LastResort at SemEval-2022 Task 5: Towards Misogyny Identification using Visual Linguistic Model Ensembles And Task-Specific Pretraining

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 68 - 74 July 14-15, 2022 ©2022 Association for Computational Linguistics MMG at SemEval-2022 Task 1: A Reverse Dictionary approach based on a review of the dataset from a lexicographic perspective Alfonso Ardoiz1,2 alfonso.ardoiz@dezzai.com aardoiz@ucm.es Miguel Ortega-Mart´ın1,2 m.ortega@dezzai.com m.ortega@ucm.es ´Oscar Garc´ıa-Sierra1,2 oscar.garcia@dezzai.com oscarg02@ucm.es Jorge ´Alvarez1 jorge.alvarez@dezzai.com Ignacio Arranz1 ignacio.arranz@dezzai.com Adri´an Alonso1,3,4 a.alonso@dezzai.com adrian.barriuso@urjc.es 1 dezzai by MMG 3 Universidad Rey Juan Carlos 2 Universidad Complutense de Madrid 4 Data Science Laboratory - Universidad Rey Juan Carlos Abstract This paper presents a novel and linguistic-driven system for the Spanish Reverse Dictionary task of SemEval-2022 Task 1. The aim of this task is the automatic generation of a word using its gloss. The conclusion is that this task results could improve if the quality of the dataset did as well by incorporat- ing high-quality lexicographic data. Therefore, in this paper we analyze the main gaps in the proposed dataset and describe how these limitations could be tackled. 1 Introduction The CODWOE (Comparison of Word Glosses and Word Embeddings) task at SemEval-2022 (Mickus et al., 2022) encouraged participants to analyze the relation between two types of semantic descrip- tions, word embeddings and dictionary glosses, by proposing two sub-tasks: Reverse Dictionary (RD) (Hill et al., 2016), in which participants must gener- ate vectors from glosses, and Definition Modeling (DM) (Noraset et al., 2017), in which participants must generate glosses from vectors. These sub- tasks aim to be useful for explainable Artificial In- telligence (AI) by including human-readable and machine-readable data. Given the didactic nature of these tasks, the out- put generated by these models should be as accurate as the most prestigious dictionaries. Hence, the pro- cess of selecting a quality dataset is a critical phase, as Garg et al. (2020) state: “a small number of data examples prevents an effective convergence to the task, while noisy data leads to incorrect conver- gence”. In this case, tasks require that the glosses used in the training represent the exact meaning of the word being defined in the context that the em- beddings were extracted. However, as per our un- derstanding, coherence, rigour and lexicographical prestige of the provided dataset should be improved; although accessing a prestigious dictionary is not an easy task. A Reverse Dictionary takes a description in nat- ural language and generates a list of words satis- fying it (Siddique and Sufyan Beg, 2018). First RD were Information Retrieval systems for Turkish (El-Kahlout and Oflazer, 2004) and Japanese (Bilac et al., 2004). Other approaches used lexical graphs (Thorat and Choudhari, 2016; Ortega-Mart´ın, 2021) that capture the relationships between the words of the definition itself and between these and others similar to them at different levels (synonymy, hyper- onymy, etc.). Other systems create a vector space from these lexical resources, such as Wordnet (Du- toit and Nugues, 2002; Calvo et al., 2016; M´endez 68 ID: Definition es.train.212: “Biolog´ıa.— Se dice de los microorganismos que no aceptan los colorantes habituales.” es.train.250: “Zoolog´ıa.— Cualquiera de …

Content truncated to 500 words