Medical Metaphors Corpus (MCC) Overview

• Medical Metaphors Corpus (MCC) is a comprehensive annotated dataset containing 792 instances of scientific metaphors from diverse registers in the biomedical domain.
• It aggregates metaphorical expressions from academic literature, news, social media, and crowdsourced contributions to support domain-specific metaphor detection and grading.
• The corpus employs both binary and graded annotations with reliability metrics, enabling nuanced evaluation of metaphorical language in complex medical discourse.

Searching arXiv for the MCC paper and closely related work to ground the article. arXiv search: "Medical Metaphors Corpus" The Medical Metaphors Corpus (MCC) is a comprehensive dataset of 792 annotated scientific conceptual metaphors spanning medical and biological domains. It aggregates metaphorical expressions from peer-reviewed literature, news media, social media discourse, and crowdsourced contributions, and encodes both binary metaphoricity judgments and perceived metaphoricity scores on a $0$–$7$ scale, together with source–target conceptual mappings. MCC is presented as the first annotated resource for computational scientific metaphor research and is intended to support metaphor detection benchmarking, quality-aware generation systems, and patient-centered communication tools (Lippolis et al., 11 Aug 2025).

1. Conceptual remit

Metaphor is treated in MCC as a fundamental cognitive mechanism that shapes scientific understanding, enabling the communication of complex concepts while potentially constraining paradigmatic thinking. The corpus is therefore not limited to rhetorical ornamentation; it targets scientific conceptual metaphors in which a biomedical target domain is structured via a distinct source domain.

The resource is explicitly positioned against an existing landscape in which metaphor detection resources primarily focus on general-domain text. MCC addresses the resulting gap for domain-specific applications by concentrating on medical and biological discourse and by preserving Conceptual Metaphor Theory-style source/target mappings. In practical terms, this means that the corpus is designed not only for classifying utterances as metaphorical or literal, but also for representing the underlying conceptual transfer that motivates the expression.

This framing has methodological consequences. Because the corpus includes both sentence-level context and conceptual mappings, it is suitable for work on detection, grading, interpretation, and controlled reuse of metaphors. A plausible implication is that MCC supports research agendas that require more than token-level figurativeness labels, particularly where scientific terminology and conventionalized figurative usage interact.

2. Construction and coverage

MCC was assembled through keyword-guided searches using terms such as “medical metaphor” and “biological metaphor” across scholarly databases and venues. Sources were included only if they provided explicit sentence-level metaphor examples with CMT-style source/target mappings and sufficient context. Sentences were extracted verbatim, provenance was tagged, and any pre-existing domain annotations were preserved (Lippolis et al., 11 Aug 2025).

The corpus combines four source families. From peer-reviewed literature, it includes 455 instances from “Metaphors in Medical Texts” (van 1997). From news media, it includes 19 metaphors from The Guardian (Camus 2009), 145 from major UK outlets (Kaikarytė 2020), and 35 from preventative-care reporting (Cheded 2022). From social media and patient narratives, it includes 27 items from a UK cancer forum (Semino 2018), 35 from a FORCE patient forum (Fereralda 2022), 50 from cancer-recovery interviews (Gibbs 2002), and 40 diabetes metaphors on Twitter (Sinnenberg 2018). Crowdsourced contributions add 16 user-submitted analogies filtered from the Metamia database.

At corpus level, MCC contains 792 total sentences, 82 distinct conceptual mappings (“metaphor types”), 38 unique source domains, and 24 unique target domains. Its register distribution by channel is 455 instances from academic literature, 199 from news, 152 from social media or patient narratives, and 16 from crowdsourced contributions.

This composition is notable because it is cross-register rather than single-genre. The same broad conceptual machinery can therefore be studied across expert scientific prose, journalistic mediation, and patient-facing or public discourse. This suggests that MCC is relevant not only for metaphor detection in scientific writing, but also for analyses of how metaphorical framing moves between technical and public registers.

3. Annotation protocol and reliability

MCC uses a two-part annotation scheme. First, each sentence receives a binary metaphoricity judgment, yes or no. Second, it receives a graded metaphoricity judgment on a $0$–$7$ scale, where $0$ is fully literal and $7$ is highly figurative. Each sentence was judged by at least two annotators, and overlaps were designed to compute reliability (Lippolis et al., 11 Aug 2025).

The annotation workforce comprised 42 annotators: 27 advanced humanities-informatics students and 15 professional linguists, all C1-level English. Before annotation, each annotator received a brief on Conceptual Metaphor Theory, definitions of “metaphor” versus “literal,” and examples illustrating borderline cases. Consistency checks were built into the workflow: literal judgments (“No”) were required to pair with a $0$ rating on the $0$–$7$ scale, and empty or inconsistent submissions were discarded.

Agreement is reported at several levels. Fleiss’ $\kappa$, described as the multi-rater generalization of Cohen’s $7$0, is $7$1, and average pairwise percent agreement is approximately $7$2. For two annotators, Cohen’s $7$3 is defined as

$7$4

where $7$5 is observed agreement and $7$6 is the expected chance agreement. On the graded scale, the average Pearson $7$7 between annotators is approximately $7$8, and Spearman $7$9 is also approximately $0$0, which the paper interprets as confirming a moderate but reliable graded signal.

The agreement profile indicates that MCC encodes interpretive difficulty rather than suppressing it. The presence of both binary and graded annotations is important here: the binary labels support standard classification settings, while the graded scores retain information about borderline cases, conventionalization, and disagreement intensity.

4. Instance structure and statistical profile

Each MCC instance comprises the source sentence, a binary label, a mean metaphoricity score on $0$1, and a source–target conceptual mapping following Conceptual Metaphor Theory. One example given is: “In theory, blocking any of the necessary steps for invasion listed in Table 7 could prevent tumor cell invasion.” Its annotation is binary: yes; metaphoricity mean: $0$2; mapping: TUMOR SPREAD (target) $0$3 MILITARY INVASION (source) (Lippolis et al., 11 Aug 2025).

The corpus-level binary decisions are distributed as follows: yes, 353 instances ($0$4); no, 305 ($0$5); ties, 134 ($0$6), where ties are cases with perfect $0$7 splits. On the graded side, ratings are heavily skewed to $0$8: approximately $0$9 of all ratings are $7$0, more than five times any other single rating. Mean ratings are $7$1 for items annotated as metaphor and $7$2 for items annotated as literal, yielding $7$3 points, which holds for $7$4 of sentence pairs. For sources with at least 40 items, source-wise means lie between $7$5 and $7$6, reflecting consistency across genres. The maximum rating variance is reported as $7$7 for boundary cases with high annotator disagreement.

MCC also formalizes disagreement. Binary disagreement is

$7$8

where $7$9 is the fraction of “yes” votes; $0$0 denotes full consensus and $0$1 a perfect split. Graded disagreement is measured as the standard deviation $0$2 of annotator scores. The correlation formula used, for example between annotator means and model predictions, is

$0$3

These statistics show that MCC is not a uniformly easy binary dataset. It contains clear-cut cases, fully literal cases, and a substantial boundary region. That structure is central to its value as an evaluation resource, because it exposes systems to the same uncertainty and conventionalization effects that human annotators encountered.

5. Baseline modeling and evaluation

The MCC paper evaluates zero-shot detection using five LLM APIs: GPT-4, o1-preview, o3-mini, DeepSeek, and Claude Opus 4. Performance is reported with standard macro-averaged yes/no metrics: $0$4 Without confidence weighting, o1-preview achieves accuracy $0$5, precision $0$6, F1 $0$7, and recall $0$8. Claude-4 is reported at accuracy $0$9, precision $7$0, F1 $7$1, and recall $7$2. o3-mini attains accuracy $7$3, precision $7$4, F1 $7$5, and recall $7$6. DeepSeek reaches accuracy $7$7, precision $7$8, F1 $7$9, and recall $0$0. GPT-4 is reported at accuracy $0$1, precision $0$2, F1 $0$3, and recall $0$4 (Lippolis et al., 11 Aug 2025).

The paper also introduces confidence-weighted metrics. For majority cases,

$0$5

and ties receive $0$6. This weighting yields a $0$7–$0$8 boost for all models; the paper gives o1-preview with weighted accuracy $0$9 as an example.

Error analysis identifies a consistent pattern. All LLMs exhibit high precision but lower recall, signalling a bias toward literal interpretations when lexical metaphoric cues are weak. Performance is best on high-consensus cases and drops steeply on boundary cases with domain-specific terminology or highly conventionalized metaphors. The authors interpret this as indicative of the need for domain-aware semantic representations beyond surface co-occurrence patterns.

The baseline results therefore function less as a leaderboard than as a characterization of task difficulty. MCC does not merely test whether a model can recognize conspicuous figurative markers; it tests whether a model can detect scientific metaphor under conditions of lexical subtlety, terminological specialization, and partial human disagreement.

6. Applications, adjacent corpora, and significance

The applications named for MCC are broad but technically specific. The corpus is intended for benchmarking domain-specific metaphor detection and grading, for fine-tuning or continued pre-training of LLMs for scientific figurative language, and for confidence-weighted evaluation frameworks in settings where human consensus is uneven. It is also proposed as a resource for patient-centered communication tools, including “metaphor menus” that adapt metaphor choice to individual preferences and emotional impact, for science-writing assistants that recommend context-appropriate, clarity-enhancing metaphors, and for contrastive studies of expert versus lay metaphor framing in public health messaging (Lippolis et al., 11 Aug 2025).

A related 2025 effort is “Dutch Metaphor Extraction from Cancer Patients' Interviews and Forum Data using LLMs and Human in the Loop,” which compiles HealthQuote.NL (Han et al., 9 Nov 2025). That corpus contains 130 verified metaphor instances drawn from 13 transcribed, anonymised oncology-setting interviews and a forum pilot of 100 posts, with annotation fields including original span, literal translation, type, source domain, function, speaker role, context sentence, and notes. Its human-in-the-loop workflow involves LLM extraction, independent verification by two expert annotators, and adjudication by a third senior linguist. Reported Cohen’s $0$0 values are approximately $0$1 on the interview set and approximately $0$2 on the forum pilot.

The comparison is instructive. MCC aggregates 792 sentence-level instances across academic, news, social, and crowdsourced registers and emphasizes binary plus graded metaphoricity together with source–target mappings. HealthQuote.NL focuses on Dutch cancer discourse and adds function-oriented labels such as explanation, coping, empowerment, relationship, prognosis, treatment, emotion, and humour. This suggests complementary trajectories within biomedical metaphor research: MCC supports computational scientific metaphor research at cross-register scale, while HealthQuote.NL concentrates on patient-produced metaphors and clinical communication settings.

By combining high-quality, cross-register examples with both binary and graded annotations, MCC fills a critical resource gap for computational and applied metaphor research in the biomedical sciences. Its importance lies not only in the size of the collection, but in the fact that it formalizes metaphor as a graded, contested, and domain-sensitive phenomenon within scientific and medically relevant discourse.