TL;DR
This paper introduces a novel machine learning approach to automatically interpret metaphors by focusing on the grammatical structure of sentences, specifically argument-structure cues. By combining these cues with a deep learning model, the researchers achieved a significant improvement in accuracy for identifying the source domain of metaphors, demonstrating the power of construction grammar in computational linguistics.
What Problem Does This Paper Address?
Understanding and interpreting metaphors computationally is a long-standing challenge in natural language processing. While metaphor detection has seen progress, the interpretation of how a metaphor works – specifically, identifying the source domain (e.g., "journey" in "love is a journey") and its mapping to the target domain (e.g., "love") – remains difficult. Existing methods often rely on large databases of metaphoric examples or are limited to specific types of metaphors. This paper aims to bridge this gap by developing a more generalizable method for metaphor interpretation that leverages linguistic theory, specifically usage-based construction grammar, to extract meaningful features from sentence structure.
Methods
The core of this research lies in a novel feature engineering approach combined with a feed-forward deep neural network (DNN).
- Feature Engineering: Instead of relying solely on word embeddings or pre-defined metaphor databases, the authors extract features directly from the grammatical structure of sentences. They use dependency parsing to identify the verb associated with the target domain referent and its arguments (subject, object, prepositional arguments, nominal modifiers). Crucially, they then create interaction features between the target domain referent and these grammatical elements (e.g., "target domain NP x subject," "target domain NP x verb x object"). These features are designed to capture the "constructional makeup" of the metaphor, reflecting how the grammatical roles and relationships within a sentence guide its interpretation.
- Deep Neural Network: These engineered features are then converted into embeddings and fed into a feed-forward DNN. The network consists of a single hidden layer with 77 units and a dropout rate for regularization. The output layer predicts one of the 77 source domain labels.
Datasets
The study utilizes the EN-Small LCC Metaphor Dataset, compiled and annotated by Mohler et al. (2016). This dataset contains government discourse examples with metaphoricity scores and, importantly for this study, annotations for source-target domain mappings. The authors filtered the dataset to include only metaphoric instances with annotated mappings, resulting in 7,941 sentences and 108 unique source domain tags. They used 85% of this data for training and 15% for testing.
Key Insights and Ideas
- Grammatical Structure as a Metaphor Interpreter: The most significant insight is that the syntactic structure and argument roles within a sentence provide powerful cues for metaphor interpretation, even without explicit semantic knowledge of the words themselves.
- Interaction Features are Key: Simply providing individual grammatical elements (like the verb or subject) wasn't enough. The interactions between the target domain referent and these elements (e.g., how "poverty" relates to the verb "pushing" and its object "people") proved to be the most predictive features.
- Construction Grammar's Practical Value: The research demonstrates the practical utility of construction grammar, a linguistic theory that emphasizes form-meaning pairings in language, for computational tasks like metaphor interpretation.
- Potential for Direct Interpretation: The authors suggest that this approach might eventually bypass the need for explicit metaphor detection, allowing direct prediction of source domains for specific target concepts.
Key Results
The proposed method, combining constructional cues with a feed-forward DNN, achieved an accuracy of 80.4% in predicting the source domain label. This represents a significant improvement over a baseline DNN that used only argument structure features without the interaction terms (achieving 78.9% accuracy). An LSTM network without these constructional features performed even lower (73.0% highest accuracy).
Limitations and Future Work
- Inter-Annotator Agreement: A significant limitation acknowledged is the relatively low average inter-annotator agreement (54.4% Cohen's Kappa) for source domain mappings in the LCC dataset. This suggests inherent ambiguity in some metaphoric interpretations, which can impact model performance.
- Exclusion of Multi-Word Expressions (MWEs): The study excluded MWEs from the dataset for simplicity, which limits the generalizability of the current results. The authors express a desire to re-test their methods with MWEs included.
- Limited Source Domain Labels: The exclusion of rare source domain labels also reduced the total number of unique labels.
- Future Work: The authors suggest integrating this system with metaphor detection or target domain identification tools to create a more comprehensive predictive tool. They also plan to explore the inclusion of MWEs and potentially other linguistic theories.
Summary updated