Verb Metaphor Detection via Contextual Relation Learning

Wei Song, Shuhui Zhou, Ruiji Fu, Ting Liu, Lizhen LiuACL 2021 · OpenReview

TL;DR

This paper proposes a novel approach to verb metaphor detection by reframing it as a "relation extraction" task, rather than traditional classification. By explicitly modeling the relationship between a target verb and its various contextual components (grammatical, sentential, and basic meaning), their MrBERT model achieves competitive results on benchmark datasets, highlighting the importance of understanding these nuanced interactions.

What Problem Does This Paper Address?

Understanding human language requires distinguishing between literal and metaphorical meanings of words. Verb metaphors, in particular, are prevalent and crucial for conveying complex ideas. While previous work has treated verb metaphor detection as a classification or sequence labeling problem, often relying on token-level representations or broad sentence context, these methods have limitations. They tend to use context primarily to enhance the target word's representation, rather than explicitly modeling the interaction between the word and its context. Furthermore, some approaches struggle to effectively integrate diverse contextual information, potentially leading to noise or loss of focus. This paper aims to bridge this gap by proposing a more explicit way to model these contextual relationships.

Methods

The core innovation of this paper is the Metaphor-relation BERT (MrBERT) model, which views verb metaphor detection as a relation extraction problem. Instead of just classifying a verb's sense, MrBERT models the relationship between a target verb and its various "context components."

Here's a breakdown of their approach:

  • Context Components: They identify three key types of context:
  • Global Context: The entire sentence.
  • Local Context: Words with close grammatical relations to the target verb, specifically verb-subject and verb-object pairs, extracted using a Stanford dependency parser.
  • Distant Context: The basic, context-independent meaning of the verb itself, derived from BERT's tokenizer.
  • Representation: These context components are represented using BERT's hidden states. Special marker tokens are used to delineate the verb, subject, and object within the input sequence fed to BERT.
  • Relation Modeling: MrBERT explicitly models the "contextual relation" between the target verb and each context component. They explore three methods for this:
  • Linear Model: A simple parameter vector.
  • Bilinear Model: A parameter matrix for more interaction.
  • Neural Tensor Model: A more complex model for richer interactions.
  • Integration Strategies: To combine the information from different context components, they use:
  • Context Concatenation: Stacking the representations.
  • Context Average: Averaging the representations.
  • Context Maxout: Taking the maximum prediction across individual contexts.
  • Prediction: The metaphoricity of the verb is predicted based on these modeled contextual relations. The model is trained using a combined loss function that includes both relation-level prediction and sequence-level prediction (for broader token-level metaphor detection).

Datasets

The paper evaluates MrBERT on three established datasets:

  • VUA (VU Amsterdam Metaphor Corpus): The largest public dataset, used extensively for benchmarking. The authors use a preprocessed version for consistent comparison.
  • MOH-X: A smaller dataset.
  • TroFi: Another smaller dataset.

Key Insights and Ideas

  • Reframing as Relation Extraction: The most significant insight is that viewing metaphor detection as a relation extraction task, rather than simple classification, allows for a more explicit modeling of how a verb's meaning interacts with its surroundings. This aligns better with linguistic theories of metaphor.
  • Leveraging Grammatical Relations: Explicitly incorporating grammatical relations (verb-subject, verb-object) as local context components significantly improves performance, especially for verbs with incomplete argument structures. This suggests that syntactic information is vital for disambiguating metaphorical usage.
  • The Power of Multiple Contexts: Integrating global, local, and distant contexts together yields better results than using any single context type. This highlights the multifaceted nature of metaphor and the need for a holistic approach.
  • Basic Meaning as a Distant Context: Using the verb's context-independent meaning as a "distant context" is shown to be effective, supporting the idea that metaphors often involve a divergence from a word's core sense.

Key Results

  • MrBERT achieves competitive and often superior performance compared to state-of-the-art methods on the VUA dataset, particularly on the Verb track.
  • For example, on the VUA Verb track, MrBERT achieves an F1 score of 75.9, outperforming several strong baselines.
  • The paper demonstrates that the bilinear model for relation modeling and the context average strategy for integration perform best.
  • On smaller datasets like MOH-X and TroFi, MrBERT shows good zero-shot transferability and competitive results, especially when fine-tuned.
  • Analysis shows that incorporating grammatical relations leads to larger improvements on sentences with more clauses, suggesting its utility in complex linguistic structures.

Limitations and Future Work

  • All-POS Track Performance: While MrBERT excels at verb metaphor detection, its performance on the All-POS (all part-of-speech tags) track is slightly lower than some competitors like MelBERT. The authors attribute this to DeepMet and MelBERT being trained for every token, potentially allowing for more sufficient training, whereas MrBERT primarily focuses on verbs.
  • Future Work: The authors suggest extending MrBERT to handle tokens of other POS tags, indicating its flexibility. Further exploration into more sophisticated relation modeling and context integration strategies is also implied.

Summary updated

BibTeX
@inproceedings{song2021verb,
  author = {Wei Song and Shuhui Zhou and Ruiji Fu and Ting Liu and Lizhen Liu},
  title = {{Verb Metaphor Detection via Contextual Relation Learning}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
  url = {https://openreview.net/forum?id=TBPnww9cE9},
}
Figure