MarSan at SemEval-2023 Task 10: Can Adversarial Training with help of a Graph Convolutional Network Detect Explainable Sexism?
24831Authors
nan
Abstract
AbstractThis paper describes SemEval-2022’s shared task “Explainable Detection of Online Sexism”. The fine-grained classification of sexist content plays a major role in building explainable frameworks for online sexism detection. We hypothesize that by encoding dependency information using Graph Convolutional Networks (GCNs) we may capture more stylistic information about sexist contents. Online sexism has the potential to cause significant harm to women who are the targets of such behavior. It not only creates unwelcoming and inaccessible spaces for women online but also perpetuates social asymmetries and injustices. We believed improving the robustness and generalization ability of neural networks during training will allow models to capture different belief distributions for sexism categories. So we proposed adversarial training with GCNs for explainable detection of online sexism. In the end, our proposed method achieved very competitive results in all subtasks and shows that adversarial training of GCNs is a promising method for the explainable detection of online sexism.
Keywords
Full Content
Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 101–107 July 13-14, 2023 ©2023 Association for Computational Linguistics ECNU MIV at SemEval-2023 Task 1: CTIM - Contrastive Text-Image Model for Multilingual Visual Word Sense Disambiguation Zhenghui Li 1, Qi Zhang 1, Xueyin Xia 2, YinXiang Ye 2, Qi Zhang2, and Cong Huang3 1School of Computer Science and Technology East China Normal University, China 2School of Communication and Electronic Engineering East China Normal University, China 3School Of Software JiangXi Agricultural University, China { zhli, xyxia, qzhang, yxye, qzhang}@stu.ecnu.edu.cn HCong@gmail.com Abstract This paper describes the performance of MIV team in SemEval-2023 Task-1-Visual Word Sense Disambiguation (Visual-WSD). This task is to give a potentially ambiguous word and some limited textual context and select among a set of candidate images the one which corresponds to the intended meaning of the tar- get word, which plays a critical role in human language understanding. Our team focuses on the multimodal domain of images and texts, we propose a model that can learn the matching relationship between text-image pairs by con- trastive learning. More specifically, We train the model from the labeled data provided by the official organizer, after pre-training, texts are used to reference learned visual concepts enabling visual word sense disambiguation tasks. In addition, the top results our teams got have been released showing the effective- ness of our solution. We release our code at https://github.com/Insaner1004/VWSD. 1 Introduction Word sense disambiguation (WSD) consists of as- sociating words in context with their most suitable entry in a predefined sense inventory, which is a historical and hot point in natural language pro- cessing and artificial intelligence. Supervised (Luo et al., 2018; Barba et al., 2021) and knowledge- based(Moro et al., 2014; Chaplot and Salakhut- dinov, 2018) are two common research methods that are widely used. However, visual word sense disambiguation aims to select among a set of can- didate images the one which corresponds to the intended meaning of the target ambiguous word, which is completely different from the previous work(Raganato et al., 2023). In the real world, in- formation often exists in different modalities, there- fore, the fusion of multimodal information plays an important role in the research of deep learning. This task is also useful for vision-and-language tasks (Lu et al., 2019) and text-to-image genera- Ambiguous word ... Limited textual Candidate images Gold image andromeda andromeda tree آندرومدا درخت آندرومدا andromeda albero di andromeda English Farsi Italian Languages Figure 1: An example of the visual word sense disam- biguation task. Given a potentially ambiguous word and some limited textual context, it is required to select a gold image form candidate images. tor(Wang et al., 2022). In this paper, we focus on the computing of similarity between text and image track. As shown in Figure 1, given the ambiguous word "andromeda" and some limited textual con- text "andromeda tree", select one of the best corre- sponding images from candidate pictures to match the semantics according to the similarity. Apart from that, it is …
Content truncated to 500 words
Labeled Text
AbstractThis paper describes SemEval-2022’s shared task “Explainable Detection of Online Sexism”. The fine-grained classification of sexist content plays a major role in building explainable frameworks for online sexism detection. We hypothesize that by encoding dependency information using Graph Convolutional Networks (GCNs) we may capture more stylistic information about sexist contents. Online sexism has the potential to cause significant harm to women who are the targets of such behavior. It not only creates unwelcoming and inaccessible spaces for women online but also perpetuates social asymmetries and injustices. We believed improving the robustness and generalization ability of neural networks during training will allow models to capture different belief distributions for sexism categories. So we proposed adversarial training with GCNs for explainable detection of online sexism. In the end, our proposed method achieved very competitive results in all subtasks and shows that adversarial training of GCNs is a promising method for the explainable detection of online sexism.
Publication Details
- Paper ID:
- 24831
- Year:
- 2023
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan