How Can Metaphor not Handle Anomaly? Metaphor Detection with Anomalous Text

ACL ARR 2024 April Submission133 Authors

Abstract: Metaphor is essentially literal shifts in meaning, which is manifested as a mismatch between the literal meaning of the target word and its contextual context. In metaphor research, the theory of Selection Preference Violation (SPV) is commonly used to identify metaphor, which the target word occurs less frequently in the surrounding words in its context, yielding a mismatch. Researchers are mainly concerned with considering such collocational mismatch as a metaphorical expression, yet they tend to overlook that collocational mismatch may also be a syntactic anomaly. We integrate syntactic anomaly into the metaphor detection. First, we use ChatGPT to construct a dataset containing syntactic anomaly, called the LMA. Second, we propose a model for enhanced literal, metaphor, and syntactic anomaly detection (MetaLA), which considers not only the target word and context in classification detection, but also adds other semantic contexts to reduce misclassifying anomaly as metaphor. We explore the relationship between literal, metaphor and syntactic anomaly, as well as the role of introducing SPV. Our experimental results show that syntactic anomaly reduce the model's correctness for metaphor detection, and that SPV reduces this correctness even further. Finally, we compare MetaLA with existing metaphor detection methods as well as other large language models (LLM) to demonstrate the effectiveness of our approach in literal, metaphor and syntactic anomaly detection.

Paper Type: Long

Research Area: Efficient/Low-Resource Methods for NLP

Research Area Keywords: Efficient/Low-Resource Methods for NLP

Contribution Types: Approaches to low-resource settings, Data resources

Languages Studied: English

Submission Number: 133