mlinflect

A rule-based Malayalam morphological synthesizer. It does forward morphological generation: given a root and grammatical features, it produces the inflected surface form (the counterpart to morphological analysis/segmentation).

from mlinflect import synthesize_noun, synthesize_verb, with_clitic, Case, Clitic, Number, VerbForm

synthesize_noun("മരം", Case.LOCATIVE).surface          # 'മരത്തിൽ'
synthesize_noun("മരം", Case.GENITIVE).surface          # 'മരത്തിന്റെ'
synthesize_noun("കുട്ടി", Case.GENITIVE).surface        # 'കുട്ടിയുടെ'
synthesize_noun("മരം", Case.NOMINATIVE, number=Number.PLURAL).surface  # 'മരങ്ങൾ'

with_clitic(synthesize_noun("കുട്ടി", Case.ACCUSATIVE), Clitic.UM).surface  # 'കുട്ടിയെയും'

synthesize_verb("ഓടുക", VerbForm.PAST).surface          # 'ഓടി'
synthesize_verb("കൊടുക്കുക", VerbForm.PAST).surface      # 'കൊടുത്തു'
synthesize_verb("ഓടുക", VerbForm.PRESENT_NEGATIVE).surface  # 'ഓടുന്നില്ല'

Why this exists

Existing Malayalam morphology tools are either copyleft (Apertium, libindic = GPL/AGPL) or, in the case of the one permissive generator (mlmorph, MIT), built on a GPL FST runtime. There is no permissive, dependency-clean, rule-based Malayalam synthesizer. mlinflect aims to fill that gap with a small pure-Python rule engine and no copyleft dependencies.

Design

• Declarative, provenance-tagged rules (mlinflect/rules.py): each rule cites the source it was drawn from and carries a verified flag that is True only when the form has been ratified by a native reviewer. Adding or correcting a paradigm is a data edit, not a code change.
• Inspectable results: every synthesize_noun(...) returns a SynthResult with the surface form, the morphemes that compose it, the stem_class, the provenance key, and verified. Feature combinations that are not yet encoded raise rather than return a silently wrong form.
• Akshara-correct joins: suffixes are represented matra-initial so concatenation produces correct conjuncts/vowel signs; the genitive uses the canonical nta form (NA + virama + RRA).

Status

Alpha. Eleven ending-conditioned noun classes across 11 cases, covering the major Malayalam noun shapes, with every encoded form native-ratified (verified=True); shapes outside the supported classes raise rather than guess. Five classes (am_neuter മരം, vowel_anuswara കലാം, i_vowel കുട്ടി/സ്ത്രീ, u_vowel പശു, ṭ_geminate വീട്) are complete in singular and plural; a_stem (അമ്മ) and the chillu classes (അവൻ, മകൾ, കാർ, കാൽ, തൂൺ) are singular-complete; their plurals are animacy-conditioned across the full paradigm (inanimate -കൾ/-ഉകൾ, human -മാർ/-ന്മാർ/-കാർ, animate -കൾ). Suppletive personal pronouns (ഞാൻ, നീ, അവർ, നാം, താൻ, ഇവൻ) are handled through an exception table rather than the rule engine. A derive_feminine helper builds a feminine lemma from a masculine base (എഴുത്തുകാരൻ → എഴുത്തുകാരി) before inflection. Includes differential object marking and a synthetic/colloquial register for the instrumental. Clitics (-ഉം, -ഓ, -തന്നെ) attach via with_clitic. Verbs (synthesize_verb) cover the finite forms: present, future, past (allomorphy by ending plus an irregular lexicon), negation, imperative, and a few moods. See LIMITATIONS.md for the precise gaps. Postpositions, stylistic variants, and verb aspects/participles/voice are future work.

Install

pip install mlinflect
# from source:
pip install -e ".[dev]"

License

Apache-2.0. See LICENSE and NOTICE. Contributions are accepted under Apache-2.0 §5 (inbound = outbound); no separate CLA is required.

Linguistic sources are credited in REFERENCES.md.