Adversarial Multi-task Learning for End-to-end Metaphor Detection

Shenglong Zhang, Ying LiuACL 2023 · DOI

Introduction

Metaphor detection (MD) is a critical task in natural language processing, essential for understanding nuanced human communication, sentiment analysis, and machine translation. A metaphor occurs when a word or phrase is used in a non-literal way to map meaning from a "source" domain to a "target" domain (e.g., "the police smashed the drug ring").

Despite its importance, MD is frequently hamstrung by a severe data scarcity problem. Annotating metaphors requires expert linguistic knowledge, resulting in datasets that are small and prone to overfitting. Conversely, Word Sense Disambiguation (WSD)—the task of identifying which specific meaning of a word is being used—enjoys an abundance of labeled data. This paper introduces AdMul, a framework designed to bridge this gap by distilling knowledge from data-rich word senses into the data-poor world of metaphors.

The AdMul Framework

The authors propose the Adversarial Multi-task Learning (AdMul) framework, which reframes WSD as an auxiliary task called Basic Sense Discrimination (BSD). Drawing on the "Metaphor Identification Procedure" (MIP), the authors assume that the most frequent sense of a word in WordNet represents its "basic" or literal meaning. BSD then becomes a binary classification task: determining if a word is used in its basic or non-basic sense.

The AdMul architecture featuring a shared feature extractor and adversarial discriminators.

As shown in the architecture above, AdMul consists of several key components:

  1. Shared Feature Extractor: The model utilizes a Transformer-based backbone (such as DeBERTa-base) to encode the target word and its surrounding context into a dense representation.
  2. Global Discriminator: To ensure the model learns features applicable to both MD and BSD, a Global Discriminator paired with a Gradient Reversal Layer (GRL) is employed. This forces the encoder to produce "task-invariant" representations, making it impossible for the classifier to distinguish whether the input originated from the MD or BSD dataset.
  3. Local Discriminators (Multi-mode Alignment): This component acts as the framework's "secret sauce." The authors argue that "literal" usage in MD corresponds to "basic" usage in BSD, while "metaphorical" usage corresponds to "non-basic." Two local discriminators are used to forcibly align these specific clusters in the feature space.

Adversarial Multi-mode Alignment

The intuition behind multi-mode alignment is that metaphors are essentially semantic conflicts between a word's contextual meaning and its most basic, concrete meaning. By training the model to recognize "basicness," it effectively learns the foundation of literalness.

Visualization of the feature space alignment between literal/basic and metaphorical/non-basic clusters.

As illustrated in the conceptual scatter plot, the adversarial process encourages the model to find a shared feature space where literal and basic senses are indistinguishable, as are metaphorical and non-basic senses. This alignment allows the model to leverage the vast linguistic patterns found in WSD data to improve its performance on the smaller MD datasets.

Experimental Setup and Datasets

AdMul was evaluated on four standard MD benchmarks:

  • VUA All & VUA Verb: The VU Amsterdam Metaphor Corpus, currently the largest available resource for this task.
  • MOH-X: A dataset sampled directly from WordNet.
  • TroFi: A dataset consisting of sentences from the Wall Street Journal.

For the auxiliary BSD task, the authors created SemCor-BSD, derived from SemCor (the largest manually annotated WSD dataset). They specifically filtered for words with at least three senses to ensure the discrimination task remained meaningful and challenging.

Key Results and Insights

AdMul achieved state-of-the-art (SOTA) performance across all evaluated datasets. Notably, it reached an F1 score of 79.0 on VUA All and 87.9 on MOH-X. The performance gains were most pronounced on smaller datasets like MOH-X and VUA Verb, confirming that knowledge transfer from the BSD task effectively mitigates data scarcity.

Furthermore, the model demonstrated remarkable zero-shot robustness. In tests where the model was trained on VUA and tested on other datasets without fine-tuning, AdMul outperformed several models that had been specifically fine-tuned on those target datasets. For instance, AdMul's zero-shot F1 on MOH-X was 83.8, surpassing the fine-tuned MisNet (83.4). This suggests that AdMul learns general linguistic principles rather than just memorizing dataset-specific quirks.

Conclusion and Future Work

AdMul successfully demonstrates that adversarial multi-task learning can unlock the potential of WSD data for metaphor detection. By aligning the concepts of "basicness" and "literalness," the model achieves superior performance and robustness.

Future work may address current limitations, such as the model's reliance on sentence-level context, which can struggle with "indirect metaphors" requiring paragraph-level understanding. Additionally, while the heuristic of using the most frequent WordNet sense as the "basic" sense is effective, exploring more nuanced ways to define literalness could further refine the framework.

Summary updated

BibTeX
@inproceedings{zhang2023adversarial,
  author = {Shenglong Zhang and Ying Liu},
  title = {{Adversarial Multi-task Learning for End-to-end Metaphor Detection}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
  doi = {10.18653/v1/2023.findings-acl.96},
}
Figure