Metaphors are a fundamental component of human language, enabling abstract reasoning, nuanced communication, and cultural expression. However, their inherent complexity poses significant challenges for natural language processing (NLP) systems, which must accurately detect, interpret, and translate metaphorical expressions across diverse linguistic and cultural contexts. This thesis addresses these challenges by developing computational methodologies for interpretable metaphor processing, bridging insights from computational linguistics, psycholinguistics, and artificial intelligence.
The work advances three core areas: 1）Enhancing metaphor detection through syntactic pruning (RoPPT), semantic frame integration (FrameBERT), and basic meaning modeling (BasicBERT), achieving state-of-the-art performance across benchmark datasets. 2）Improving cross-linguistic metaphor translation via MMTE – a novel evaluation framework combining human and automatic metrics to assess emotional salience and translation quality in English, Chinese, and Italian. 3）Leveraging interpretability in large language models (LLMs) for metaphor understanding through sparse autoencoders and dictionary learning, decomposing latent representations to extract monosemantic features that improve metaphor transparency.
The findings contribute to the development of more linguistically sophisticated, contextually adaptive, and culturally aware NLP systems. By advancing metaphor processing methodologies and introducing interpretability techniques, this research provides a foundational exploration for applications in machine translation, sentiment analysis, and explainable AI. The proposed models and evaluation frameworks not only improve metaphor understanding in computational settings but also provide a foundation for future work in cross-linguistic and cross-cultural NLP.
