Enhancing Hyperbole and Metaphor Detection with Their Bidirectional Dynamic Interaction and Emotion Knowledge¶

Conference: ACL 2025 (acl-long.23)
arXiv: 2506.15504
Code: None
Area: Others
Keywords: hyperbole detection, metaphor detection, emotion guidance, bidirectional dynamic interaction, LLM reasoning

TL;DR¶

This paper proposes the EmoBi framework, which utilizes a three-stage prompting process consisting of emotion analysis, emotion-guided domain mapping, and bidirectional dynamic interaction. By leveraging LLMs to uncover emotional cues behind hyperbole and metaphor as well as their mutually reinforcing relationship, the method substantially outperforms state-of-the-art (SoTA) approaches across four datasets (achieving a 28.1% F1 gain for hyperbole detection on TroFi and a 23.1% F1 gain for metaphor detection on HYPO-L).

Background & Motivation¶

Hyperbole and metaphor are two of the most common figures of speech in natural language, vital for downstream tasks such as sentiment analysis and dialogue systems. However, existing methods suffer from two core limitations:

Neglect of Emotional Factors: Most methods only focus on surface-level lexical and syntactic features, whereas the use of figures of speech is fundamentally emotion-driven. Without understanding the cruel emotion underlying "the butcher's knife," it is challenging to determine that "time is a butcher's knife" is a metaphor.
Neglect of the Interaction Between Hyperbole and Metaphor: Existing methods either detect the two independently or perform simple implicit feature sharing (e.g., through multi-task learning), failing to explicitly model the bidirectional reinforcing relationship between them. For instance, in "his determination is like a steel fortress," the metaphor (determination \(\rightarrow\) fortress) grounds the hyperbole (extreme sturdiness), while the hyperbole makes the metaphorical mapping more vivid.

Core Problem¶

How to leverage emotional knowledge to guide hyperbole and metaphor detection, and model the bidirectional dynamic interaction between them to improve detection accuracy?

This problem is crucial because: (1) emotion is a key bridge to understanding rhetorical effects, yet prior work has rarely integrated emotional knowledge systematically into rhetoric detection; (2) both hyperbole and metaphor involve deviating from literal meaning to achieve expressive effects, establishing a natural semantic reinforcement between them, but existing multi-task methods only share shallow features without explicitly modeling this interaction.

Method¶

EmoBi is an LLM-based multi-stage prompting framework. It does not involve model fine-tuning; instead, it guides the LLM through a step-by-step reasoning process via carefully designed prompt chains.

Overall Architecture¶

Given an input sentence, the goal is to predict both its hyperbole and metaphor labels. The overall process is divided into three stages:

Emotion Analysis \(\rightarrow\) Obtains the emotional information of the sentence.
Emotion-Guided Domain Mapping \(\rightarrow\) Identifies the source and target domains based on the emotional analysis.
Bidirectional Dynamic Interaction \(\rightarrow\) Employs hyperbole information to guide metaphor detection, and metaphor information to guide hyperbole detection, followed by a verification mechanism.

Each stage queries the LLM with a specific prompt, where the input of a subsequent stage incorporates the output of the previous stage, forming a progressive reasoning chain.

Key Designs¶

Emotion Analysis Module: Prompt 1 is used to prompt the LLM to analyze the emotional connotations of the sentence. This step connects surface-level language with deep rhetorical effects, providing emotional cues for subsequent detection. For example, identifying that "the butcher's knife" implies a cruel, ruthless emotion helps in determining its rhetorical usage.
Emotion-Based Domain Mapping Module: The sentence and its emotion analysis are fed together into the LLM. Through Prompt 2, the LLM is guided to identify the source domain and target domain from an emotional perspective and analyze their emotional connection. The source domain serves as the conceptual basis of the rhetorical expression, and the target domain serves as the destination of the semantic transfer. Pinpointing these two endpoints allows for a precise examination of semantic exaggeration in hyperbole and cross-domain mapping in metaphor.
Bidirectional Dynamic Interaction Module: This is the core innovation. Taking "metaphor-guided hyperbole detection" as an example: the model first analyzes metaphor based on emotional knowledge and domain mapping, yielding metaphorical information. This metaphorical information, alongside emotion and domain mapping knowledge, is then used as context to guide the LLM in determining whether hyperbole is present. Conversely, "hyperbole-guided metaphor detection" follows a similar workflow. The core insight is that the intense emotions and scale variation in hyperbole provide richer semantic extension directions for metaphors, while metaphors establish the semantic framework and emotional tone for hyperbole.
Verification Mechanism: After the initial detection, if inconsistencies or errors are detected in the identification results, the model re-evaluates and adjusts them to ensure the accuracy and reliability of the final detection.

Loss & Training¶

This method does not require model training or fine-tuning, operating entirely on the in-context reasoning of LLMs. Llama3-8b is used as the default backbone LLM, and its effectiveness is also verified on GPT-4o.

Key Experimental Results¶

Dataset	Task	Metric (F1)	EmoBi	MTL-F-RoBERTa (Prev. SOTA)	CoT-based	Gain (vs Prev. SOTA)
HYPO	Hyperbole	F1	90.8	88.1	83.2	+2.7
HYPO	Metaphor	F1	84.5	78.7	77.2	+5.8
HYPO-L	Hyperbole	F1	79.3	68.7	72.8	+10.6
HYPO-L	Metaphor	F1	80.3	57.2	72.6	+23.1
LCC	Hyperbole	F1	84.9	65.9	77.5	+19.0
LCC	Metaphor	F1	91.3	80.5	83.6	+10.8
TroFi	Hyperbole	F1	84.2	56.1	78.5	+28.1
TroFi	Metaphor	F1	76.6	57.3	70.7	+19.3

Compared with the fine-tuning-based state-of-the-art (MTL-F-RoBERTa), the performance improvements on the HYPO dataset are moderate (+2.7 / +5.8), but the gains are substantial on datasets like TroFi and LCC (+28.1 / +19.0). This demonstrates that LLM reasoning far surpasses small model fine-tuning in terms of cross-domain transfer capability.

Ablation Study¶

Variant	HYPO (Hyp/Met)	HYPO-L (Hyp/Met)	LCC (Hyp/Met)	TroFi (Hyp/Met)
Full Model	90.8 / 84.5	79.3 / 80.3	84.9 / 91.3	84.2 / 76.6
w/o Emotion Analysis	-4.6 / -5.1	-4.6 / -5.7	-5.3 / -5.4	-4.4 / -4.4
w/o Bidirectional Interaction	-3.4 / -3.8	-3.5 / -5.0	-4.0 / -4.1	-3.6 / -3.0
w/o Domain Mapping	-2.6 / -3.3	-2.7 / -4.3	-3.5 / -3.2	-2.8 / -2.7
w/o Verification Mechanism	-1.5 / -1.4	-1.2 / -1.9	-1.5 / -1.4	-1.3 / -1.4

The Emotion Analysis module contributes the most: Removing it leads to an average F1 drop of 4.4-5.7, with the most pronounced impact on metaphor detection, verifying that emotion is key to understanding rhetoric.
The Bidirectional Interaction module is the second largest contributor: Removing it leads to a drop of 3.0-5.0, proving that the mutually reinforcing relationship between hyperbole and metaphor provides substantial help for detection.
Domain Mapping is the third: Removing it yields a decrease of 2.6-4.3, indicating that source/target domain identification aids semantic understanding.
The Verification Mechanism has the smallest but stable impact: It accounts for a drop of 1.2-1.9 across all tasks, indicating a consistent positive contribution from result validation.

Highlights & Insights¶

Emotion as a bridge for rhetorical understanding: This insight is highly convincing—rhetorical devices are inherently designed to convey specific emotions, making it natural to understand rhetoric from an emotional standpoint. This "emotion-first" design paradigm can be transferred to tasks like sarcasm detection and irony understanding.
Bidirectional interaction paradigm: Instead of simple multi-task representation sharing, this approach explicitly feeds the detection results of one task as contextual input to the other, enabling directional information flow. This "task A informs task B, and vice-versa" design can be generalized to other multi-task scenarios with semantic associations.
Pure prompting outperforming fine-tuning: On TroFi, the F1 score increases from 56.1 to 84.2 (+28.1). Relying purely on prompt engineering vastly outperforms fine-tuned BERT/RoBERTa models, showing that for tasks requiring deep semantic reasoning like rhetorical understanding, the zero-shot capability of LLMs combined with structured reasoning chains can significantly exceed traditional fine-tuning.
Progressive prompt chain design: The emotion \(\rightarrow\) domain mapping \(\rightarrow\) bidirectional detection paradigm, where the output of each step serves as the input for the next, forms an information-rich reasoning chain. This is more targeted than standard CoT and can be applied to other multi-step reasoning tasks.

Limitations & Future Work¶

Error propagation: The authors acknowledge that multi-step reasoning suffers from error cascading—errors in emotion analysis will inevitably propagate to subsequent domain mapping and detection.
Unstable emotional analysis quality: The current emotion module might fail to capture subtle emotions accurately, especially in complex or mixed emotional contexts.
High computational overhead: Processing each sentence requires multiple LLM calls (emotion analysis, domain mapping, bidirectional detection, and verification), resulting in a much higher inference cost than fine-tuned small models.
Lack of code and reproduction details: The paper does not provide public code, and details regarding the exact prompt phrasing and the implementation of the verification mechanism are insufficient.
Small dataset scales: All four datasets are academic benchmarks, lacking validation in large-scale real-world scenarios.
English-bound: The method has not been validated in multilingual settings, while rhetoric is highly language- and culture-dependent.

vs Badathala et al. (2023) MTL-F: The previous SoTA, a BERT/RoBERTa-based multi-task fine-tuning method that only establishes shallow feature sharing. EmoBi comprehensively outperforms it through emotion guidance and explicit-interaction dimensions, showing massive gaps on TroFi (+28.1 F1) and HYPO-L (+23.1 F1), which reveals the performance ceiling of fine-tuning small models in rhetoric understanding.
vs CoT prompting: Standard CoT simply prompts the LLM to think "step by step" without domain-specific structured guidance for rhetoric tasks. EmoBi's three-stage design provides a task-specific reasoning framework for the LLM, significantly outperforming general CoT.
vs Tian et al. (2024) Domain Mining Method: Focuses on interpretable domain pair mining in metaphor detection but lacks emotional information and hyperbole-metaphor interaction modeling. EmoBi encapsulates both dimensions.

On the joint understanding of multiple rhetorical devices in NLP: The bidirectional interaction scheme of hyperbole and metaphor can be extended to the joint detection of more rhetorical devices, such as sarcasm, irony, and euphemism.
Emotion-driven semantic understanding paradigm: Utilizing emotion analysis as a precursor step for deep semantic understanding rather than an independent task is a design methodology that could be effective in tasks like stance detection and hate speech detection.
Methodology for LLM reasoning chain design: EmoBi demonstrates a task-customized prompt chaining methodology—first analyzing supporting knowledge (emotions, domain mapping), and then leveraging this knowledge for target detection, rather than prompting the LLM to output answers directly.

Rating¶

Novelty: ⭐⭐⭐⭐ The framework design of emotion guidance and bidirectional dynamic interaction is innovative, but the overall framework is a combination of prompt engineering.
Experimental Thoroughness: ⭐⭐⭐⭐ The evaluations across four datasets, ablation studies, model size/type comparisons, and case studies are comprehensive.
Writing Quality: ⭐⭐⭐⭐ The motivation is clearly articulated and the methodology is well-structured, but prompt details and description of the verification mechanism are not sufficiently detailed.
Value: ⭐⭐⭐ The design ideas of emotion guidance and bidirectional interaction are highly transferrable, but the specific task domain is relatively narrow.