Interpretable Chinese Metaphor Identification via LLM-Assisted MIPVU Rule Script Generation: A Comparative Protocol Study

Metaphor pervades human language, structuring how we reason about abstract concepts through concrete experience. Automatic metaphor identification, the task of determining whether a given linguistic expression is used metaphorically, has received sustained attention in computational linguistics, driven by applications in sentiment analysis, machine translation, and discourse understanding.

Despite considerable progress, the field faces a persistent challenge. State-of-the-art neural classifiers based on pre-trained language models achieve strong performance on benchmark datasets but provide no structured explanation for their decisions. A model may correctly flag a token as metaphorical yet offer no insight into the conceptual mapping, basic-meaning contrast, or figurative mechanism that justifies the label. This opacity limits both scientific understanding of what these models learn and practical deployment in educational or annotation-support settings where users need to know.

In Chinese, this interpretability problem is compounded by several challenges. First, Chinese lacks the morphological inflections and derivational patterns that provide surface cues for metaphor in Indo-European languages; the distinction between literal and figurative senses must be resolved almost entirely through context and world knowledge. Second, Chinese figurative language encompasses diverse phenomena, including conceptual metaphor, simile, metonymy, and culture-specific figures of speech, that do not map neatly onto annotation frameworks developed for English. Third, annotated resources for Chinese metaphor remain relatively scarce and fragmented across incompatible annotation schemes.

The core problem in metaphor identification lies in balancing between opaque classifiers and interpretability. Existing neural network models, while performing excellently, lack transparency in their decision-making processes, which is particularly pronounced in Chinese metaphor identification.

The complexity of Chinese lies in its lack of morphological markers, making metaphor identification almost entirely reliant on context and world knowledge. Additionally, Chinese figurative language phenomena are diverse, encompassing conceptual metaphor, simile, metonymy, etc., which do not easily correspond to annotation frameworks developed for English. The scarcity and incompatibility of annotated resources further exacerbate the difficulty of identification.

Therefore, achieving high performance while providing interpretability in metaphor identification remains a pressing challenge.

The core innovation of this paper is the introduction of an LLM-assisted pipeline for metaphor identification, operationalizing four protocols as executable rule scripts. This approach not only improves identification accuracy but also enhances the interpretability and reproducibility of results.

• �� LLM-assisted rule script generation: Using LLMs to generate executable metaphor identification rule scripts achieves full transparency of the identification process.

• �� Cross-protocol comparison: Conducting the first cross-protocol comparison in Chinese metaphor identification reveals that protocol choice is the largest source of variation.

• �� Modular design: Each protocol is a modular chain of deterministic steps interleaved with controlled LLM calls, producing structured rationales alongside every classification decision.

The methodology of this paper consists of the following steps:

• �� Preprocessing: Text is segmented, POS-tagged, and normalized to ensure consistency.

• �� Candidate Selection: Depending on the protocol, analysis targets are selected. For example, Protocol A selects all content words, while Protocol B selects sentences containing potential cross-domain expressions.

• �� Semantic Analysis: LLMs are used for contextual meaning and basic meaning retrieval and contrast, or for evaluating emotional valence.

• �� Classification Decision: Binary or multi-class decisions are made based on protocol criteria. For example, Protocol A labels a token as metaphorical if the contextual meaning contrasts with the basic meaning.

• �� Rationale Generation: A structured explanation is generated for each decision, including the specific protocol step that triggered the decision, key evidence, and a confidence indicator.

The experimental design includes evaluation on seven Chinese metaphor datasets, covering token-, sentence-, and span-level annotation. The benchmark datasets used include PSU CMC, CMC, CMDAG, Chinese Simile, NLPCC 2024 T9, ConFiguRe, and ChineseMCorpus.

Each protocol is evaluated on its most closely aligned dataset using standard train/dev/test splits. For cross-protocol evaluation, all four protocols are applied to a common subset of PSU CMC, converted to sentence-level labels.

Experiments use GPT-4 as the underlying LLM, with temperature set to 0 to maximize determinism. Evaluation uses standard metrics: precision, recall, and F1 score for binary classification; partial-match F1 for span extraction.

Experimental results show that Protocol A (MIP) achieves an F1 of 0.472 on token-level identification, with an overall accuracy of 0.898. Register-level analysis reveals that academic prose performs best with an F1 of 0.598.

Protocol B (CMDAG) achieves an F1 of 0.347 on sentence-level identification, showing high precision but low recall, indicating that while it correctly identifies metaphors, it misses many metaphorical sentences.

Protocol C (Emotion) mirrors Protocol B's performance, suggesting that emotion-based and conceptual-mapping approaches capture a similar subset of metaphors in this dataset. Protocol D (Simile) achieves an F1 of 0.392 in binary classification, showing moderate precision.

This method can be directly applied to Chinese text metaphor identification, especially in scenarios requiring high transparency and interpretability, such as educational and annotation support environments.

Additionally, the method can be used to develop more complex figurative language processing systems, integrating multiple protocols to improve identification accuracy and consistency.

In industrial applications, this method can be used in sentiment analysis, machine translation, and discourse understanding, providing more accurate and interpretable results.

Despite the excellent performance of this method in metaphor identification, there are some limitations. Firstly, due to the complexity of Chinese, metaphor identification relies almost entirely on context and world knowledge, increasing complexity.

Secondly, differences between protocols may lead to inconsistent identification results, especially when dealing with complex metaphor structures.

Finally, although rule scripts provide transparency, the LLMs they rely on may introduce biases, particularly when handling unseen corpora. Future research could explore better integration of multiple protocols to improve the accuracy and consistency of metaphor identification.