Unsupervised Metaphor Identification Using Hierarchical Graph Factorization Clustering

Ekaterina Shutova, Lin SunNAACL 2013

TL;DR

This paper introduces a novel, fully unsupervised method for identifying metaphorical associations and expressions in text. By leveraging hierarchical graph factorization clustering and cognitive science findings on abstract vs. concrete concept organization, the system achieves competitive precision and recall without requiring labeled training data.

What Problem Does This Paper Address?

Metaphor is a fundamental aspect of human language and reasoning, allowing us to understand abstract concepts in terms of more concrete ones (e.g., "argument" as "war"). While computational approaches to metaphor identification exist, they often rely on supervised or semi-supervised methods, demanding extensive labeled datasets that are costly and time-consuming to create. Furthermore, many existing methods treat all concepts similarly, neglecting the cognitive evidence that abstract and concrete concepts are processed differently in the brain. This paper aims to bridge this gap by developing an unsupervised method that accounts for these cognitive distinctions and can identify both conceptual metaphors and their linguistic expressions in unrestricted text.

Methods

The core of the proposed method is Hierarchical Graph Factorization Clustering (HGFC). Here's a breakdown of the key steps:

  1. Concept Graph Construction: HGFC is applied to a set of frequent nouns to build a hierarchical graph of concepts. Unlike traditional hierarchical clustering that forms strict trees, HGFC creates a more flexible graph where concepts can have multiple parents at different levels. This is inspired by cognitive science findings suggesting abstract concepts have more complex association patterns than concrete ones.
  2. Metaphorical Association Identification: The system traverses this concept graph, using edge weights (representing association strength) to identify potential metaphorical links between clusters of concepts.
  3. Salient Feature Extraction: For identified metaphorical associations, the system extracts "salient features" (primarily verbs and their grammatical relations) that strongly link the source and target concepts. These features are ranked based on their probability of co-occurrence with both the source and target.
  4. Metaphorical Expression Identification: The extracted salient features are then used to search a large corpus (the British National Corpus - BNC) for actual metaphorical expressions that exemplify the identified conceptual metaphors. This step involves filtering features using selectional preference to retain those that are truly descriptive of the source domain.

Datasets

  • Training/Clustering: The 2000 most frequent nouns from the British National Corpus (BNC) were used for clustering.
  • Feature Extraction: Grammatical relations were extracted from the Gigaword corpus using the RASP parser.
  • Evaluation: The British National Corpus (BNC) was used for identifying metaphorical expressions.

Key Insights and Ideas

  • Cognitive Plausibility: The method explicitly incorporates cognitive science findings on the differential organization of abstract and concrete concepts in the brain, leading to a more nuanced representation of conceptual relationships.
  • Unsupervised Learning: The entire pipeline is unsupervised, eliminating the need for costly manual annotation and making it more robust and applicable to real-world scenarios.
  • Hierarchical Graph Factorization Clustering (HGFC): This technique is novel for noun clustering in NLP and proves effective in building a flexible, multi-level concept graph that captures complex associations.
  • Salient Features for Metaphor: Identifying verbs and their grammatical relations as "salient features" provides a direct link between abstract conceptual metaphors and concrete linguistic expressions.

Key Results

  • Metaphorical Association Identification: The system achieved a precision of 0.69 and a recall of 0.61 in identifying metaphorical associations, which is significantly higher than the AGG (P=0.36, R=0.11) and WN (P=0.29, R=0.03) baselines.
  • Metaphorical Expression Identification: The system's performance was close to previous supervised systems, with a precision of 0.65 against human annotations, outperforming AGG (P=0.47) and WN (P=0.12).
  • Human Ceiling: The human ceiling for judging metaphorical associations was P=0.80, indicating the inherent difficulty of the task.

Limitations and Future Work

  • Lexicalized Metaphors: The system struggles with identifying already lexicalized metaphors (e.g., verbs like "impose," "decline"), which can lead to disagreements with human annotators.
  • Multiple Word Senses: Ambiguity in word senses for salient features or source/target concepts can lead to errors.
  • Chains of Mappings: The current system does not explicitly handle chains of metaphors (e.g., A is B, and B is C).
  • Future Work:
    • Investigate using the created conceptual graph for metaphor interpretation.
  • Extend the approach to cover nominal and adjectival metaphors by harvesting salient nominal and adjectival features.
  • Address the issue of lexicalized metaphors by incorporating topical similarity filtering.

Summary updated

BibTeX
@inproceedings{shutova2013unsupervised,
  author = {Ekaterina Shutova and Lin Sun},
  title = {{Unsupervised Metaphor Identification Using Hierarchical Graph Factorization Clustering}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2013},
}
Figure