Leveraging Syntactic Constructions for Metaphor Identification

Kevin Stowe, Martha PalmerNAACL 2018

TL;DR

This paper argues that the way words are arranged in a sentence (syntax) is a powerful, yet often overlooked, signal for identifying metaphorical language. By analyzing predicate-argument constructions, the authors demonstrate that syntactic patterns can significantly improve the accuracy of metaphor detection and help distinguish between literal and figurative language.

What Problem Does This Paper Address?

Understanding metaphor is crucial for natural language understanding (NLU). While many computational approaches to metaphor identification rely on word meanings (lexical semantics) and basic context, they often miss the subtle cues provided by sentence structure. Metaphors, like "argument is war," where we talk about "defending" a position or "attacking" a viewpoint, are deeply embedded in how we construct sentences. Existing methods struggle to leverage this syntactic information effectively, leading to limitations in accurately identifying metaphorical language and its underlying source and target domains. This paper aims to bridge this gap by exploring how specific syntactic constructions can be used as features for metaphor identification.

Methods

The core of this research lies in analyzing predicate-argument constructions – the relationships between a verb (predicate) and its subjects and objects. The authors propose that these constructions, rather than just individual words, carry significant metaphorical weight.

Their approach involves:

  • Corpus Analysis: Examining large, annotated corpora (LCC and VUAMC) to understand the distribution of metaphoric language within different syntactic constructions.
  • Syntactic Feature Engineering: Developing novel ways to represent syntactic structures as features for machine learning models. This includes:
  • Predicate-Argument Structure: Identifying the type of construction (intransitive, transitive, ditransitive, equative, simile) a verb belongs to.
  • Head and Dependent Features: Incorporating information about the word being classified and its related words.
  • Dependency Relations: Using specific grammatical relationships (like subject, direct object) to capture syntactic nuances.
  • VerbNet Classes: Leveraging a lexical resource that categorizes verbs based on their syntactic behavior.
  • Machine Learning Classification: Training classifiers (specifically Support Vector Machines) to identify metaphoric words, source words, and target words using a combination of lexical semantics (word embeddings) and the engineered syntactic features.

Datasets

The paper utilizes two primary datasets:

  • LCC Metaphor Corpus: This corpus contains approximately 7,500 source/target metaphor pairs with sentential context, rated for metaphoricity. It was used for analyzing source and target domain identification.
  • VUAMC Corpus: A larger corpus (approximately 200,000 words) where each word is tagged as either metaphoric or non-metaphoric. This was used for large-scale analysis of metaphoricity at the word level and for metaphor identification tasks.

Key Insights and Ideas

  • Syntax is a Strong Predictor: The arrangement of words in a sentence is not just grammatical; it's a powerful indicator of metaphorical language. Certain syntactic constructions are inherently more likely to be used metaphorically.
  • Verbs as Metaphor Drivers: Verbs, especially in their predicate-argument structures, are key players in metaphor. The type of construction a verb participates in significantly influences its likelihood of being used metaphorically.
  • Source vs. Target Domain Differences: Identifying source domains (the conceptual domain used to understand another) is generally harder than identifying target domains, likely due to the wider variety of source domains.
  • Constructional Patterns Matter: While lexical semantics are important, they are insufficient on their own. Syntactic patterns, like the presence of direct objects in equative constructions, show clear parallels with metaphorical usage.
  • Combining Syntax and Semantics is Key: The best performance is achieved when both lexical semantics (word embeddings) and rich syntactic features are combined, highlighting the synergistic nature of language understanding.

Key Results

  • Improved Classification: The addition of syntactic features consistently improved classification performance for identifying metaphoric words, source words, and target words compared to a baseline using only lexical semantics.
  • LCC Analysis: In the LCC corpus, metaphoric verbs were predominantly source domain items (92%), while subjects and objects were more often target domain items (61% each).
  • VUAMC Analysis: Verbs in ditransitive constructions were found to be more likely to be metaphoric than those in intransitive or transitive constructions.
  • Best Combination: The paper reports that combining all tested syntactic features with lexical semantics yielded the best classification results across various tasks. For example, in the LCC corpus, the "Best Combination" achieved scores of .551 for Source, .600 for Target, and .505 for Metaphor classification.

Limitations and Future Work

  • Limited Construction Coverage: The automatic identification of constructions was limited to basic types (intransitive, transitive, etc.) due to reliance on dependency parses. More complex constructions, which can be highly metaphoric, were not fully captured.
  • Syntactic Feature Inconsistency: While some syntactic features showed consistent improvement, others were less reliable, particularly in predicting verb metaphoricity.
  • Future Directions: The authors suggest exploring more advanced representations of syntax-semantics interaction, such as syntactic tree kernels and compositional distributional semantic models. They also propose using ensemble methods to select the most effective classifiers for different syntactic constructions and extending the approach to identify source and target domains more comprehensively.

Summary updated

BibTeX
@inproceedings{stowe2018leveraging,
  author = {Kevin Stowe and Martha Palmer},
  title = {{Leveraging Syntactic Constructions for Metaphor Identification}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
}
Figure