Bridging Word-Pair and Token-Level Metaphor Detection with Explainable Domain Mining

Yuan Tian, Ruike Zhang, Nan Xu, Wenji MaoACL 2024 · OpenReview

Abstract

Metaphor detection aims to identify whether a linguistic expression in text is metaphorical or literal. Most existing research tackles this problem either using word-pair or token-level information as input, and thus treats word-pair and token-level metaphor detection as distinct subtasks. Benefited from the simplified structure of word pairs, recent methods for word-pair metaphor detection can provide intermediate explainable clues for the detection results, which remains a challenging issue for token-level metaphor detection. To mitigate this issue in token-level metaphor detection and take advantage of word pairs, in this paper, we make the first attempt to bridge word-pair and token-level metaphor detection via modeling word pairs within a sentence as explainable intermediate information. As the central role of verb in metaphorical expressions, we focus on token-level verb metaphor detection and propose a novel explainable Word Pair based Domain Mining (WPDM) method. Our work is inspired by conceptual metaphor theory (CMT). We first devise an approach for conceptual domain mining utilizing semantic role mapping and resources at cognitive, commonsense and lexical levels. We then leverage the inconsistency between source and target domains for core word pair modeling to facilitate the explainability. Experiments on four datasets verify the effectiveness of our method and demonstrate its capability to provide the core word pair and corresponding conceptual domains as explainable clues for metaphor detection.

Introduction: The Explainability Gap

Metaphor detection in Natural Language Processing (NLP) generally falls into two categories: word-pair detection and token-level detection. Word-pair models classify specific pairs (e.g., sweet child) and are relatively easy to explain using "source" and "target" domains. In contrast, token-level models identify metaphorical words (usually verbs) within full sentences. While more practical, token-level models often act as "black boxes," lacking the clear, theory-driven explanations found in word-pair models.

Conceptual Domain Inconsistency in Metaphor Detection

This research introduces Word Pair based Domain Mining (WPDM) to bridge this gap. By using word pairs as explainable intermediate information, the model identifies metaphors based on domain inconsistency—for example, recognizing that "digesting news" involves a clash between the INFORMATION domain (target) and the FOOD domain (source).

The WPDM Framework

The WPDM method operates through a four-stage pipeline designed to mimic human-like conceptual reasoning:

  1. Literal Example Construction: The system retrieves example sentences from dictionaries where the focus verb is used in its primary, literal sense (e.g., "digesting meat").
  2. Semantic Role Mapping: Using tools like VerbNet, the model identifies the "arguments" (Agent and Theme) for the verb in both the input sentence and the literal examples.
  3. Conceptual Domain Mining: Leveraging WordNet and cognitive resources, the model maps these arguments to conceptual domains. It identifies the appropriate level of abstraction (e.g., mapping "news" to INFORMATION).
  4. Word Pair Modeling & Detection: Inspired by Conceptual Metaphor Theory (CMT), the model calculates the "inconsistency" between the target and source domains. This inconsistency score weights the importance of different word pairs during final classification by a RoBERTa-based neural network.
Overview of the WPDM Framework

Experimental Evaluation

The model was evaluated on four major datasets: MOH-X (WordNet-based), TroFi (Wall Street Journal), VUA-Verb (VU Amsterdam Corpus), and a new benchmark, LCC-Verb, specifically filtered for verb instances.

WPDM achieved state-of-the-art performance, outperforming specialized linguistics-enhanced networks and large language models. Notably, WPDM achieved F1 scores of 83.8 on MOH-X, 84.9 on LCC-Verb, and 73.4 on TroFi. In direct comparisons with GPT-3.5 and GPT-4, WPDM significantly outperformed both in zero-shot and few-shot settings. For example, on LCC-Verb, WPDM's F1 of 84.9 far exceeded GPT-4's best effort of 69.7.

Key Insights and Discussion

The success of WPDM highlights several critical aspects of metaphor detection:

  • Domain Conflict: Metaphor is fundamentally about domain inconsistency. Quantifying how much a target word deviates from a verb's typical literal arguments provides a powerful detection signal.
  • Explainability: By identifying a "core word pair" and its associated domains, the system provides human-readable justifications (e.g., "I flagged 'digests' because 'news' is INFORMATION, not FOOD"). Human validation confirmed these explanations are accurate over 80% of the time.
  • Contextual Focus: While a sentence may contain multiple word pairs, the model successfully identifies the "core" pair that carries the primary metaphorical meaning.

Limitations and Future Work

Currently, WPDM is verb-centric and does not handle noun-noun or adjective metaphors. It also relies on external knowledge bases like WordNet and VerbNet, which may limit its performance for words not present in these databases. Future work aims to expand the framework to other parts of speech and explore ways to reduce reliance on static external resources.

Summary updated

Abstract

Metaphor detection aims to identify whether a linguistic expression in text is metaphorical or literal. Most existing research tackles this problem either using word-pair or token-level information as input, and thus treats word-pair and token-level metaphor detection as distinct subtasks. Benefited from the simplified structure of word pairs, recent methods for word-pair metaphor detection can provide intermediate explainable clues for the detection results, which remains a challenging issue for token-level metaphor detection. To mitigate this issue in token-level metaphor detection and take advantage of word pairs, in this paper, we make the first attempt to bridge word-pair and token-level metaphor detection via modeling word pairs within a sentence as explainable intermediate information. As the central role of verb in metaphorical expressions, we focus on token-level verb metaphor detection and propose a novel explainable Word Pair based Domain Mining (WPDM) method. Our work is inspired by conceptual metaphor theory (CMT). We first devise an approach for conceptual domain mining utilizing semantic role mapping and resources at cognitive, commonsense and lexical levels. We then leverage the inconsistency between source and target domains for core word pair modeling to facilitate the explainability. Experiments on four datasets verify the effectiveness of our method and demonstrate its capability to provide the core word pair and corresponding conceptual domains as explainable clues for metaphor detection.
BibTeX
@inproceedings{tian2024bridging,
  author = {Yuan Tian and Ruike Zhang and Nan Xu and Wenji Mao},
  title = {{Bridging Word-Pair and Token-Level Metaphor Detection with Explainable Domain Mining}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
  url = {https://openreview.net/forum?id=q6Mv82c0B4},
}
Figure