Lexical Conceptual Structure of Literal and Metaphorical Spatial Language: A Case Study of “Push”

Bonnie Dorr, Mari OlsenNAACL 2018

TL;DR

This paper proposes a unified framework, Lexical Conceptual Structure (LCS), to represent both literal and metaphorical meanings of spatial language. By analyzing the verb "push" across a large corpus, they demonstrate that metaphorical extensions retain structural similarities to their literal counterparts, enabling more robust natural language understanding.

What Problem Does This Paper Address?

Understanding and processing language that uses spatial concepts metaphorically (e.g., "her job pushed her over the edge") is a significant challenge in natural language processing (NLP). Traditionally, literal spatial language and its metaphorical extensions have been treated separately. This disconnect hinders the development of NLP systems that can accurately interpret the nuances of human language, especially when abstract concepts are expressed through concrete spatial metaphors. The paper aims to bridge this gap by providing a unified representational framework that captures the underlying semantic structure of both literal and metaphorical expressions.

Methods

The core of the paper's methodology lies in leveraging Lexical Conceptual Structure (LCS), a linguistic theory that breaks down verb meanings into a set of primitive components. The authors extend this by:

  • Building on MetaNet (MN): They utilize MN, a repository of conceptual metaphors, to identify and categorize spatial metaphors.
  • Systematic LCS Representation: They map both literal (spatial) and metaphorical senses of verbs to LCS representations. This involves identifying common LCS primitives that underlie both types of meaning.
  • Corpus Analysis: They analyze a corpus of "Push Sentences" (sentences containing the word "push") to identify real-world instances of literal and metaphorical usage.
  • Derived Classes: They create 16 semantically motivated classes that group pairs of literal and metaphorical LCS representations. These classes help organize and understand the relationships between different uses of a verb.

The novelty lies in the systematic application of LCS to bridge the literal-metaphorical divide, demonstrating that the syntactic patterns and structural components of literal spatial language are often preserved in their metaphorical extensions.

Datasets

  • MetaNet (MN): A repository of conceptual metaphors, from which 197 spatially grounded metaphors were identified and categorized.
  • "Push Sentences" Corpus: A collection of approximately 1655 sentences containing the word "push," mined from a larger corpus of 30k Microsoft language resource library documents. This corpus was used to find natural occurrences of "push" and to supplement MN with examples not found in the repository.

Key Insights and Ideas

  • Unified Representation: LCS provides a powerful framework for representing both literal and metaphorical meanings under a single, consistent structure. This is a significant step towards a unified theory of language meaning.
  • Structural Similarities: Metaphorical extensions of spatial language often retain the core structural components and syntactic patterns of their literal counterparts. For example, "push over the edge" (literal) and "Mary's job pushed her over the edge" (metaphorical) share underlying LCS structures related to causation and change of state/location.
  • Predictive Power: By systematically mapping metaphors to LCS, the framework can predict potential metaphorical extensions and identify gaps in existing metaphor repositories like MN.
  • Beyond Spatial: The LCS framework, with its extensions to different "fields" (e.g., Identificational, Circumstantial), allows for the systematic mapping of spatial meanings to abstract domains, explaining how spatial concepts are used to understand non-spatial phenomena.

Key Results

  • The analysis of "Push Sentences" revealed that 52% of verb occurrences were metaphorical, highlighting the prevalence of metaphorical language.
  • Across all parts of speech, almost 40% of "push" uses were metaphorical.
  • The study identified 16 semantically motivated classes of literal/metaphorical LCS pairings, demonstrating a structured way to organize and understand these relationships.
  • The research successfully identified metaphorical instances of "push" not present in the MN repository, which were then classified using the LCS-based framework.

Limitations and Future Work

  • Disambiguation: The paper focuses on representation and doesn't delve deeply into the processes required for disambiguating literal from metaphorical senses, suggesting this as a direction for future work, potentially involving context, visual grounding, or belief systems.
  • Completeness of MN: The study revealed that MN, while extensive, is not exhaustive, and the "Push Sentences" corpus helped to identify missing examples.
  • Multilingual Representation: The authors propose future work on identifying multilingual equivalents of the identified spatial expressions to build cross-lingual mappings and explore how metaphorical extensions vary across languages.
  • Continuous Dimension: The research suggests a continuum between literal and metaphorical language, advocating for future work that incorporates a more continuous dimension in representation beyond discrete categories.

Summary updated

BibTeX
@inproceedings{dorr2018lexical,
  author = {Bonnie Dorr and Mari Olsen},
  title = {{Lexical Conceptual Structure of Literal and Metaphorical Spatial Language: A Case Study of “Push”}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
}
Figure