Word Formation in Computational Linguistics

Where does word formation information play a role in computational linguistics? Think of information retrieval where the parts of a word might contain the important information. If we want to find information in German on relaxation techniques, we need to look for the verb entspannen 'to relax' as well as for the derived noun Entspannung 'relaxation' and compounds containing Entspannung, such as Entspannungsantwort 'relaxation response', Tiefenentspannung 'deep relaxation' or Entspannungsübung 'relaxation exercise' etc. Another example are Text-To-Speech systems where the structure of a word can tell us where the stress goes. Consider English words containing so-called neoclassical affixes (affixes of Latin or Greek etymology). Some of these affixes, such as –ation influence the stress of a word: re'lax vs. relax'ation. Even such seemingly 'basic' components as part-of-speech taggers can (and often do) use word formation information, if only as heuristics. If a tagger encounters an unknown English word ending in the letters , for example, it can guess that this must be an adjective. Word formation information is thus important for many computational linguistics applications. But why do we need a word formation component? There are large machine-readable lexicons available today. Why can't we just use these? First, we have to distinguish between computational linguistic applications (by application we mean 'higher' systems like machine translation systems, text understanding systems etc. that include a number of components) that deal with a fixed text/set of texts (which can, of course, work with a finite lexicon) and applications that deal with unseen text. These applications make use of basic components such as taggers, lemmatizers / stemmers, parsers etc. Often these components (and accordingly the application) – no matter