[Paper Review] Generating Conceptual Metaphors from Proposition Stores

Ekaterina Ovchinnikova, Vladimir Zaytsev+2 more|arXiv (Cornell University)|Sep 25, 2014

Language, Metaphor, and Cognition34 references7 citations

TL;DR

This paper presents an unsupervised method for automatically generating conceptual metaphors from proposition stores—structured linguistic data representing semantic relations—without relying on manually coded resources or annotated training data. The approach identifies potential source domains by matching property similarities between target concepts and generates corresponding linguistic metaphors, which are validated by experts, achieving high accuracy in both English and Russian evaluations.

ABSTRACT

Contemporary research on computational processing of linguistic metaphors is divided into two main branches: metaphor recognition and metaphor interpretation. We take a different line of research and present an automated method for generating conceptual metaphors from linguistic data. Given the generated conceptual metaphors, we find corresponding linguistic metaphors in corpora. In this paper, we describe our approach and its evaluation using English and Russian data.

Motivation & Objective

• To develop an automated, unsupervised method for generating conceptual metaphors from raw linguistic data without reliance on manually curated lexical-semantic resources.
• To enable the discovery of linguistic metaphors in corpora based on generated conceptual metaphors, supporting both metaphor recognition and interpretation.
• To evaluate the method’s effectiveness using both English and Russian language data, ensuring cross-linguistic applicability.
• To provide an open-source toolset and validated resources for community use in metaphor generation and analysis.
• To assess the correlation between computed metaphor ranking scores and human judgments of metaphor salience.

Proposed method

• The method constructs proposition stores from parsed sentences, extracting weighted tuples of words with syntactic relations to represent common-sense knowledge about concepts.
• For each target concept, it collects a weighted list of propositions describing its properties, then identifies potential source domains with similar property profiles.
• It computes similarity between target and candidate source concepts based on shared propositional features, using a weighted matching mechanism to rank potential conceptual metaphors.
• Given a generated conceptual metaphor, the system searches corpora for linguistic realizations (e.g., phrases like 'the economy moves forward') that express the mapping.
• The method does not require predefined source or target domains, nor does it depend on selectional preferences or hand-coded knowledge.
• Validation is performed by expert linguists who assess the relevance and naturalness of found linguistic metaphors.

Experimental results

Research questions

• RQ1 Can conceptual metaphors be reliably generated from proposition stores without relying on manually coded lexical-semantic resources?
• RQ2 How well do the automatically generated conceptual metaphors align with established conceptual metaphors from the Master Metaphor List and CorMet?
• RQ3 To what extent can the system discover valid linguistic metaphors in corpora based on generated conceptual metaphors?
• RQ4 How effective is the proposed ranking mechanism in identifying the most plausible source domains for a given target concept?
• RQ5 How well does the similarity in propositional properties between target and source domains predict the perceived salience of a conceptual metaphor?

Key findings

• The system achieved a 76% success rate in validating Russian linguistic metaphors for top-ranked sources, with 141 out of 185 total metaphors judged correct by experts.
• For English, 83% of the linguistic metaphors found for top-ranked sources were validated as correct, with 163 out of 197 metaphors judged accurate.
• The method successfully generated 12 out of 14 conceptual metaphors from the Master Metaphor List, demonstrating strong alignment with established metaphorical mappings.
• The system outperformed CorMet in identifying mappings such as 'water → emotion' and 'fight → love', with higher confidence scores and better coverage of natural expressions.
• The proposed method achieved a 99% confidence score for the mapping 'fight → emotion', indicating strong alignment with natural metaphorical usage in corpora.
• The approach demonstrated cross-linguistic robustness, successfully generating and validating metaphors in both English and Russian with consistent performance.