Contextual Modulation for Relation-Level Metaphor Identification

Omnia Zayed, John P. McCrae, Paul BuitelaarEMNLP 2020 · DOI

Introduction

Metaphor identification is a critical task in Natural Language Processing (NLP), underpinning applications from sentiment analysis to machine translation. Traditionally, this task is approached either at the word level (labeling individual tokens) or the relation level (analyzing pairs like adjective-noun or verb-object). However, existing relation-level models often suffer from "tunnel vision," focusing exclusively on the interaction between the two words in a pair—the "tenor" and the "vehicle"—while neglecting the broader sentence context.

Context is frequently the deciding factor in metaphoricity. For instance, the phrase "the rooms communicated" is literal in an architectural context but metaphorical in a literary one. To address this, researchers have developed a new neural architecture that identifies metaphors by examining specific grammatical relationships while simultaneously "tuning" the model based on the surrounding sentence context. By adapting Feature-wise Linear Modulation (FiLM)—a technique originally designed for visual reasoning—the proposed model achieves state-of-the-art results across several benchmark datasets.

Methodology: The Contextual Modulation Pipeline

The core innovation of this work is the Contextual Modulation Pipeline, which utilizes FiLM to bridge the gap between local relations and global context. The architecture consists of two parallel tracks:

  1. The Linguistic Pipeline: This primary track processes the entire sentence. Words are embedded using GloVe and passed through a Long Short-Term Memory (LSTM) network to capture sequential dependencies.
  2. The Contextual Modulation Pipeline: This auxiliary track focuses on the specific candidate expression (e.g., a verb-noun pair). It generates a deep contextualized representation using ELMo embeddings.
The dual-pipeline architecture featuring the Linguistic Pipeline and the Contextual Modulation Pipeline using FiLM.

The integration occurs in the Affine Transformation layer. The modulation pipeline generates two parameters, $\gamma$ (scaling) and $\beta$ (shifting), based on the candidate expression. These parameters are applied to every hidden state of the main LSTM. In effect, the candidate expression acts as a "remote control" that dynamically adjusts the neural features of the entire sentence, highlighting contextual clues relevant to metaphor identification. While an optional attention mechanism was tested, the FiLM-based modulation was found to capture the majority of necessary information.

Datasets and Evaluation

The model was evaluated on five distinct datasets to ensure robustness:

  • TSV: Focused on adjective-noun relations.
  • ZayTw: A new dataset of 2,531 tweets annotated for verb-direct object metaphors, spanning general and political topics.
  • Adapted Word-Level Datasets: Filtered versions of MOH, TroFi, and VUAMC, adapted to focus on relation-level (verb-noun) dependencies.

Key Results and Insights

The proposed AffineTrans model consistently outperformed the previous state-of-the-art, the Supervised Similarity Network (SSN).

Performance comparison of the AffineTrans model against baselines on ZayTw and TSV datasets.

On the TSV dataset, AffineTrans achieved an F1-score of 0.834 (vs. 0.811 for SSN). On the new ZayTw dataset, it reached an F1-score of 0.786 (vs. 0.704 for SSN). The model showed statistically significant improvements ($p < 0.01$) on three of the five benchmarks.

Key takeaways include:

  • FiLM vs. Attention: FiLM proved more expressive than standard attention. By using more parameters to scale and shift features, it conditions the model's computation more effectively on the specific candidate.
  • The Necessity of Context: Removing the modulation layer led to a significant drop in performance, confirming that the interaction between specific relations and wider context is essential.
  • Robustness to Noise: The strong performance on the ZayTw dataset suggests the architecture can handle the informal and idiosyncratic language of social media.

Limitations and Future Directions

Despite its success, the model faces challenges with abstract verbs related to internal cognitive or emotional states (e.g., "believe," "experience"), where literal and metaphorical uses are subtly distinguished. Additionally, some benchmark test sets are small, making metrics sensitive to individual errors. Future work may explore replacing the LSTM base with more complex transformer architectures like BERT to further enhance the encoding of contextual information.

Summary updated

BibTeX
@inproceedings{zayed2020contextual,
  author = {Omnia Zayed and John P. McCrae and Paul Buitelaar},
  title = {{Contextual Modulation for Relation-Level Metaphor Identification}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  doi = {10.18653/v1/2020.findings-emnlp.36},
}
Figure