UnedMediaBiasTeam @ SemEval-2023 Task 3: Can We Detect Persuasive Techniques Transferring Knowledge From Media Bias Detection?

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 83–87 July 13-14, 2023 ©2023 Association for Computational Linguistics MarsEclipse at SemEval-2023 Task 3: Multi-Lingual and Multi-Label Framing Detection with Contrastive Learning Qisheng Liao MBZUAI Qisheng.Liao@mbzuai.ac.ae Meiting Lai Liaoning University Lai.Meiting@outlook.com Preslav Nakov MBZUAI Preslav.Nakov@mbzuai.ac.ae Abstract This paper describes our system for SemEval- 2023 Task 3 Subtask 2 on Framing Detection. We used a multi-label contrastive loss for fine- tuning large pre-trained language models in a multi-lingual setting, achieving very compet- itive results: our system was ranked first on the official test set and on the official shared task leaderboard for five of the six languages for which we had training data and for which we could perform fine-tuning. Here, we de- scribe our experimental setup, as well as vari- ous ablation studies. The code of our system is available at https://github.com/QishengL/ SemEval2023 1 Introduction Framing involves highlighting certain elements of the perceived reality over others, which can promote problem definition, causal interpretation, moral evaluation, and/or treatment recommenda- tion (Entman, 1993). For example, events such as the outbreak of the COVID-19 pandemic can be discussed from different perspectives: health, eco- nomic, political, legal, quality of life, etc. Recently, machine learning has led to significant progress in framing detection. Previously, Naïve Bayes, Sup- port Vector Machines, and deep learning models based on recurrent neural networks (RNNs), long short-term memory networks (LSTM), and trans- formers have been used to analyze large amounts of text data and to capture complex patterns of framing (Morstatter et al., 2018; Liu et al., 2019a). Here, we describe our system for SemEval-2023 Task 3 (Piskorski et al., 2023) Subtask 2 on Fram- ing Detection, which is defined as a multi-label text classification task at the article level. The par- ticipants were required to build a system that can identify the frames used in an input article. Below, we describe our approach, which combines pre- trained language models (PLMs) and contrastive learning. Our contributions are as follows: • We are the first to use contrastive learning in a multi-label framing detection setting. • We achieve very strong results: our system is ranked first for five out of the six languages for which training data was available, and it was third for English, which shows the effective- ness of contrastive learning for multi-lingual multi-label framing detection. • We perform ablation experiments to study the impact of different elements of our system. 2 Related Work Contrastive learning has been used in computer vi- sion for many years. It works as follows: a random example is picked as an anchor, and then positive and negative examples are selected with respect to that anchor. Next, the positive examples are pulled towards the anchor, while the negative ones are pushed away from it. SimCLR is a popular contrastive learning model (Chen et al., 2020), introduced in computer vision. It uses two views of the same image, generated using augmentations, and these are considered as positive examples. All other images are negative …

Content truncated to 500 words