🏔️ Rumantsch Grischun Morphology & Disambiguation

Finite-State Morphological Pipeline with CRF Disambiguation for Rumantsch Grischun (ISO 639-3: roh)
Department of Computational Linguistics (Text Technology Group), University of Zurich & Lia Rumantscha — Version 1.2.0

Contextual Morphological Disambiguation

The finite-state analyzer first produces the possible morphological analyses for each word. Because many word forms are ambiguous, a statistical sequence model (Wapiti CRF) then uses the surrounding words to choose the analysis that best fits the sentence context.

The selected analysis gives the lemma and morphological features of the word. Alternative analyses proposed by the finite-state analyzer remain available for inspection.

When possible, the system selects among analyses licensed by the finite-state morphology. For words that cannot be analyzed by the finite-state system, the statistical model can provide a fallback prediction.
Examples

Result Columns

  • Token: The word form as it occurs in the input text.
  • Selected Lemma: The dictionary or citation form associated with the selected analysis.
  • Selected Analysis: The morphological analysis selected in context. For example, esser + Verb + ImpInd + 3P + Sg means esser, verb, imperfect indicative, third person singular.
  • Model Confidence: How strongly the statistical sequence model favors the selected analysis in this context. This is useful for identifying uncertain decisions, but it should not be interpreted as a probability that the analysis is linguistically correct.
  • Selection Method: How the final analysis was obtained:
    • FST-Licensed (CRF Rank 1): The contextual model’s preferred analysis is also licensed by the finite-state morphology.
    • FST-Licensed (CRF Rank 2/3): The model’s first choice was incompatible with the finite-state analyses, so the best higher-ranked licensed analysis was selected.
    • FST Candidate (Feature Match): An FST analysis was selected because its main morphological features agree with the statistical prediction, although there was no exact match.
    • CRF Fallback: The finite-state analyzer provided no suitable analysis, so the result comes from the statistical sequence model alone.
  • Alternative Analyses: Other morphological analyses produced by the finite-state analyzer for this word form but not selected in the current sentence.

Concrete Example: La vulp era puspè ina giada fomentada.

In isolation, era has several possible analyses, including the noun era ("garden bed"), the adverb era ("also"), and forms of the verb esser. In this sentence, the contextual model selects esser + Verb + ImpInd + 3P + Sg ("was"). The alternative analyses remain visible in the final column so that the disambiguation decision can be inspected.

The analyzer uses Xerox/Foma-style feature tags prefixed with +. Common elements:

Category Tag Examples Meaning
Part of Speech +Noun, +Verb, +Adj, +Adv, +Art, +Pron, +Prep, +Conj, +Subj, +Num, +Prop Primary syntactic category
Gender +Masc, +Fem, +MF Masculine, Feminine, Invariant/Either
Number +Sg, +Pl Singular, Plural
Person +1P, +2P, +3P First, Second, Third person
Verb Tense & Mood +PresInd, +ImpInd, +PastPart, +Inf, +Gerund, +Con, +Cond, +Impv Present Indicative, Imperfect, Past Participle, Infinitive, Gerund, Subjunctive/Conjunctive, Conditional, Imperative
Pronoun Type +Pers, +Poss, +Dem, +Indef, +Rel, +Refl, +Interrog Personal, Possessive, Demonstrative, Indefinite, Relative, Reflexive, Interrogative
Case / Function +Nom, +Acc, +Dat, +AccDat, +Aton, +Ton Nominative, Accusative, Dative, Weak/Atonic, Stressed/Tonic
Orthography +Apo, * prefix Elided with apostrophe (e.g. l', d'), Capitalized surface form

(See the Versions & Architecture tab for full linguistic documentation and references).