📄 SMOR: A German Computational Morphology Covering Derivation, Composition and Inflection

Summary

SMOR addresses a gap in German computational morphology. Earlier finite-state systems either depended mainly on large word lists (WordManager, GerTWOL) or handled only part of the problem: DMOR has no derivation component, and DeKo has no inflection. SMOR covers productive derivation, composition and inflection in a single analyser, built with the open SFST-Tools. The lexicon gives every stem and affix feature labels: entry type, word class, stem type (base/derivation/compounding), origin and complexity. Affixes use these labels to restrict which bases they attach to. The transducer is built in stages. Suffixes, then one prefix, then more suffixes are concatenated and checked by feature filters. Stems are then joined into compounds, inflectional endings are added, and morphophonological rules map the result to surface spelling. The authors tested it on 900 unseen word forms from the ELNC corpus, sampled across three frequency bands. SMOR analysed 85% of them, and about 96% of those analyses were correct. It runs at over 4000 words per second. This makes it a practical, freely available tool for German text, where ad hoc compounds and derivations are frequent.

Key points

• Implementation: about 1500 lines of SFST code, not counting the lexicon. SFST-Tools are released under the GNU GPL.
• Inflection: handled by continuation classes that were converted automatically from DMOR's code.
• Overall test results: on 900 ELNC word forms (LDC95T11, 300 per frequency band), 767 were analysed and 31 of those analyses were wrong.
• Results by frequency band: high-frequency forms were 96.7% analysed and 99.3% correct. Low-frequency forms were only 71% analysed.
• Corpus coverage: across all alphabetic-only types in ELNC, SMOR analysed 66.8%.
• Speed: more than 4000 words/s with fst-infl2 on a 750 MHz Sun Blade 1000.

Interesting ideas

• I1. Affixes as lexicon entries with selection features. Affixes are listed in the lexicon like ordinary morphemes, with features stating what bases they accept. Word formation then becomes plain concatenation followed by filtering. Grammar stays declarative and easy to extend, which suits richly derivational languages such as Ukrainian.
• I2. Three distinct stem types. The lexicon separates base, derivation and compounding stems, and default rules derive the regular variants (for example, the nominative singular also serving as a compounding stem). Only irregular stems need to be listed by hand, which keeps the lexicon small.
• I3. Rules composed in sequence. Phonological rules are chained by composition rather than run in parallel. This avoids interactions between rules and makes debugging much easier.
• I4. A pragmatic limit on prefixes. Allowing several prefixes, whose features must match distant suffixes, made the number of transducer states explode. The authors limited words to one prefix and backed the choice with a linguistic argument: apparent double prefixes are really lexicalised stems.
• I5. Frequency-stratified evaluation with tolerant gold labels. The test set is split into high, medium and long-tail bands, and any reasonable segmentation depth counts as correct. This shows where coverage drops (names, misspellings, rare formations) and avoids unproductive arguments over a single correct analysis.