Conceptual metaphor processing and its application

Abstract

Metaphors, as figurative expressions embedded within natural language, are often used unconsciously and implicitly by individuals. The rich sentiments and implicit connotations in metaphors influence metaphor processing in downstream natural language processing tasks, such as sentiment analysis, dialogue systems, and machine translation.

Most metaphor processing research focuses on linguistic metaphor processing, which involves identifying and understanding metaphors by comparing their basic meanings with their contextual meanings. With the increase in model complexity and popularity of deep learning models, linguistic metaphor processing models can achieve growingly accurate performance. However, they usually regard metaphors more as linguistic phenomena, ignoring their important role in human cognition. Most of the previous research in conceptual metaphor processing mainly depended on domain-specific knowledge to develop clusters or generate concept mappings that could only represent limited metaphors, leading to difficulties in generalization and restricted roles in downstream tasks.

This thesis focuses on conceptual metaphor processing and its application. Our main contributions are listed as follows:

•We first attempted to propose an explainable metaphor identification model, which can generate concept mappings at the same time. Without golden labels of concept mappings, we utilized a statistical algorithm and a dynamic reward mechanism to select concepts from a structural knowledge base. The generated concept mappings can enhance the performance of metaphor identification and reflect human underlying cognition by explaining why it is a metaphor and why people compare one concept to another. We also designed multiple settings to evaluate and demonstrate that the generated concept mappings can deliver the meanings of the input metaphors.

•To extend the usage of generated concepts, we evaluated our algorithm proposed before to conceptualize an entity into a concept. First, we asked annotators to annotate concepts for 1,000 entities (300 nouns, 400 verbs, 150 adjectives, and 150 adverbs). Then, we evaluated the accuracy of our algorithm compared with GPT 4.0. The results showed that our algorithm greatly exceeded GPT 4.0 in the conceptualization task. Finally, we analyzed the tendencies of concepts generated by human annotators, our algorithm, and GPT 4.0 from frequency, abstractness, multi-word, and diversity.

•To explore the application of concept mappings in linguistics and cognitive science, we analyzed the cognitive patterns of toxic language based on the concept mappings generated from metaphors. First, we processed a large-scale public toxic language dataset with our conceptual metaphor processing tool. Then, we compared the differences and analyzed the associations and dependencies of source concepts, target concepts, and concept mappings between toxic and non-toxic language, multiple levels and subtypes of toxic language, as well as toxic language mentioning different genders, sexual orientations, and races. Our findings have verified some previous linguistics and cognition studies. Besides, some of our exciting insights can provide hypotheses for linguistics and cognition fields.