Unified Benchmark for Metaphor Identification (UBMI)

The Unified Benchmark for Metaphor Identification (UBMI) gathers a wide range of metaphor and idiomaticity datasets into a single, harmonised format. The collected datasets were not all created for the automatic identification of metaphors: some come from psycholinguistic studies designed to collect human ratings or neurological measurements (e.g. CARD, JANK), some were originally created in other languages and later manually translated by their authors (e.g. WANG, BAMB), some are collections of metaphors obtained with concordancers (e.g. CCM), and others are full documents annotated to provide a standard annotation scheme (e.g. VUAC). Even within NLP, the datasets serve varied purposes — studying compositionality (GUT), evaluating metaphor generation (CHAK), analysing metaphoricity (DUNN), or disambiguating potentially idiomatic expressions.

Because of these differences, UBMI is not designed to merge all labelled phrases into one large dataset. What the collected datasets have in common is that they encode the metaphoric or idiomatic properties of expressions in context — or, conversely, their absence. UBMI re-encodes them with unified field names to facilitate experiments and comparisons across this diversity. In the open benchmark, only datasets with open licences are included.

---

Loading the dataset

Each source dataset is exposed as a separate subset (config), named <dataset>_<method>, where <method> is one of full, lexical or random:

• random and lexical subsets have train / validation / test splits.
• full subsets are the whole dataset merged into a single split named full (no train/test separation — useful as a reference set, but do not train on a full subset and then evaluate on the matching random/lexical test split).

from datasets import load_dataset, get_dataset_config_names
import json

# See every available subset name
configs = get_dataset_config_names("USERNAME/UBMI")

# Load one subset
ds = load_dataset("USERNAME/UBMI", "TroFi_random")
example = ds["test"][0]

# Dataset-specific / optional fields are stored as a JSON string — parse them:
extra = json.loads(example["additional_information"])
print(extra["lemma"], extra["pos"])

Note on additional_information. Across the benchmark this field holds ~80 different keys with mixed and sometimes incompatible types. To keep every subset loadable with a single, stable schema, it is stored as a JSON-encoded string. Decode it with json.loads(...) to recover the original nested object.

Config names vs. abbreviations. Config names use each dataset's internal id, which occasionally differs from the abbreviation in the table below (e.g. TSVET_AN for TSV_A, VUAC_MW for VUAC_BO, SE2013_ALL_WORDS for SE2013_ALL). Use get_dataset_config_names(...) to list the exact names.

---

One UBMI entry

Every example is stored with a unique tagged Potentially Metaphoric Expression (PME) per sentence, together with the minimal information needed for a binary classification task (the position of the expression in its context and its label), plus optional unified and dataset-specific fields kept under additional_information.

An example UBMI entry from the CARD_N dataset (Cardillo et al., 2010). The core fields (ref_id, dataset_id, id, context, pme, position, label, split) locate and label the PME, while additional_information preserves unified optional fields (e.g. lemma, pos, mscore, tenor, tenor_position, set_id, five-fold splits) and the original dataset-specific ratings.

Top-level fields (columns)

Field	Description
ref_id	Name of the original dataset (first author's abbreviated name if no other name exists). Several UBMI datasets may come from the same reference dataset.
dataset_id	Unique id for a UBMI dataset, often composed of the reference id and a split (random, lexical, etc.).
id	Entry id for a single example.
context	Text containing the PME (most often a sentence).
pme	The Potentially Metaphoric Expression: the labelled words in the context.
position	Offset [start, end] of the labelled expression within context.
label	Only metaphorical (m), literal (l) and other (o) labels appear here; finer labels are kept in original_label inside additional_information.
split	Whether the example belongs to the train, validation or test set (present in lexical / random subsets).
data_split	The sampling method — random, lexical, original split, etc. (present in lexical / random subsets).
additional_information	JSON-encoded string holding the fields below plus any non-unified, dataset-specific fields.

Inside additional_information (after json.loads)

Field	Description
lemma	Lemma(s) of the PME, obtained with spaCy.
pos	A single PoS, or a list of PoS for PMEs longer than one word.
pos_pattern	The PoS pattern of the PME.
five_folds	Train/validation/test membership across 5 folds (for cross-validation).
long_context	Additional context when the original dataset provides more than one sentence.
mscore	Metaphoricity or confidence score scaled to [0, 1].
tenor / tenor_lemma / tenor_pos / tenor_position	The topic word(s) of the metaphor and their lemma, PoS and position.
source_domain / target_domain	Conceptual source-to-target domain mapping.
original_label	The original (usually more fine-grained) label when it differs from the binary label.
set_id	Present when the sentence is grouped with others (usually pairs or triples).
other	Non-unified fields are stored here as original dataset information.

---

Splits

All datasets containing both literal and metaphorical labels are split into training (70%), development (10%) and test (20%) sets. Collections of metaphors without literal examples are not split. Splits come in two main sampling methods:

• Random splits — simple random shuffles of the data.
• Lexical splits — ensure a PME present in one split does not appear in the others, probing model generalisation.

Some datasets contain duplicated sentences (e.g. MIPVU datasets) or duplicated contexts where only the tagged expression changes (e.g. CHAK, GUT, TONG); for these, additional context-based splits keep the same context out of more than one split. Fixed 5-fold cross-validation splits are provided for both random and lexical sampling, and any original split shipped with a dataset is preserved.

---

UBMI datasets

The table below lists all datasets in the open benchmark, with the number of entries, number/percentage of metaphors, number of distinct expressions and contexts, the context span, and the sizes of the random-split train/dev/test sets. The Available? column indicates whether the dataset is already part of this release (✅ Available) or will be added shortly (⏳ Soon).

Dataset Name	Available?	N Entry	N Met (%)	N dist. Expr.	N dist. Ctxt.	Ctxt. Span	Train	Dev	Test
Nominal PME
CARDN	✅ Available	512	256 (50)	256	511	sentence	358	51	103
JANK	✅ Available	360	120 (33)	120	360	sentence	252	36	72
BAMB	⏳ Soon	115	115 (100)	91	84	noun	0	0	115
WANG	⏳ Soon	240	120 (50)	120	240	sentence	128	24	48
Fig-QA	✅ Available	8922	8922 (100)	6845	4454	sentence	0	0	8922
2x2Meta	⏳ Soon	691	329 (48)	547	203	paragraph	484	69	138
Adjectival PME
GUT	✅ Available	8591	4601 (54)	23	3479	noun	6014	859	1718
NEU	✅ Available	100	56 (56)	5	93	noun	70	10	20
TSVA	✅ Available	1945	979 (50)	687	1072	noun	1362	195	388
Verbal PME
CARDV	✅ Available	280	140 (50)	140	280	sentence	196	28	56
CHAK	✅ Available	468	312 (67)	355	155	sentence	328	47	93
MOH	✅ Available	1632	407 (25)	438	1630	sentence	1142	163	327
DUNN	✅ Available	60	40 (67)	20	60	sentence	0	0	100
TSVV	⏳ Soon	222	111 (50)	119	222	sentence	0	0	222
TroFi	✅ Available	3642	2098 (58)	50	3642	sentence	2549	364	729
NewsMet	⏳ Soon	1205	594 (49)	680	1205	sentence	844	120	241
Various
TONG	⏳ Soon	1428	655 (46)	1031	739	sentence	1000	142	286
CCM	✅ Available	8492	8492 (100)	590	8490	sentence	0	0	8492
GORD	⏳ Soon	1771	1771 (100)	574	1771	sentence	0	0	1771
ATT-META-S	⏳ Soon	500	500 (100)	500	500	sentence	0	0	500
PIEs
VNC	✅ Available	2568	2017 (79)	53	2534	3 sentences	1798	257	513
PVC	✅ Available	1348	878 (65)	23	1348	3 sentences	944	135	269
SE2013ALL	✅ Available	1969	1172 (60)	55	1939	3 sentences	1378	197	394
SE2013LEX	✅ Available	2371	1199 (51)	10	2339	3 sentences	1660	237	474
PIE	✅ Available	3025	1434 (47)	1608	2204	3 sentences	2118	303	604
MAD	✅ Available	4558	2190 (48)	443	4554	3 sentences	3191	456	911
MAGPIE	✅ Available	48395	36328 (75)	9462	47280	5 sentences	33876	4839	9680
PARSEME	⏳ Soon	1114	1114 (100)	530	958	sentence	0	0	1114
MIPVU
VUACBO	✅ Available	39223	20350 (52)	8722	39108	3 sentences	27456	3922	7845
VUACST1	✅ Available	23113	6554 (28)	2382	22894	3 sentences	16179	2311	4623
VUACST2	✅ Available	94807	15026 (16)	13509	93382	3 sentences	66365	9481	18961
NACEY	⏳ Soon	498	249 (50)	244	498	3 sentences	349	49	100
JUL	⏳ Soon	4268	2134 (50)	1649	4268	3 sentences	2988	426	854

---

Dataset descriptions by category

The datasets are grouped along two dimensions: the syntactic form of the labelled expression, and the granularity of the annotation. The first three groups are organised by syntactic form (nominal; adjectival and verbal; mixed), and the last two by the type of expression and annotation procedure (multi-word expressions; MIPVU). Each table below lists the UBMI datasets of that category with their original reference and original licence.

Nominal metaphors

Datasets designed for the study of the metaphorical usage of nouns or noun phrases (NPs). The figurative examples are direct metaphors, where the tenor NP and the vehicle NP are explicitly related within the sentence (e.g. Man is a wolf, an ocean of happiness). Most were created for psycholinguistic or neurolinguistic studies; two come from NLP projects.

Dataset	Reference	Original licence
CARDN	Cardillo et al. (2010); Cardillo et al. (2017)	CC BY-NC
JANK	Jankowiak (2020)	CC BY 4.0
BAMB	Bambini et al. (2014)	CC BY 4.0
WANG	Wang et al. — Chinese metaphor norms	Open
Fig-QA	Liu et al. (2022)	MIT License — copy, modification and redistribution allowed; must always include the original license: https://github.com/nightingal3/Fig-QA/blob/master/LICENSE
2x2Meta	Boisson et al. (2025)	CC BY-NC

Adjectival and verbal (indirect) metaphors

Indirect metaphors — the source concept is suggested by the words used metaphorically but not explicitly named (e.g. tasty metaphor, pour money, economy flourishes). They are the most frequent type of metaphor in natural text, and the datasets were created in majority for NLP studies. Adjectival datasets annotate the literal/metaphorical use of an adjective in the context of a noun (often adjective-noun pairs); verbal datasets are collections of full sentences.

Adjectival PME

Dataset	Reference	Original licence
GUT	Gutiérrez et al. (2016)	AFL-3.0
NEU	Turney et al. (2011); Assaf et al. (2013)	CC BY-NC 4.0
TSVA	Tsvetkov et al. (2014)	Redistribution allowed; license at https://github.com/ytsvetko/metaphor/blob/master/LICENSE.md

Verbal PME

Dataset	Reference	Original licence
CARDV	Cardillo et al. (2010); Cardillo et al. (2017)	CC BY-NC
CHAK	Chakrabarty et al. (2021)	No explicit licence; original data shared on GitHub: https://github.com/tuhinjubcse/MetaphorGenNAACL2021
MOH	Mohammad et al. (2016)	Redistribution allowed; license at https://saifmohammad.com/WebPages/metaphor.html
DUNN	Dunn (2014)	CC BY-SA 3.0
TSVV	Tsvetkov et al. (2014)	Redistribution allowed; license at https://github.com/ytsvetko/metaphor/blob/master/LICENSE.md
TroFi	Birke & Sarkar (2006)	CC BY-NC 4.0
NewsMet	Joseph et al. (2023)	Apache-2.0

Mixed metaphors

Datasets labelled with various types of metaphoric expressions, with (mostly) no syntactic constraint on the context or the PME, and expressions that may span more than one word. They include collections of metaphors focused on mapping analysis and datasets built for metaphor identification. (In the UBMI table these appear under the Various group.)

Dataset	Reference	Original licence
TONG	Tong et al. (2024)	CC BY 4.0
CCM	MacWhinney & Fromm (2014); Levin et al. (2014)	CC BY-NC 4.0
GORD	Gordon et al. (2015)	Redistribution allowed with required attribution. Any product, report, publication, presentation or document including or referencing the data must contain the text: "This effort contains or makes use of the IARPA-funded Metaphor Program USC/ISI annotated metaphorical language collection, release iarpa_metaphor_isi.edu_metaphor_corpus_20150403"
ATT-META-S	ATT-Meta Databank — Barnden et al.	Non-commercial research/instructional use; most pages of the databank may be freely copied for that purpose: https://www.cs.nmsu.edu/atmet/Databank/root.html.OLD

Multi-Word Expressions (MWE / PIEs)

Datasets where all labelled expressions are multi-word expressions. Most are Potentially Idiomatic Expressions (PIEs) — the same surface form occurs both literally and idiomatically — compiled and annotated for binary classification. Some target MWEs with specific patterns (noun compounds, verb-preposition compounds); others impose no restriction on the compound form. PARSEME 1.3 is an exception: it was created for verbal-MWE identification (negative cases are not annotated by default). (In the UBMI table these appear under the PIEs group.)

Dataset	Reference	Original licence
VNC	Cook et al. (2008)	No specific licence stated (no entry in the licence appendix)
PVC	Tu & Roth (2012)	No explicit licence; original dataset: https://cogcomp.seas.upenn.edu/page/resource_view/26
SE2013ALL	Korkontzelos et al. (2013)	CC BY-SA 4.0; redistribution authorised directly by the authors. Original download: https://www.inf.uni-hamburg.de/en/inst/ab/lt/resources/data.html
SE2013LEX	Korkontzelos et al. (2013)	CC BY-SA 4.0; redistribution authorised directly by the authors. Original download: https://www.inf.uni-hamburg.de/en/inst/ab/lt/resources/data.html
PIE	Haagsma et al. (2019)	CC BY 4.0
MAD	Tayyar Madabushi et al. (2021, 2022)	GPL-3.0
MAGPIE	Haagsma et al. (2020)	CC BY 4.0
PARSEME	Savary et al. (2023)	CC BY-SA 4.0 / CC BY-SA 3.0 / CC BY 4.0

MIPVU datasets

Datasets annotated with the MIPVU procedure (and its extensions): richly annotated documents in which each lexical unit is labelled metaphoric or literal. The benchmark retains three open-source MIPVU datasets (VUAC, NACEY, JUL); the VUAC is additionally sampled and reformatted for binary classification — balanced multi-word sampling (VUAC_BO) and the shared-task samplings (VUAC_ST1, VUAC_ST2). The licence shown for the VUAC samplings is that of the underlying VUAC corpus.

Dataset	Reference	Original licence
VUACBO	Boisson et al. (2023) sampling of Steen et al. (2010) — VUAC	CC BY-SA 3.0
VUACST1	Leong et al. (2020) sampling of Steen et al. (2010) — VUAC	CC BY-SA 3.0
VUACST2	Leong et al. (2020) sampling of Steen et al. (2010) — VUAC	CC BY-SA 3.0
NACEY	Nacey (2019)	CC0 1.0
JUL	Julich (2022)	CC BY-SA 4.0

---

Licensing

UBMI aggregates datasets released under different original licences (listed per dataset above); only datasets with open licences are included in this benchmark. The repository's license: other tag reflects this mixture — please consult and respect each dataset's individual licence for any redistribution or derivative use.

Citation

@misc{ubmi,
  title  = {Unified Benchmark for Metaphor Identification (UBMI)},
  author = {Joanne Boisson, Luis Espinosa-Anke and Jose Camacho-Collados},
  year   = {2026}
}