Sandhi-Morph-1.0

Malayalam word analyses: 921,531 training rows, plus five held-out evaluation sets. It accompanies the model Sandhi-1.0-8M.

Each row is one word (not a sentence) with its sandhi split, morphemes and their grammatical function, root lemma, grammatical features, and for some rows IPA and syllables.

Configs and splits

Config	Split	Rows	What
default	train	903,078	Training rows
default	dev	18,453	Development rows. No word appears in both train and dev
eval	loanword_stems	5,000	Words built on unseen loanword stems
eval	roots	2,761	Words built on unseen dictionary roots
eval	depth_stacks	1,549	Real words of 4–6 morphemes
eval	compound_pairs	414	Unseen real compounds
eval	tag_stacks	324	Unseen ending combinations
valid_analyses	train	24,839	Every valid reading of each eval word, for set-match scoring

The eval words are not in train or dev. Please do not train on them.

Training rows

Three tasks share one file, told apart by the task column:

task	Rows	Contains
<FULL>	729,407	The complete analysis
<SPLIT>	148,916	The sandhi split only
<G2P>	43,208	IPA and syllables only

The slice column groups rows by kind:

slice	Rows	Share	Kind
S1	402,482	43.7%	Generated inflection paradigms
S2	158,700	17.2%	Word forms attested in text
S3	179,831	19.5%	Compounds
S4	134,403	14.6%	Words with roots absent from the analyser's lexicon
S6	43,208	4.7%	Pronunciation (IPA, syllables)
S5	2,907	0.3%	Partial parses where the root is unanalysable

(Counts are for train and dev together.) Most <FULL> rows are shallow: 86.1% have two morphemes, 9.7% three, and about 2.1% four or more.

Columns

Column	Description
row_id	Row index in the source corpus. Not contiguous. Null in eval
input	The word
task	<FULL>, <SPLIT> or <G2P>
slice	S1–S6. In eval this is the set name
split	Sandhi split, pieces joined by +
ipa	IPA, where the row carries it
root_lemma	Root lemma
root_type	known, compound or unanalysable
pos	Part of speech
n_morphemes	Number of morphemes
confidence	high, medium or low
target	The full target as a JSON string: morphemes with underlying forms and features, root, features, IPA, syllables, confidence, alternatives
provenance	JSON string describing how the row was produced

target and provenance are JSON strings because their keys differ between slices. Parse them with json.loads.

import json
from datasets import load_dataset

train = load_dataset("nithinmanoj10/Sandhi-Morph-1.0", "default", split="train")
row = next(r for r in train if r["task"] == "<FULL>")
print(row["input"], "->", row["split"], "|", row["root_lemma"])
print(json.loads(row["target"])["features"])

loans = load_dataset("nithinmanoj10/Sandhi-Morph-1.0", "eval", split="loanword_stems")

Scoring

A Malayalam word form can serve several grammatical functions, so one word can have more than one valid analysis. Score against the set of valid analyses in valid_analyses (keyed by surface) rather than a single gold string, and report strict match too.

Notes

• Labels were produced automatically, not annotated by hand, so they carry the errors of the tools that generated them. This is most visible in depth_stacks.
• Words only: there are no sentences or context.
• There are no English meanings.

Licence and attribution

CC BY 4.0. Attribution to the sources it draws on:

• mlmorph (MIT) and Mlphon (MIT)
• IndicCorpV2, AI4Bharat (CC0 1.0)
• sangraha, AI4Bharat (CC BY 4.0)