The research, conducted by Changyong Niu, Hongde Liu, Senbin Zhu, Xingren Wang, Chenyuan He, and Yuxiang Jia of the School of Computer and Artificial Intelligence at Zhengzhou University in China, introduces a framework that jointly models the three core components of any metaphor: the source-domain word, the target-domain word, and the metaphorical context. In the classic structure of metaphor, the source is the concrete concept borrowed for comparison—say, “lion”—while the target is the abstract entity being described—say, “the general”—and the context is the surrounding text in which the mapping unfolds. Existing sentiment analysis models, the authors contend, typically focus on modeling only part of this triad, ignoring the interactions among the elements that determine whether “He is a lion in battle” conveys admiration or menace.

The team’s second criticism of prior work concerns the diversity of metaphor types themselves. Most computational metaphor research has been built around a single pattern of figurative mapping, which means models trained on one variety of metaphor often fail to generalize to others. A system tuned to noun-based metaphors may falter when the sentiment is carried by a verbal metaphor such as “the crisis swallowed the economy.” Because different metaphor types encode sentiment through different structural routes, a model that captures shared representations across types should, in principle, be more robust. Testing that hypothesis required data, and data was part of the problem: annotated resources that record not only the sentiment polarity of a metaphor but also its source, target, and type have been scarce, particularly for languages other than English.

To close that gap, the researchers constructed two new benchmark datasets. The first, named EMSA, is an English metaphor sentiment corpus; the second, CMSA, is its Chinese counterpart. Both datasets are annotated with the full complement of information the framework demands: the source-domain word, the target-domain word, the sentiment polarity of the expression, and the metaphor type. This rich labeling is what makes joint modeling possible in the first place, because a model cannot learn to weigh the interactions among source, target, and context unless training examples explicitly mark where each element begins and ends. The bilingual design also allows the team to probe whether their approach depends on the idiosyncrasies of one language or reflects something deeper about how figurative sentiment works.

On the modeling side, the researchers propose an architecture called MTCL-IT, which stands for Multi-Task Contrastive Learning combined with Instruction Tuning. The contrastive learning component is designed to sharpen the model’s internal representations by pulling together examples that share the same sentiment or metaphor structure while pushing apart those that differ. In a supervised contrastive setup, the model learns an embedding space in which semantically and emotionally similar metaphorical expressions cluster together, making the eventual classification decision easier and more reliable. Multi-task learning, meanwhile, allows the model to pursue several objectives simultaneously—identifying sentiment polarity, recognizing metaphor type, and grounding the source and target elements—so that the representations serving one task are enriched by the demands of the others.

The instruction tuning component brings the framework into the era of large language models. Rather than treating sentiment classification as a bare labeling problem, the approach reformulates the task as natural-language instructions, a strategy that has proven effective in related areas such as aspect-based sentiment analysis. The published work draws on a family of openly available language models in its experiments, including Llama-3.1-8B-Instruct, Qwen2.5, GLM-4, DeepSeek-V3, InternLM2.5, Flan-T5, and the BERT family of encoders, suggesting that the framework is designed to be backbone-agnostic: the three-element joint modeling philosophy can be layered onto different underlying architectures. Low-rank adaptation, a parameter-efficient fine-tuning technique known as LoRA, features among the tools used to adapt large models without retraining them from scratch.

The experimental results, the authors report, validate the framework from multiple perspectives. Across the English EMSA and Chinese CMSA datasets, the MTCL-IT model outperformed baselines in identifying the sentiment of metaphorical texts, and—perhaps more interestingly—the joint modeling approach revealed something about the pathway by which sentiment travels through a metaphor. By explicitly representing the source and target domains and their interaction with context, the model can trace how affective charge moves from the borrowed concept to the thing being described. When “steel” appears as a source domain, the model can register the connotations of hardness and resolve that the word carries, and observe how those connotations transfer to whatever target it modifies. This ability to expose the mechanism of sentiment transfer is a step beyond simple polarity prediction, moving toward genuine understanding of figurative language.

The dataset construction itself is methodologically noteworthy. The authors built the corpora using established procedures for linguistic metaphor identification, drawing on the widely cited MIP-style methodology developed by Gerard Steen and colleagues, and employed inter-annotator agreement measures rooted in Cohen’s kappa to verify that human labelers judged sentiment and metaphor structure consistently. Some of the annotation pipeline reflects the current moment in natural language processing: the study references human-LLM collaborative annotation, in which large language models propose labels that human annotators then verify, a hybrid strategy that has gained traction for reducing cost without sacrificing quality. The team’s broader research program also includes work on metaphor component identification using feedback-enhanced in-context learning, indicating that the new datasets and model are part of a sustained effort to make machines fluent in figurative language.

One of the more striking findings appears in an appendix analyzing sociological and cultural patterns in the Chinese CMSA dataset. The metaphors in that corpus predominantly involve mappings to the concept of “HUMAN,” and the choice of source-domain imagery differs sharply by gender. The male pronoun “He” is frequently associated with animal concepts—mouse, pig, eagle, snake, and lion—yielding both negative comparisons, such as calling someone as stupid as a pig, and positive ones, such as praising eyes as sharp as an eagle’s. The female pronoun “She,” by contrast, is overwhelmingly mapped to aesthetically charged imagery: angel, fairy, doll, sunshine, and flower. The distribution, the authors note, mirrors traditional cultural expectations in which male figures are associated with strength and competence while female figures are praised for beauty and kindness. That a sentiment analysis dataset can surface such patterns illustrates a broader point: computational models of figurative language are also instruments for studying culture, because the metaphors a community favors encode its values and biases.

The implications stretch well beyond linguistics departments. Sentiment analysis underpins real-world systems that monitor brand reputation, gauge public opinion during elections and public health crises, and detect distress in online communications. Metaphors saturate precisely the texts these systems read—social media posts, product reviews, political speeches—and research in marketing has already shown that metaphor-enabled sentiment analysis can change how consumer attitudes are measured. A model that misreads “this phone is a brick” as neutral commentary about construction materials will miscalculate customer satisfaction; one that understands the source-target mapping will correctly register frustration. The bilingual nature of the new resources matters too, since most figurative-language datasets have been English-only, and metaphor conventions vary across cultures in ways that a system trained solely on English data cannot capture.

The work also contributes to an ongoing debate about how large language models handle figurative meaning. Despite their fluency, LLMs remain brittle with metaphors, idiom, sarcasm, and other non-literal constructions, and a line of recent research has sought to scaffold metaphor reasoning with theory-guided prompts and chain-of-cognition prompting. The Zhengzhou team’s approach differs by insisting on structural completeness: rather than prompting a model to guess at the feeling of a sentence, they argue the model should first be given, or should learn to extract, the source, the target, and the context, and should be trained to weigh their interactions explicitly. The contrastive learning objective provides a complementary pressure, encouraging the model to organize metaphors in its embedding space according to the features that actually drive sentiment rather than surface similarity.

The paper passed through peer review quickly by the standards of the field—received in late March 2026, revised by mid-July, and accepted days later—before appearing in the journal in August 2026. The authors have released their data and code publicly through a GitHub repository, lowering the barrier for other groups to build on the datasets and to test the joint modeling framework against their own architectures. The work was supported by the Henan Provincial Science and Technology Research Project, a key program of the National Natural Science Foundation of China, and an open project of the Sichuan Provincial Key Laboratory of Philosophy and Social Science for Language Intelligence in Special Education.

Whether the three-element framework becomes a standard ingredient in sentiment systems remains to be seen, but the study makes a clear and testable claim: metaphorical sentiment is not a property of any single word or even of the sentence as an unanalyzed whole, but an emergent product of structured cross-domain mapping unfolding in context. Machines that ignore that structure will keep misreading the emotional temperature of human language—and as more of the world’s reading is done by algorithms, the cost of that misreading only grows.