UTSA NLP at SemEval-2022 Task 4: An Exploration of Simple Ensembles of Transformers, Convolutional, and Recurrent Neural Networks

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 379 - 386 July 14-15, 2022 ©2022 Association for Computational Linguistics UTSA NLP at SemEval-2022 Task 4: An Exploration of Simple Ensembles of Transformers, Convolutional, and Recurrent Neural Networks Xingmeng Zhao and Anthony Rios Department of Information Systems and Cyber Security University of Texas at San Antonio San Antonio, TX {xingmeng.zhao, anthony.rios}@utsa.edu Abstract The act of appearing kind or helpful via the use of but having a feeling of superiority con- descending and patronizing language can have have serious mental health implications to those that experience it. Thus, detecting this conde- scending and patronizing language online can be useful for online moderation systems. Thus, in this manuscript, we describe the system de- veloped by Team UTSA SemEval-2022 Task 4, Detecting Patronizing and Condescending Language. Our approach explores the use of several deep learning architectures including RoBERTa, convolutions neural networks, and Bidirectional Long Short-Term Memory Net- works. Furthermore, we explore simple and effective methods to create ensembles of neu- ral network models. Overall, we experimented with several ensemble models and found that the a simple combination of five RoBERTa models achieved an F-score of .6441 on the development dataset and .5745 on the final test dataset. Finally, we also performed a compre- hensive error analysis to better understand the limitations of the model and provide ideas for further research. 1 Introduction Patronizing and condescending language (PCL) generally appears as an act to hold a superior at- titude, resulting in language that “talks down” to others. For instance, PCL may describe someone in a power position as the potential “savior” of a vul- nerable community (e.g., ”A donation of one dollar can save a life”), masquerading a sense of superior- ity as compassion. There has been recent research suggesting that PCL can have adverse effects on the mental health of individuals (Giles et al., 1993; Shaw and Gordon, 2021), particularly in the con- text of ageism. While there has been substantial research on PCL in various contexts (Huckin, 2002; Komrad, 1983; Giles et al., 1993; Shaw and Gor- don, 2021), unfortunately, there have been few ef- forts to develop PCL detectors in the field of Natu- ral Language Processing (NLP). Hence, this paper describes our team’s (UTSA NLP) contributions on the SemEval-2022 Task 4 (Pérez-Almendros et al., 2022) that introduced a new dataset for detecting PCL language. NLP has investigated a broad spectrum of prob- lematic language usages, such as hate speech (Vid- gen et al., 2021), sarcasm language (Bamman and Smith, 2015), fake news (Hu et al., 2021), and the spread of rumors and disinformation. However, PCL has only recently been explored in the NLP community (?). To alleviate this issue, SemEval- 2022 Task 4 expanded on the work by ?, releasing a large PCL dataset for two PCL subtasks. Sub- task 1 focuses on detecting the presence of PCL in news stories. PCL detection consists of a variety of sub-problems, for instance, identifying the exact PCL type expressed post …

Content truncated to 500 words