Amsqr at SemEval-2022 Task 4: Towards AutoNLP via Meta-Learning and Adversarial Data Augmentation for PCL Detection

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 485 - 489 July 14-15, 2022 ©2022 Association for Computational Linguistics Amsqr at SemEval-2022 Task 4: Towards AutoNLP via Meta-Learning and Adversarial Data Augmentation for PCL Detection Alejandro Mosquera Broadcom Corporation / 1320 Ridder Park Drive San Jose, 95131 California, USA alejandro.mosquera@broadcom.com Abstract This paper describes the use of AutoNLP tech- niques applied to the detection of patronizing and condescending language (PCL) in a binary classification scenario. The proposed approach combines meta-learning, in order to identify the best performing combination of deep learning architectures, with the synthesis of adversar- ial training examples; thus boosting robustness and model generalization. A submission from this system was evaluated as part of the first sub- task of SemEval 2022 - Task 4 and achieved an F1 score of 0.57%, which is 16 percentage points higher than the RoBERTa baseline pro- vided by the organizers. 1 Introduction The harmful use of language in social media can have negative and long-lasting effects such as ex- clusion and unfair treatment, specially when tar- geting vulnerable communities. For this reason, the detection of toxic, hateful and abusive com- ments has been the central topic of several work- shops and tool evaluations, drawing a lot of atten- tion from the Natural Language Processing (NLP) research community in the last years. However, while toxic language has a clear intent and is usu- ally obvious to the reader, patronizing and conde- scending language (PCL) is more subtle and likely used in a subconscious manner even in traditional media (Perez Almendros et al., 2020). The afore- mentioned characteristics and its subjective nature makes PCL harder to identify than abusive com- ments by both humans (Sap et al., 2019) and NLP applications. The continuously increasing taxonomies of lan- guage misuse poses new challenges to social media platforms, thus not only requiring more effort and cost in order to identify abuse across different lan- guages and textual genres but also having to keep a balance between aggressive and conservative fil- tering strategies. On the one hand, users eventually devise ways of evading automatic content mod- eration (Gerrard, 2018), while on the other hand, policing that restricts freedom of speech can lead to distrust (Kirk and Schill, 2021). For these reasons, content filters usually rely on the latest advances in NLP research, dominated in the recent years by deep learning architectures. Despite the com- petitive scores achieved via transfer learning and models such as the Transformer (Vaswani et al., 2017) in this area, choosing and optimizing the right modeling framework for a given NLP task is still a non-trivial problem. Automated Natural Language Processing (Au- toNLP), the equivalent of Automated Machine Learning (AutoML) for NLP, is a relatively new field of study that aims to automate the iterative components of developing a NLP model given a specific input data and task without requiring any special domain expertise. By building upon exist- ing concepts such as transfer learning, data augmen- tation and meta-learning the author …

Content truncated to 500 words