Introduction
Metaphor detection (MD) involves identifying when a word is used in a non-literal, analogical sense—for instance, distinguishing between "curing juvenile delinquency" and "curing a disease." This task is a cornerstone of natural language understanding but remains challenging due to the heavy reliance on context and semantic nuance.
Existing deep learning approaches often suffer from two primary limitations: contextual noise and task isolation. Standard sequence models like LSTMs treat all surrounding words with similar linear weight, potentially missing "clue" words that are syntactically close but linearly distant. Furthermore, most models treat metaphor detection in isolation, ignoring its inherent relationship with Word Sense Disambiguation (WSD).
The MUL_GCN Architecture
To address these gaps, this paper introduces MUL_GCN, a multi-layered model that integrates syntactic structure and multi-task learning.
The architecture consists of four primary modules:
- Encoding Module: Utilizes a combination of static (GloVe) and contextual (ELMo) embeddings, along with a specialized "index embedding" to highlight the target word. These are processed through a BiLSTM layer.
- Graph Convolution Module: Employs Graph Convolutional Networks (GCNs) over the sentence's dependency parse tree. This allows the model to capture long-range syntactic dependencies that are often crucial for identifying metaphors.
- Control Module: A novel filtering mechanism that generates a "control vector" based on the target word. This vector filters the GCN and BiLSTM representations, ensuring the model retains only the information most relevant to the specific word being analyzed.
- Multi-Task Learning (MTL) Module: The model is trained simultaneously on metaphor detection and WSD. A Mean Squared Error (MSE) loss term encourages the encoders for both tasks to produce similar representations for the same input, effectively transferring semantic knowledge from the broader WSD task to metaphor detection.
Why Syntax Matters
A key insight of this work is that syntactic neighbors often provide more reliable clues for metaphoricity than linear neighbors. In complex sentences, the words that determine a target's meaning might be separated by several modifiers.
As shown in the example above, the word "house" is the syntactic subject of "lay," providing the essential context to determine that "lay" is used metaphorically. While an LSTM must process the intervening phrase "that at first seemed to float on a raft of golden mist," the GCN directly connects "house" to "lay" via the dependency tree.
Experimental Evaluation
The model was evaluated on four standard benchmarks: VUA (All POS and Verb), MOH-X, and TroFi. For the multi-task component, the Semcor dataset provided the WSD annotations.
Key Results
MUL_GCN achieved state-of-the-art (SOTA) performance across all datasets:
- VUA Verb: Reached an F1 score of 71.7, improving upon the previous best of 70.5.
- TroFi: Achieved an F1 of 73.2, a significant jump from the previous SOTA of 72.4.
Ablation studies confirmed the necessity of each component. Removing the GCN, Control Module, or Multi-task learning led to measurable performance declines. Notably, the multi-task learning approach provided up to a 2.7% absolute improvement in F1 score compared to single-task baselines, validating the hypothesis that metaphor detection benefits from WSD-informed semantic representations.
Conclusion and Future Work
By combining syntactic graph convolutions with multi-task learning, MUL_GCN effectively captures the complex contextual relationships required for metaphor detection. Future research may explore expanding the multi-task framework to other related NLP tasks and addressing the challenges of detecting multi-word metaphorical expressions.
Summary updated