are the following: 1. Stem-grade changes. Stem can occur either in a strong or a weak grade; the grades are differentiated first of all by phonetic quantity (2nd or 3rd degree of quantity marked by') that may be accompanied by various sound changes enfblding the medial sounds. For instance members of the stem pair h6ive-h'~ive are distinguished only by the different phonetic quantity; in case of couple aat2e'aal2e the rewriting rule b --+ p is concurrent with the phonetic quantity change. 2. Stem-end changes. Stem can appear either as a lemmatic stem or an inflection stem; stem variants are differentiated by changes enfolding the final sounds ( e.g. 'aadel-aadli, j'alg \~bot\-j'alga, sipelgas" \ant\-sipelga). 3. Secondary changes. These changes are conditioned by the certain context arising after either the stem-end or the stemgrade change (e.g. k'uppel \dome\ --~ * k'uppli --+ k'upli). About 20 % of stems stay changeless, mostly take place the stem-end or stem-grade changes or both at the same time. Formally the recognition of the stem change rules can be reduced to the classification task with string pairs as the objects to classify and possible rules of stem changes as the classes. System has to create class descriptions from the 'available' data: characters and their belongness to the sound classes. The important demand to the classification system is the linguistical