Using Conceptual Norms for Metaphor Detection

Mingyu Wan, Kathleen Ahrens, Emmanuele Chersoni, Menghan Jiang, Qi Su, Rong Xiang, Chu-Ren HuangACL 2020

TL;DR

This paper introduces a linguistically-enriched approach to automatically detect metaphors at the word level. By leveraging conceptual norms related to sensory modalities and embodiment, the authors significantly improve metaphor detection performance, outperforming strong baselines and even some deep learning models without using complex architectures.

What Problem Does This Paper Address?

Metaphor detection is a crucial task for natural language processing (NLP) applications like machine translation and sentiment analysis, as metaphors are fundamental to how humans understand and express complex ideas. However, automatically identifying metaphors is challenging due to the subtle semantic and ontological differences between literal and figurative language, and the subjective nature of metaphor perception. Existing methods often struggle with generalization or lack a deep connection to the underlying cognitive properties of metaphors. This paper aims to address this by developing a more robust metaphor detection system that incorporates insights from cognitive linguistics.

Methods

The core of this paper's contribution lies in its novel feature engineering approach. Instead of relying solely on lexical or distributional features, the authors incorporate "conceptual norms" to represent words. Specifically, they use:

  • Modality Exclusivity (ME) and Dominant Modality (DM) norms: These capture how sensory modalities (touch, hearing, smell, taste, vision, interception) and action effectors are associated with words. The idea is that metaphors often involve a mismatch or transfer of properties across these sensory domains.
  • Embodiment (EB) and Embodiment Differences (EB-diff) norms: These focus on how physically involved a word's meaning is, particularly for verbs. The hypothesis is that verbs with different levels of bodily involvement might be more prone to metaphorical usage.

These conceptual features are then combined with traditional NLP features like n-grams, word embeddings (Word2Vec, GloVe), and cosine similarity. The system uses standard classifiers like Logistic Regression, Linear SVC, and Random Forest, with careful parameter tuning.

Datasets

The paper uses two corpora for the shared task on Metaphor Detection:

  • VUA: The Victoria University of Wellington corpus.
  • TOFEL: The Toronto English Corpus.

These corpora are divided into two tracks: "Verbs" (focusing on verb metaphors) and "All-POS" (considering metaphors across all parts of speech). The authors participated in all four phases of the competition.

Key Insights and Ideas

  • Conceptual Norms are Powerful: The most significant finding is that incorporating modality exclusivity and embodiment norms dramatically improves metaphor detection. This highlights the importance of cognitive linguistic theories in building NLP systems.
  • Modality Exclusivity Shines: The Modality Exclusivity (ME) feature set emerged as the single best performing feature, demonstrating its strong predictive power for metaphoricity.
  • Embodiment Data Sparseness is a Hurdle: While embodiment norms are conceptually relevant, their application was limited by data sparseness, especially for non-verb words. This suggests a need for more comprehensive embodiment datasets.
  • Simplicity Can Be Effective: The authors achieved state-of-the-art results without resorting to complex deep learning architectures, showcasing the efficiency and effectiveness of their linguistically-enriched feature sets with traditional classifiers.
  • Metaphor as Concept Mismatch: The results strongly support the hypothesis that metaphor detection is fundamentally about identifying mismatches between source and target conceptual domains.

Key Results

The proposed system achieved strong performance on the VUA dataset, with an F-score of 0.652 for the VUA Verbs track, which was 5% higher than strong baselines. When combining features, the "Selected" feature set (which included linguistic and collocational features but excluded word embeddings) achieved the best F1-score of 0.672 for Logistic Regression on the VUA Verbs phase. The paper also reports competitive results on the test sets across all four phases, outperforming baseline methods and demonstrating the effectiveness of their approach.

Limitations and Future Work

The authors acknowledge that the embodiment features performed poorly due to data sparseness, as the available norms only covered a limited set of English verbs. This suggests a need for more comprehensive embodiment datasets that include a wider range of word types.

For future work, the authors plan to integrate their current feature sets with state-of-the-art deep learning models to further explore their effectiveness and potentially achieve even higher performance. They also aim to investigate the contribution of other conceptual features for metaphor detection.

Summary updated

BibTeX
@inproceedings{wan2020using,
  author = {Mingyu Wan and Kathleen Ahrens and Emmanuele Chersoni and Menghan Jiang and Qi Su and Rong Xiang and Chu-Ren Huang},
  title = {{Using Conceptual Norms for Metaphor Detection}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure