Bidirectional Interaction Module in EmoBi Framework

• The Bidirectional Interaction Module is a neuro-symbolic prompting mechanism that enables mutual reinforcement between hyperbole and metaphor detection using staged LLM calls.
• It integrates sequential pipelines of emotion analysis, domain mapping, and bidirectional reasoning to boost detection precision, as demonstrated by significant F1 improvements.
• Empirical results show that removing the module decreases performance by up to 5 F1 points, underscoring its critical role in enhancing multi-task rhetorical device detection.

The Bidirectional Interaction Module is the component of the EmoBi framework that operationalizes mutual promotion between hyperbole detection and metaphor detection through staged LLM reasoning rather than through a parameterized neural interaction block (Zheng et al., 18 Jun 2025). In EmoBi, the module is embedded in an emotion-guided pipeline in which emotion analysis and emotion-based domain mapping first produce intermediate semantic signals, after which bidirectional reasoning transfers task-specific knowledge in both directions: metaphor knowledge informs hyperbole prediction, and hyperbole knowledge informs metaphor prediction. The module is therefore defined not by attention matrices, gates, or trainable interaction losses, but by prompt-conditioned LLM calls that compose emotion, domain, and cross-task information into final labels for the two rhetorical-device detection tasks (Zheng et al., 18 Jun 2025).

1. Definition and scope

Within EmoBi, the goal of the Bidirectional Interaction Module is to enable mutual promotion between hyperbole detection and metaphor detection by exchanging task-specific knowledge in both directions, guided by emotion analysis and emotion-based domain mapping (Zheng et al., 18 Jun 2025). The task is formalized for a sentence $x$ as simultaneous prediction of a hyperbole label $y_h$ and a metaphor label $y_m$, namely $f(x) \to (y_h, y_m)$.

This formulation distinguishes the module from both single-task and unidirectional approaches. Single-task methods detect hyperbole or metaphor independently, thereby missing mutual reinforcement. Unidirectional interaction would pass information only one way, whereas the EmoBi design explicitly includes both metaphor-guided hyperbole reasoning and hyperbole-guided metaphor reasoning. A plausible implication is that the module treats the two rhetorical devices as semantically entangled rather than merely co-occurring.

Crucially, the paper states that EmoBi’s “bidirectional dynamic interaction” is not instantiated as a neural layer. It does not define neural interaction weights, attention matrices, gates, or training losses. Instead, it formalizes interaction through carefully staged LLM prompts that inject emotion and domain knowledge and then transfer reasoning in both directions between hyperbole and metaphor (Zheng et al., 18 Jun 2025).

2. Position within the EmoBi framework

The module appears after two upstream components: the emotion analysis module and the emotion-based domain mapping module. The overall pipeline is described as emotion analysis $\to$ domain mapping $\to$ bidirectional dynamic interaction $\to$ verification. EmoBi is therefore a multi-step LLM prompting pipeline rather than a trained neural network (Zheng et al., 18 Jun 2025).

Its inputs are the raw sentence $x$, the emotion analysis output $x_e$, the emotion-based domain mapping output $x_d$, and intermediate knowledge from the opposite task: metaphor information $y_h$0 for metaphor-guided hyperbole, or hyperbole information $y_h$1 for hyperbole-guided metaphor. Its outputs are the final labels consumed by the classification stage, namely the hyperbole label $y_h$2 and the metaphor label $y_h$3.

The two upstream modules supply the semantic substrate on which the interaction operates. Emotion analysis is defined as

$y_h$4

Emotion-based domain mapping is then defined as

$y_h$5

Semantically, $y_h$6 includes the source domain and target domain along with the emotion connection. The paper emphasizes that emotion knowledge enriches the target domain’s semantic representation and sets the tone that improves metaphor mapping and hyperbolic degree reasoning (Zheng et al., 18 Jun 2025). This suggests that the module’s bidirectionality is semantically conditioned rather than purely structural.

3. Bidirectional reasoning paths and mathematical formulation

The core of the module consists of two complementary knowledge-transfer paths. The first is metaphor-guided hyperbole detection:

$y_h$7

The second is hyperbole-guided metaphor detection:

$y_h$8

In the first direction, metaphor knowledge is extracted as

$y_h$9

and hyperbole classification is then conditioned on emotion, domain, and metaphor knowledge:

$y_m$0

In the reverse direction, hyperbole knowledge is extracted analogically as

$y_m$1

and metaphor classification is conditioned on emotion, domain, and hyperbole knowledge:

$y_m$2

The effective update dependencies are therefore explicit. Metaphor-to-hyperbole transfer means that $y_m$3 depends on $y_m$4, while hyperbole-to-metaphor transfer means that $y_m$5 depends on $y_m$6 (Zheng et al., 18 Jun 2025). Unlike implicit feature fusion in conventional multi-task models, this design exposes the reasoning path by which one task conditions the other.

The paper’s description also clarifies what the module is not. There are no equations for cross-entropy over $y_m$7 or $y_m$8, no $y_m$9-weighted interaction consistency terms, and no differentiable verification penalties. Evaluation uses Precision, Recall, and F1 only. Accordingly, the module is best understood as a symbolic-prompted interaction mechanism implemented through sequential LLM calls rather than as a trainable differentiable block.

4. Verification, inference procedure, and operational character

After the bidirectional reasoning stage, EmoBi applies a verification mechanism. This mechanism is described qualitatively: if an error is detected in the identified hyperboles or metaphors, the model re-evaluate and adjust the results (Zheng et al., 18 Jun 2025). No equations, explicit thresholds, confidence calibration functions, or auxiliary classifiers are specified.

Conceptually, verification acts as an additional LLM check over the intermediate outputs $f(x) \to (y_h, y_m)$0 and the final labels $f(x) \to (y_h, y_m)$1. A plausible implication is that the mechanism functions as a consistency filter over the staged reasoning trace, rather than as a probabilistic confidence estimator.

The inference process is sequential. Emotion analysis is executed first, followed by domain mapping, followed by the two bidirectional reasoning directions, and then verification. Optimization algorithm, hyperparameters, number of interaction layers or iterations, and initialization are reported as not applicable or not provided. Computational complexity is also not provided. The paper further reports comparisons between Llama3-8b and GPT-4o, with GPT-4o consistently outperforming Llama3-8b (Zheng et al., 18 Jun 2025).

These properties define the module’s operational identity. It is an inference-time interaction procedure, not a separately optimized network component. Its reproducibility depends on prompt templates, LLM choice, and the staged execution order more than on architectural depth or objective design.

5. Empirical contribution within EmoBi

The paper provides direct evidence for the contribution of the Bidirectional Interaction Module through an ablation labeled “w/o interaction.” Removing the module consistently degrades performance across all reported datasets (Zheng et al., 18 Jun 2025).

Dataset	Hyperbole F1	Metaphor F1
HYPO	90.8 → 87.4	84.5 → 80.7
HYPO-L	79.3 → 75.8	80.3 → 75.3
LCC	84.9 → 80.9	91.3 → 87.2
TroFi	84.2 → 80.6	76.6 → 73.6

These drops range from $f(x) \to (y_h, y_m)$2 to $f(x) \to (y_h, y_m)$3 F1, indicating that bidirectional interaction contributes materially to both tasks rather than only to one direction. The largest reported degradation from removing interaction is on HYPO-L metaphor detection, where F1 falls from 80.3 to 75.3.

The module’s contribution is also situated within stronger end-to-end results. EmoBi outperforms all baselines, including the MTL-F-RoBERTa SoTA. The paper reports a $f(x) \to (y_h, y_m)$4 F1 increase for hyperbole detection on TroFi and a $f(x) \to (y_h, y_m)$5 F1 increase for metaphor detection on HYPO-L relative to the current SoTA (Zheng et al., 18 Jun 2025).

Additional ablations indicate that the interaction module operates most effectively in conjunction with the other semantic components. Removing emotion analysis causes the largest drops, such as HYPO-L metaphor F1 changing from 80.3 to 74.6. Removing domain mapping also reduces performance, for example HYPO-L metaphor F1 changing from 80.3 to 76.0. Removing verification leads to smaller but consistent drops, such as HYPO-L metaphor F1 changing from 80.3 to 78.4. This pattern suggests that the interaction module is not a self-sufficient mechanism; it is strengthened by emotion-grounded and domain-grounded inputs.

6. Relation to prior interaction mechanisms and limitations

The EmoBi module differs sharply from bidirectional interaction mechanisms in other contemporary architectures. In BIFIT for referring video object segmentation, the bidirectional vision-language interaction module is a lightweight, plug-and-play pre-aligner built from multi-head attention and element-wise multiplication, with explicit query-key-value structure and end-to-end training objectives (Lan et al., 2023). In CroBIM for referring remote sensing image segmentation, the mutual-interaction decoder uses cascaded bidirectional cross-attention together with residual connections and multi-scale deformable attention (Dong et al., 2024). In both cases, bidirectional interaction is realized as a parameterized neural computation.

By contrast, EmoBi implements bidirectionality as prompted reasoning over intermediate semantic artifacts. Its comparative advantages, as stated in the paper, are explicit, knowledge-centric interaction, emotion-guided domain mapping, and a verification step absent from typical multi-task architectures (Zheng et al., 18 Jun 2025). Prior multi-task baselines such as MTL-F with BERT, ALBERT, or RoBERTa rely on shared representations and joint training without explicit emotion/domain knowledge or explicit bidirectional reasoning steps.

The module’s limitations follow directly from its multi-step structure. Error propagation is a central concern: early errors in emotion analysis $f(x) \to (y_h, y_m)$6 or domain mapping $f(x) \to (y_h, y_m)$7 can cascade into $f(x) \to (y_h, y_m)$8, $f(x) \to (y_h, y_m)$9, and the final labels. The paper explicitly notes sensitivity to inaccurate emotion detection and misidentified source or target domains, and the ablations without emotion analysis or domain mapping empirically support this vulnerability (Zheng et al., 18 Jun 2025). Generalization beyond hyperbole and metaphor is also not evaluated; the paper notes that other rhetorical devices such as irony are not studied.

Taken together, these properties place the Bidirectional Interaction Module in a distinct category. It is neither a generic attention block nor a simple multi-task sharing strategy. It is a staged neuro-symbolic-style prompting procedure in which task interaction is mediated by emotion connotations, source-target domain mapping, and cross-task reasoning traces, with final reliability supported by a qualitative verification pass (Zheng et al., 18 Jun 2025).