yuanxin112/morphbench-de

MorphBench — German Morphology Evaluation Suite

A benchmark for morphology-aware tokenizers and language models in German. Eight numbered tasks probe inflection, segmentation, derivation, compounding, affix semantics, and whole-word meaning. Each task is a config; difficulty/generalization tiers are splits.

from datasets import load_dataset
ds = load_dataset("yuanxin112/morphbench-de", "task3_derivation", split="test_main")

Tasks

task1_inflection

Produce the inflected form from a lemma + features. input interagieren + v;ind;pl;1;pst → output interagierten columns: lemma, feats, target, status

task2_deriv_segmentation — derivational segmentation

Split a derived word into base + affix. input Dreher → output drehen|er columns: word, segmentation, status

task3_derivation

Build the derived word from a base + affix. input Apotheke + pflichtig → output apothekenpflichtig columns: base, affix, target, status

task4_compound

Split a compound into its parts. input Freiheitsheld → output Freiheit|Held columns: word, segmentation, difficulty, source

task5_affix_function

Given a derived word, classify the semantic function of its affix (13 classes: negation, without, agent_person, make_become, …). input stimmlos → output without columns: word, function, freq, status

task6a_definition

Generate the dictionary gloss of a derived word. input Malocher → output Arbeitnehmer, der überwiegend körperlich hart … arbeitet columns: word, gloss, base, affix, function, derived_freq, base_exposure, status

task6b_definition_mcq

Multiple-choice version of 6a: pick the correct gloss among hard distractors. columns: word, base, affix, function, gold_gloss, distractors, pretrain_status, derived_freq, base_exposure

Splits — how items are bucketed

Every test split is defined by the pretraining exposure of the item's parts (frequency in the BabyLM-style German training corpus): seen = frequency above a threshold, unseen = frequency 0. The headline idea everywhere is the target is UNSEEN while its base/lemma IS seen (compositional generalization); memorization and rare/OOV/hardest are the seen / harder controls. train, dev are training / validation.

task1_inflection — by (lemma seen?) × (this form seen?):

split	condition
test (main)	lemma seen, form unseen
test_memorization	lemma seen, form seen
test_rare	lemma unseen, form unseen

task2_deriv_segmentation, task3_derivation — by (base seen?) × (derived seen?), where seen = corpus freq ≥ 1:

split	condition
test_main	base seen, derived unseen (compositional generalization)
test_memorization	base seen, derived seen
test_OOV	base unseen, derived unseen

task4_compound — by (compound word seen?) × (all components seen?), where compound seen = freq ≥ 1 and a component is seen = freq ≥ 5:

split	condition	what it tests
test_main	compound unseen, all components seen	compositional generalization
test_memorization	compound seen, all components seen	rote recall
test_rare	compound seen, some component rare	mixed
test_hardest	compound unseen, some component unseen	hardest

task6a_definition, task6b_definition_mcq — by the derived word's corpus frequency derived_freq, splitting the unseen case by base_exposure:

split	condition
memorization	derived_freq ≥ 20
rare	derived_freq 1–5
main	derived_freq = 0 and base_exposure ≥ 300 (base familiar)
main_oov	derived_freq = 0 and base_exposure < 300 (base also OOV)

task5_affix_function uses a plain train / dev / test split (no difficulty buckets). The status / pretrain_cell column on each row records the exposure cell directly, e.g. base_seen+derived_unseen, lem_seen+form_unseen.

Note: the German "seen" thresholds are lower than the English suite (derivation/ segmentation use freq ≥ 1; definition uses derived_freq ≥ 20, base_exposure ≥ 300).

Provenance

Built from German Wiktionary (via kaikki.org / wiktextract) and UniMorph. Companion lexical resource: yuanxin112/wiktionary-morph. English counterpart: yuanxin112/morphbench-en.

License

Derived from Wiktionary — CC BY-SA 4.0; attribute Wiktionary and its contributors.