yuanxin112/morphbench-en

MorphBench — English Morphology Evaluation Suite

A benchmark for morphology-aware tokenizers and language models in English. Eight numbered tasks probe inflection, segmentation, derivation, affix semantics, and whole-word meaning. Each task is a config; difficulty/generalization tiers are splits. Related tasks are paired (a = generation / from word, b = recognition / from meaning).

from datasets import load_dataset
ds = load_dataset("yuanxin112/morphbench-en", "task3a_derivation", split="test_main")

Tasks

task1_inflection

Produce the inflected form from a lemma + UniMorph features. input goggle + V;V.PTCP;PST → output goggled columns: lemma, feats, target, infl_type, status

task2_segmentation

Split a word into its morphemes. input creekline → output creek line (space-separated morphemes) columns: word, segmentation, deriv_type, affix_position, status

task3a_derivation — generation

Build the derived word from a base + affix. input pledge + in- → output impledge columns: base, affix, target, deriv_type, affix_position, status

task3b_derivation_mcq — recognition

Multiple-choice version of 3a: pick the correct derived form from 4 options (with demo pairs of the same affix). Easier recognition counterpart to 3a's free generation. columns: query_base, query_derived, query_affix, query_pos, affix_subset, pretrain_cell, options, correct, demos

task4a_affix_function — from the word

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

task4b_affix_function_paraphrase — from the meaning

The same 13-way classification, but the input is a paraphrased meaning instead of the word, so the surface affix string is removed. This is the control for 4a — it tests whether the model understands the affix's meaning rather than reading its spelling. input "lacking practical skill" → output without columns: meaning, function, freq

Models score ~98% on 4a but collapse to ~22% on 4b — i.e. 4a is solved almost entirely by the surface affix cue.

task5a_definition — generation

Generate the dictionary gloss of a derived word. input steadfastness → output Loyalty in the face of trouble and difficulty columns: word, gloss, base, affix, function, derived_freq, base_exposure, status

task5b_definition_mcq — recognition

Multiple-choice version of 5a: pick the correct gloss among hard distractors. columns: word, base, affix, function, gold_gloss, distractors, pretrain_status, derived_freq, base_exposure

Task pairs

Pair	a	b
3	derivation — generate the form	derivation — recognize (4-choice)
4	affix function — from the word	affix function — from meaning only (surface-cue control)
5	definition — generate the gloss	definition — recognize (4-choice)

Splits — how items are bucketed

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

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

split	condition	what it tests
test (main)	lemma seen, form unseen	inflect a known lemma into an unseen form
test_memorization	lemma seen, form seen	recall a seen form
test_rare	lemma unseen, form unseen	hardest — neither seen

task2_segmentation, task3a_derivation — by (base seen?) × (derived word seen?), where seen = corpus freq ≥ 10:

split	condition	what it tests
test_main	base seen, derived unseen	compositional generalization
test_memorization	base seen, derived seen	memorization
test_rare	base unseen, derived unseen	rare / OOV

(The base_unseen + derived_seen cell is dropped as too rare.)

task5a_definition, task5b_definition_mcq — by the derived word's corpus frequency derived_freq, splitting the unseen case further by base_exposure (how attested the base is):

split	condition	what it tests
memorization	derived_freq ≥ 50	recall the gloss of a frequent word
rare	derived_freq 1–5	word seen only a few times
main	derived_freq = 0 and base_exposure ≥ 500	define an unseen derived word from a familiar base
main_oov	derived_freq = 0 and base_exposure < 500	hardest — derived word unseen and base barely seen

task3b_derivation_mcq uses dev / test; task4a_affix_function and task4b_affix_function_paraphrase use train / dev / test (no difficulty buckets — each row still carries an exposure status where applicable).

The status / pretrain_cell column on each row records the exposure cell directly, e.g. base_seen+derived_unseen, lem_seen+form_unseen.

Provenance

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

License

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