Enriching Conceptual Knowledge in Language Models through Metaphorical Reference Explanation

Zixuan Zhang, Heng Ji

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

In this paper, we explore a novel approach to concept enrichment in language models (LMs) by leveraging the fundamental similarities between conceptual knowledge enrichment and metaphorical reference resolution. While previous knowledge editing (KE) methods predominantly focus on factual updates, we introduce a method that trains LMs to not only incorporate new conceptual meanings but also generatively explain the connections between original and enriched definitions through metaphorical analogies. To achieve this, we develop a new dataset tailored for concept enrichment tasks and apply it to train an LM capable of updating and reasoning about conceptual knowledge. The proposed method was evaluated on both "is-a" relation classification and metaphorical reference detection.

Original languageEnglish (US)
Title of host publicationProceedings of the 7th Workshop on Computational Models of Reference, Anaphora and Coreference, CRAC 2024
EditorsMaciej Ogrodniczuk, Anna Nedoluzhko, Massimo Poesio, Sameer Pradhan, Vincent Ng
PublisherAssociation for Computational Linguistics (ACL)
Pages18-22
Number of pages5
ISBN (Electronic)9798891761711
DOIs
StatePublished - 2024
Event7th Workshop on Computational Models of Reference, Anaphora and Coreference, CRAC 2024 - Miami, United States
Duration: Nov 15 2024 → …

Publication series

NameProceedings of the 7th Workshop on Computational Models of Reference, Anaphora and Coreference, CRAC 2024

Conference

Conference7th Workshop on Computational Models of Reference, Anaphora and Coreference, CRAC 2024
Country/TerritoryUnited States
CityMiami
Period11/15/24 → …

ASJC Scopus subject areas

  • Artificial Intelligence
  • Modeling and Simulation

Fingerprint

Dive into the research topics of 'Enriching Conceptual Knowledge in Language Models through Metaphorical Reference Explanation'. Together they form a unique fingerprint.

Cite this