mlseg

A rule-based Malayalam surface segmenter. Given an inflected word, it splits it into a stem plus its case / number suffixes. It is the rough inverse companion to mlinflect (which generates inflected forms): mlseg takes one apart.

from mlseg import segment

segment("കുട്ടികൾ")        # ['കുട്ടി', 'കൾ']
segment("വീട്ടിൽ")          # ['വീട്', 'ഇൽ']
segment("പുസ്തകങ്ങളിൽ")  # ['പുസ്തകം', 'ങ്ങൾ', 'ഇൽ']
segment("മരം")             # ['മരം']  (no recognised suffix, returned whole)

What it does

segment(word) applies a small set of native-ratified surface rules: it strips a recognised case suffix (locative, sociative, directive, genitive) and a plural (കൾ / ങ്ങൾ), reverting the well-known geminate (…ട്ടി -> …ട്) and restoring chillu and anusvaram lemmas where the alternation is sound and reversible. Suffixes come back in their lemma spelling, in surface order (stem first, outermost suffix last).

Conservative by construction

This is a best-effort fallback for words a full morphological analyser cannot handle, not an analyser itself. It splits only on a rule whose surface-to-lemma realisation has been ratified by a native reviewer. If no rule matches, it returns the word whole ([word]); it never invents a split and never fabricates Malayalam (no guessed viramas, no fabricated stems). A minimum-stem floor blocks over-stripping of real words whose tail merely looks like a suffix (മകൾ "daughter", കോട്ട് "coat"). segment is idempotent on bare stems.

Limitations

There is no lexicon, so a proper name whose tail happens to match a suffix surface and clears the 2-akshara floor can be false-split, e.g. the name അനിൽ (Anil) becomes ['അൻ', 'ഇൽ']. Distinguishing such a name from a true locative needs a lexicon or NER. Callers that need certainty should prefer a full analyser; this fills the gap when one is unavailable.

Install

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

mlseg depends on mlnormalize for input normalisation, so geminate and chillu comparisons are byte-stable. It is on PyPI, so pip install mlseg pulls it in automatically.

License

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

Linguistic sources are credited in REFERENCES.md.