PaperlayerView exact revision on arXiv

Source-linked AI summary

AraMIP: Extending MIPVU Towards Metaphor Identification in Arabic

Mandar Marathe, Manar Ali, Sara Nabhani, Raia Abu Ahmad, Ibrahim Baroud, Omar Momen

arXiv:2609.17235v1cs.CL

TL;DR

Arabic lacks a systematic, standardized procedure for identifying metaphor that accounts for its linguistic and rhetorical properties. AraMIP adapts MIPVU with a second layer for Arabic rhetorical classification and applies it to a pilot corpus. The study reports that reliable Arabic figurative-language annotation is more difficult than in several other MIPVU adaptations and identifies sources of that difficulty.

  • Problem

    Arabic lacks a systematic, theoretically grounded procedure for identifying metaphors, despite existing studies and datasets with limitations in annotation transparency and scope.

    Source: §1 · p. 2, §2.3 · p. 3

  • Method

    AraMIP adapts MIPVU with a second annotation layer that maps metaphor-related words to Arabic rhetorical categories, and the authors apply it to a 300-sentence pilot corpus.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    The pilot’s IAA analysis finds Arabic figurative-language annotation more difficult than reported for several other MIPVU adaptations and identifies sources including limited annotation rounds, dictionary inconsistencies, and uncontrolled contextual materials.

    Source: §8 · p. 9

  • Takeaways & Limitations

    The pilot provides a diagnostic roadmap for future iterations of Arabic figurative-language annotation.

    Source: §8 · p. 9

    The 300-sentence MSA-only pilot is too small to characterize metaphor use across MSA reliably and does not establish how the guidelines generalize to dialectal Arabic.

    Source: Limitations · pp. 9–10

Abstract

from arXiv · show

Metaphor research has gained increasing attention due to its relevance to linguistic creativity, language use, cognitive processes, and related areas. While many efforts have been devoted to metaphor identification and annotation in English and other languages, Arabic remains under-resourced in this area. In this work, we propose the Arabic Metaphor Identification Procedure (AraMIP), a novel guideline for Arabic metaphor annotation. AraMIP builds on the widely used Metaphor Identification Procedure Vrije Universiteit (MIPVU) framework, incorporating adaptations that accounts for the language-specific properties of Arabic. We distinguish three major types of Arabic figurative language: Isti'ara (metaphor), kinaya (metonymy/indirect expression), and tashbih (simile), and annotate a pilot dataset of 300 sentences (5277 words). Our analysis reveals key challenges specific to Arabic, including morphological complexity, inconsistencies in dictionary sense ordering, and the absence of standardized contextual materials for annotators. This work contributes a first step toward standardized Arabic figurative instances and facilitates the development of larger annotated resources, thereby supporting future research on figurative language in Arabic.

1 Introduction

Metaphor research spans linguistic creativity, language use, cognition, and NLP, but Arabic lacks a standardized, replicable procedure that addresses its linguistic and rhetorical properties. AraMIP adapts MIPVU for Arabic and applies it to a 300-sentence pilot corpus, surfacing challenges for future annotation.

1 Introduction

  • Metaphors involve cross-domain mapping, and metaphorical framing can shape how people conceptualize abstract ideas and reason about them.In Conceptual Metaphor Theory, concrete source concepts map onto more abstract target domains; framing affects conceptualization and reasoning.Source: §1 · p. 1
  • Metaphor research includes annotation, language-model studies, and downstream NLP applications, while annotation resources have also been developed for languages such as Spanish, Chinese, and German.Arabic remains under-resourced in comparison to these broader efforts.Source: §1 · pp. 1–2
  • Arabic remains under-resourced for metaphor annotation, with no standardized, replicable procedure and challenges that existing frameworks do not address.These include rich morphology, root-based lexicography, and rhetorical categories that do not map directly onto categories developed for other languages.Source: §1 · p. 2
  • AraMIP adapts MIPVU into a two-stage Arabic annotation framework and is applied to 300 sentences (5277 words) from BAREC-10M.The work releases its guidelines and annotations and identifies challenges including dictionary coverage across the MSA/Classical Arabic boundary and root-to-derivative sense resolution.Source: §1 · p. 2

2 Background and Related Work

The background traces metaphor from competing theoretical accounts to operational identification procedures, then reviews Arabic studies and identifies a gap in systematic, theoretically grounded annotation. It motivates adapting MIP for Arabic rather than assuming existing categories and resources transfer directly.

2.1 Metaphor Theories

  • Metaphor theories range from Aristotle’s substitution view and Black’s interaction account to Searle’s distinction between sentence and speaker meaning and CMT’s domain mapping.CMT describes understanding one domain, typically abstract, in terms of another, typically concrete.Source: §2.1 · p. 2

2.2 Operationalizing Metaphors

  • MIP operationalizes metaphor identification through a lexical unit’s contextual meaning contrasting with a more basic meaning that can be understood in comparison.MIPVU extends MIP, and MIP-based procedures have supported annotated corpora in English and other languages; no systematic MIP adaptation for Arabic had been carried out at writing time.Source: §2.2 · p. 3

2.3 Related Work

  • Arabic metaphor studies include a 1,000-sentence book-review corpus, but its annotation procedure is unspecified; other datasets also have collection or language-source limitations.The related work includes Arabic writing-assistance and figurative-speech datasets, while the paper identifies a gap in systematic, theoretically grounded Arabic metaphor identification validated across languages.Source: §2.3 · p. 3

3 AraMIP: Annotation Framework

AraMIP combines MIPVU’s lexical-unit metaphor identification with a second layer grounded in Arabic rhetorical categories. The resulting annotations record both word-level metaphor labels and rhetorical-construction spans, including conventionality.

3 AraMIP: Annotation Framework

  • AraMIP retains MIPVU’s core annotation categories and decision principles while adding an Arabic-rhetoric layer to describe the rhetorical nature of identified expressions.The two stages are lexical-unit-level identification followed by rhetorical classification.Source: §3 · p. 3

3.1 Stage 1: MIPVU Annotation

  • Stage 1 follows MIPVU to identify metaphor-related words by comparing contextual and basic meanings, using an Arabic dictionary to determine basic meanings.MIPVU codes include direct, indirect, and implicit metaphor, MFlag for simile markers, and MRW-negative; WIDLII and DFMA cover uncertain or indiscernible cases.Source: §3.1 · pp. 3–4

3.2 Stage 2: Arabic Rhetorical Classification

  • Stage 2 maps non-MRW-negative words to Arabic rhetorical constructions, which often—but not invariably—align direct metaphors with tashbih and indirect metaphors with isti‘ara or kinaya.Both basic and contextual meanings are possible in kinaya, but not in isti‘ara.Source: §3.2 · p. 4
  • Stage 2 annotates complete rhetorical-construction spans, allowing multi-word and overlapping spans rather than limiting annotation to individual lexical units.MFlag words belong to the tashbih span they introduce, while implicit MRWs are not annotated in Stage 2.Source: §3.2 · p. 4
  • Stage 2 records each rhetorical construction’s conventionality on a five-point scale, which Stage 1 does not capture.The corpus therefore combines lexical-unit-level metaphor annotations with span-level rhetorical annotations grounded in Arabic rhetorical tradition.Source: §3.2 · p. 4

4 Corpus Selection

The pilot corpus samples diverse MSA sentences from BAREC-10M, aiming for coverage across text categories, domains, and readership groups. It contains 300 sentences totaling 5,277 words.

4 Corpus Selection

  • The study selects MSA text to include different genres and domains, using BAREC-10M as a diverse source with additional annotation layers.BAREC-10M contains about 10 million words and labels domain, readership group, and text category.Source: §4 · p. 4
  • The pilot uses 300 sentences and 5,277 words sampled from BAREC-10M, with 50 random sentences from each text category and sentence lengths restricted to 8–30 words.Sampling sought equal distribution across domains and readership groups when possible; average sentence length was 17.58 words.Source: §4 · pp. 4–5

5 Annotation Pilot Study

The pilot used trained annotators to independently label Arabic sentences, measured agreement across word- and span-level stages, and curated the resulting annotations through majority agreement and targeted second-round review. Curation left very few unresolved disagreements, while span curation required matching labels and exact or contained boundaries.

5.1 Annotation Setup

  • Six annotators—five native Arabic speakers and one L2 speaker—trained together on 10 sentences before independently annotating the 300-sentence sample, with each sentence assigned to three annotators.The sentences were partitioned to keep text-domain distributions approximately equal; annotation used INCEpTION.Source: §5.1 · p. 5

5.2 Inter-annotator Agreement

  • Stage 1 agreement measured MRW-positive versus MRW-negative status and full MRW labels using Fleiss’ κ, reporting average, weighted-average, and per-label scores.Table 1 presents agreement for the word-level annotation layer.Source: §5.2 · p. 5, Table 1 · p. 5
  • Stage 2 agreement assessed exact-match F1, IoU-based F1 at a 0.5 threshold, and Soft γ under three span-versus-label weighting settings.Exact-match F1 required both span boundaries and label to match; scores were averaged across annotator pairs.Source: §5.2 · p. 5, Table 2 · p. 5

5.3 Automatic Curation

  • After the second annotation round, curation yielded 90.49% clear agreement, 9.48% majority agreement cases (75.8% are MRW-Negative), and 0.04% disagreement cases. Tables 3 and 4 describe the annotated dataset before and after curation.Source: §5.3 · p. 6
  • Stage 2 curation retained spans with matching labels and exact or fully contained boundaries, excluding mere partial overlaps and averaging annotators’ conventionality scores.Tables 5 and 6 describe span annotations before and after curation; Figure 8 reports the lowest conventionality scores.Source: §5.3 · p. 6, Table 5 · p. 6, Table 6 · p. 6

6 Disagreement Analysis

Disagreements were concentrated in deciding whether words or spans were figurative, especially at the metaphor-related versus negative boundary. The analysis also identifies recurring lexical, register, expression-level, dictionary, and cultural factors behind those decisions.

Stage 1 Disagreements

  • Stage 1’s dominant confusion was MRW-indirect versus MRW-negative (812 pairs), indicating that positive-versus-negative status was the central difficulty.MRW-direct versus MRW-negative occurred 51 times and WIDLII versus MRW-negative 38 times; within positive labels, MRW-direct versus MRW-indirect was most frequent at 31 pairs.Source: §6 · p. 7
  • Stage 1 disagreements included ambiguity over whether a word’s senses constitute distinct domains, as illustrated by “step” in a self-appreciation context.The analysis identifies polysemy and domain-boundary ambiguity as one recurring source.Source: §6 · p. 7
  • Classical or poetic passages could require specialized knowledge of Classical Arabic or external exegesis, contributing to disagreement over readings such as “blazes.”The authors describe this as a probable source of disagreement.Source: §6 · p. 7
  • When figurativity attaches to a multiword expression, annotators may disagree about which constituent token or tokens carry the figurative meaning.The analysis identifies multiword expressions as a recurring disagreement category.Source: §6 · p. 7
  • Other reported sources were missing or unattested dictionary entries and interpretations dependent on annotators’ cultural or religious background.The authors note that an absent dictionary item may reflect an error in the original corpus, while a literal-versus-figurative reading of “the doors of God” may depend on religious views.Source: §6 · p. 7

Stage 2 Disagreements

  • Of 136 curated Stage 2 spans, 45 had full agreement, 91 were resolved by majority vote, and four complete disagreements were excluded.The most frequent disagreement was labeled versus unlabeled for isti‘āra and kināya; among labeled cases, isti‘āra versus kināya accounted for 44 of 55 disagreements, while tashbīh showed comparatively high agreement, likely due to its explicit comparator.Source: §6 · p. 7

7 Discussion

The discussion identifies dictionary coverage, Arabic root-based morphology, inconsistent sense ordering, sample composition, and annotation practices as constraints on AraMIP’s pilot findings. It also reports that annotator backgrounds showed no consistent pattern of higher or lower agreement.

7.1 Dictionary-related Challenges

  • Classical-register words, including Quranic and traditional-poetry terms, were missing or inadequately treated in the selected modern dictionary, creating gaps in basic-meaning retrieval.The authors suggest broader approved reference resources or explicit fallback procedures for dictionary omissions.Source: §7.1 · p. 7
  • Arabic root-based lookup can mismatch a root’s primary sense with a derived form’s conventional meaning, while MIPVU leaves annotators to judge the transfer.The authors identify explicit criteria for root-to-derivative meaning transfer as a future adaptation need.Source: §7.1 · pp. 7–8

7.3 Inter-Annotator Agreement

  • Figure 4 presents pairwise confusion matrices for Stage 1 MRW labels and Stage 2 rhetorical labels; “none” marks unassigned labels.Source: Figure 4 · p. 8

7.1 Dictionary-related Challenges

  • The dictionary’s stated general-to-specific sense ordering was inconsistently applied across entries, making basic-sense selection ambiguous.The authors suggest more reliable, empirically validated resources or supplementary procedures that do not rely on ordering alone.Source: §7.1 · p. 8

7.2 Distribution of Metaphor-Related Words

  • The low proportion of MRW-positive labels was expected in a sample containing denotative genres, but the small sample cannot represent the full range of MSA text types.The authors call for larger annotations across a more balanced domain selection to characterize figurative density more reliably.Source: §7.2 · p. 8

7.3 Inter-Annotator Agreement

  • Pilot inter-annotator agreement scores were below those reported for other MIPVU studies, including Fleiss’ κ 0.84 for the original VUAMC.More than 90% of examples received annotator agreement on the negative label, raising chance agreement and potentially keeping Fleiss’ κ low despite high observed agreement.Source: §7.3 · pp. 8–9
  • Limited annotation rounds left some rules inconsistently interpreted, including whether an MFlag-less direct metaphor should receive MRW-direct or MRW-indirect.The authors note that further discussion rounds would likely resolve the disagreement over the “Ahmad is a lion on the field” example.Source: §7.3 · p. 9
  • Annotators lacked shared contextual materials and could consult different external resources, introducing systematic variance in how they interpreted tokens.The annotator team also varied in discipline and Arabic variety, but comparisons found similar average IAA across backgrounds and no consistent agreement pattern by dialect.Source: §7.3 · p. 9

8 Conclusion

AraMIP is presented as a systematic, two-stage procedure for identifying and classifying metaphor-related expressions in Arabic, applied to a 300-sentence pilot. The pilot identifies annotation challenges and points to concrete revisions for future iterations.

8 Conclusion

  • AraMIP adds span-level classification into isti‘āra, kināya, and tashbīh, plus a five-point conventionality rating, to MIPVU’s lexical-unit-level identification.Source: §8 · p. 9
  • The pilot identifies five sources of annotation difficulty: one annotation round, dictionary coverage and sense-ordering inconsistencies, unresolved root-to-derivative meaning transfer, and uncontrolled contextual-resource access.The authors propose disagreement-informed guideline revisions, standardized contextual materials, and explicit root-based sense-lookup criteria for future iterations.Source: §8 · p. 9
  • The pilot samples 5,277 words from BAREC-10M across a balanced range of domains, readership levels, and text categories, and reports that reliably annotating Arabic figurative language is more difficult than in several other MIPVU adaptations.Source: §8 · p. 9
  • The authors release the pilot corpus and guidelines to support future work scaling to larger corpora, extending to dialectal Arabic, and training detection models.Source: §8 · p. 9

Limitations

The pilot’s size and MSA-only scope limit what it can establish, while dictionary and annotation-process choices leave further constraints. The authors report no significant risks from annotating existing data without potentially hurtful content.

Limitations

  • The 300-sentence (5,277-word) pilot is too small to characterize metaphor use across MSA reliably.Source: Limitations · pp. 9–10
  • AraMIP was applied only to MSA; dialectal generalization remains open, and one modern dictionary leaves Classical Arabic coverage and root-to-derivative meaning transfer unresolved.The authors also report inconsistent sense ordering, no iterative guideline refinement, and no standardized contextual or exegetical materials.Source: Limitations · p. 10

A Risks and Ethical Considerations

B MIPVU Procedure

The appendices describe the MIPVU procedure, the annotation interface, materials used, and annotator recruitment and training.

B MIPVU Procedure

  • Steen et al.’s standard procedure for identifying metaphors in written and spoken discourse is presented in Figure 5.Source: B MIPVU Procedure · p. 12

C Annotation Tool

  • The INCEpTION interface shows word-level MIPVU coding in Stage 1 and span-level rhetorical annotation in Stage 2.Source: C Annotation Tool · p. 12

D Scientific Artifacts

G Recruitment Consent

  • The six author-annotators received AraMIP training and practice, were recruited for annotation ability, consented freely, and received no financial compensation.Source: G Recruitment Consent · p. 12

H IAA Results Across Disciplines and Dialects

The section reports annotator backgrounds and pairwise Cohen’s κ agreement, summarized both by annotator and pair.

H IAA Results Across Disciplines and Dialects

Loading 2609.17235v1…