📄 Two-level Description of Turkish Morphology

Summary

This short poster paper presents what the author calls the first full-scale computational account of Turkish word structure in the two-level morphology framework of Koskenniemi and Karttunen. Turkish is agglutinative: long chains of derivational and inflectional suffixes attach to roots, and they can move a word between nominal, verbal and adverbial categories. The morphotactics are finite-state but intricate, and phonological processes such as vowel harmony change how each suffix is pronounced and spelled. The author built the description in the PC-KIMMO environment on a lexicon of roughly 23,000 roots. Nearly all special cases and exceptions to the phonological and morphological rules are covered. Worked examples show that one surface word can receive several competing analyses. The analyzer has served as the morphological front end for LFG parsing, ATN parsing and semantic analysis of Turkish sentences. This makes it a reusable building block for Turkish NLP. The extracted text is fragmentary, so finer details of the rule set are not available here.

Key points

• Framework: two-level morphology (Koskenniemi 1983; Karttunen 1983), implemented in PC-KIMMO (Antworth 1990).
• Lexicon and coverage: about 23,000 root words, with almost all exceptions to the phonological and morphotactic rules handled.
• Ambiguity: a word like alınmış gets several analyses:
  • noun readings built on al ("red") and alın ("forehead");
  • a passive verb from al- ("take"), read either as a finite form or as a -mış adjective;
  • a verb alın- ("be offended").
• Lexical forms: suffixes are written with underspecified symbols (e.g. H in +yH, +sH), which surface rules realize according to harmony. The boynu example shows the stem boyun losing a vowel and being read either as possessive (3PS-POS) or accusative (ACC).
• Downstream use: the analyzer has served as the morphology module for LFG and ATN parsers and for semantic analysis.

Interesting ideas

• I1. Underspecified suffix vowels: writing suffix vowels as abstract symbols and letting rules fill them in keeps the lexicon compact. This could carry over to other harmony-rich or agglutinative languages, Ukrainian alternations included.
• I2. Separate lexicon and rule layers: keeping morphotactics (which suffix may follow which) apart from phonology lets each layer be developed and debugged on its own. This is useful whenever a morphological analyzer is built by hand.
• I3. Return every analysis: the analyzer outputs all readings, including ones that switch part of speech, and leaves the choice to later stages. This is a sound design for any pipeline where context, not morphology alone, settles ambiguity.
• I4. Make exceptions a priority: covering irregular cases such as vowel drop in boyun is what turns a toy analyzer into a full-scale tool. When planning a resource, budget for exceptions explicitly.
• I5. Test in real applications: the analyzer was plugged into several parsers rather than judged in isolation. Downstream use is a practical check of whether a morphology resource is actually usable.