An analogy based approach for solving target sense verification - Grid'5000 Access content directly
Conference Papers Year : 2022

An analogy based approach for solving target sense verification

Abstract

Contextualized language models have emerged as a de facto standard in natural language processing due to the vast amount of knowledge they acquire during pretraining. Nonetheless, their ability to solve tasks that require reasoning over this knowledge is limited. Certain tasks can be improved by analogical reasoning over concepts, e.g., understanding the underlying relations in "Man is to Woman as King is to Queen". In this work, we propose a way to formulate target sense verification as an analogy detection task, by transforming the input data into quadruples. We present AB4TSV (Analogy and BERT for TSV), a model that uses BERT to represent the objects in these quadruples combined with a convolutional neural network to decide whether they constitute valid analogies. We test our system on the WiC-TSV evaluation benchmark, and show that it can outperform existing approaches. Our empirical study shows the importance of the input encoding for BERT. This dependence gets alleviated by integrating the axiomatic properties of analogies during training, while preserving performance and improving interpretability.
Fichier principal
Vignette du fichier
zervakis_NLPIR22.pdf (757.01 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-03792071 , version 1 (29-09-2022)

Identifiers

  • HAL Id : hal-03792071 , version 1

Cite

Georgios Zervakis, Emmanuel Vincent, Miguel Couceiro, Marc Schoenauer, Esteban Marquer. An analogy based approach for solving target sense verification. NLPIR 2022 - 6th International Conference on Natural Language Processing and Information Retrieval, Dec 2022, Bangkok, Thailand. ⟨hal-03792071⟩
344 View
162 Download

Share

Gmail Facebook X LinkedIn More