No clues good clues: out of context Lexical Relation Classification

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5607–5625 July 9-14, 2023 ©2023 Association for Computational Linguistics No clues, good clues: Out of context Lexical Relation Classification Lucía Pitarch1, Jorge Bernad1, Licri Dranca2, Carlos Bobed Lisbona1, and Jorge Gracia1 1University of Zaragoza, Zaragoza, Spain 2 Centro Universitario de la Defensa (CUD), Zaragoza, Spain {lpitarch,jbernad,licri,cbobed,jogracia}@unizar.es Abstract The accurate prediction of lexical relations be- tween words is a challenging task in Natural Language Processing (NLP). The most recent advances in this direction come with the use of pre-trained language models (PTLMs). A PTLM typically needs “well-formed" verbal- ized text to interact with it, either to fine-tune it or to exploit it. However, there are indica- tions that commonly used PTLMs already en- code enough linguistic knowledge to allow the use of minimal (or none) textual context for some linguistically motivated tasks, thus no- tably reducing human effort, the need for data pre-processing, and favoring techniques that are language neutral since do not rely on syn- tactic structures. In this work, we explore this idea for the tasks of lexical relation classification (LRC) and graded Lexical Entailment (LE). After fine- tuning PTLMs for LRC with different verbal- izations, our evaluation results show that very simple prompts are competitive for LRC and significantly outperform graded LE SoTA. In order to gain a better insight into this phe- nomenon, we perform a number of quantitative statistical analyses on the results, as well as a qualitative visual exploration based on embed- ding projections. 1 Introduction Lexical Relation Classification (LRC) is the task of predicting which lexical relation exists between two given words (e.g., ‘tall’ and ‘small’ are related by the antonymy relation), from a finite catalogue of lexical relations. Discovering lexico-semantic re- lations between words has received attention in the NLP community since Hearst’s seminal research in 1992 on the automatic acquisition of hyponyms from large text corpora based on pre-designed pat- terns (Hearst, 1992). Despite many recent advance- ments, LRC continues to be an open research topic in the NLP field (Wang et al., 2021; Ushio et al., 2021). Applications of the task are numerous: auto- matic thesauri creation, paraphrasing, textual entail- ment, sentiment analysis, ontology learning, and ontology population, among others (Weeds et al., 2014; Cimiano, 2006). The most recent advances in LRC come with the use of pre-trained language models (PTLMs) based on the transformers architecture (Vaswani et al., 2017), which have been proven to capture a large amount of lexico-semantic knowledge from text successfully. One of the main benefits of the adoption of PLTMs is that, while they were trained for a general task (text generation) following a masked language model (MLM) objective in an unsupervised way, they can be easily adapted to different downstream tasks (e.g., text classification, text summarization, sentiment analysis) by intro- ducing additional parameters and fine-tuning them using objective functions specific to the task. That avoids the need to train the model from scratch, still obtaining SoTA results, while …

Content truncated to 500 words