Sefamerve ARGE at SemEval-2021 Task 5: Toxic Spans Detection Using Segmentation Based 1-D Convolutional Neural Network Model

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 909–912 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 909 Sefamerve ARGE at SemEval-2021 Task 5: Toxic Spans Detection Using Segmentation Based 1-D Convolutional Neural Network Model Selman Delil∗ Birol Kuyumcu∗ Sefamerve R&D Center Istanbul, Turkey {selman.delil, birol.kuyumcu, cuneyt.aksakalli}@sefamerve.com C¨uneyt Aksakallı ∗ Abstract This paper describes our contribution to SemEval-2021 Task 5: Toxic Spans Detec- tion. Our approach considers toxic spans de- tection as a segmentation problem. The sys- tem, Waw-unet, consists of a 1-D convolu- tional neural network adopted from U-Net ar- chitecture commonly applied for semantic seg- mentation. We customize existing architec- ture by adding a special network block con- sidering for text segmentation, as an essential component of the model. We compared the model with two transformers-based systems RoBERTa and XLM-RoBERTa to see its per- formance against pre-trained language models. We obtained 0.6251 f1 score with Waw-unet while 0.6390 and 0.6601 with the compared models respectively. 1 Introduction Unlike the text classification problems targeting to classify whole documents (Borkan et al., 2019; Schmidt and Wiegand, 2017; Pavlopoulos et al., 2019), toxic span detection is an NLP task focus- ing on capturing granular contents that make a text toxic. Proposed solutions may contribute to man- aging semi-automated moderations such as online discussions or news portals that are open to large participation and user comments. Therefore, the evaluation of systems that could accurately locate toxic spans within a text is considered a crucial step for this task (Pavlopoulos et al., 2021). We adapt two solution approaches for the task. For the first approach, we consider toxic spans de- tection as a segmentation problem while in the second one we use transformers-based models. Our proposed model for the first approach uses character-based tokenized chunks as an input and outputs segmented text. The system uses a 1- dimensional (1-D) convolutional neural network adopted from U-Net architecture (Ronneberger et al., 2015) commonly applied for semantic seg- mentation. We previously studied this approach, as we call Waw-unet, on text parsing problems for unstructured postal addresses, and achieved re- markable results (Delil et al., 2020). In our second approach, we consider toxic spans as a single label Named-Entity Recognition prob- lem. We employ several different transformers- based models and obtained better scores with RoBERTa (Liu et al., 2019) and XLM-RoBERTa (Conneau et al., 2020) network architectures. To see the difference between our main system and transformers-based models, we compared model achievements, and obtained 0.6251 f1 score with Waw-unet while 0.6390 and 0.6601 with the other models respectively. Following the final rankings, our best score ranked 44th among 91 submissions1. 2 Waw-unet Architecture Waw-unet is a fully convolutional neural network architecture we designed by taking inspiration from U-Net architecture which was firstly developed for segmentation problems (Ronneberger et al., 2015). The U-Net network architecture is composed of two symmetric parts, which uses dimension reduc- tion for the first half of the network, and then in- creases its dimension in the second …

Content truncated to 500 words