NL4IA at SemEval-2023 Task 3: A Comparison of Sequence Classification and Token Classification to Detect Persuasive Techniques

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 794–799 July 13-14, 2023 ©2023 Association for Computational Linguistics NL4IA at SemEval-2023 Task 3: A Comparison of Sequence Classification and Token Classification to Detect Persuasive Techniques Albert Pritzkau Fraunhofer Institute for Communication, Information Processing and Ergonomics Wachtberg, Germany albert.pritzkau@fkie.fraunhofer.de Abstract The following system description presents our approach to the detection of persuasion tech- niques in online news. The given task has been framed as a multi-label classification prob- lem. In a multi-label classification problem, each input chunk—in this case paragraph—is assigned one of several class labels. Span level annotations were also provided. In or- der to assign class labels to the given docu- ments, we opted for RoBERTa (A Robustly Optimized BERT Pretraining Approach) for both approaches—sequence and token classi- fication. Starting off with a pre-trained model for language representation, we fine-tuned this model on the given classification task with the provided annotated data in supervised training steps. 1 Introduction Political rhetoric, propaganda, and advertising are all examples of persuasive discourse. As defined by Lakoff (1982), persuasive discourse is the non- reciprocal "attempt or intention of one party to change the behavior, feelings, intentions, or view- point of another by communicative means". Thus, in addition to the purely content-related features of communication, the discursive context of utter- ances plays a central role. SemEval-2023 Task 3 (Piskorski et al., 2023) considers persuasion as a communication phenomenon. With this approach, it is assumed that communication depends not only on the meaning of words in an utterance, but also on what speakers intend to communicate with a particular utterance. This concept is from the lin- guistic subfield of pragmatics. It is not always pos- sible to derive the function of an utterance from its form and additional contextual information is often needed. Recent research like (Tenney et al., 2019) (Jawahar et al., 2019) indicates the possibility that transformer-based networks capture structural in- formation about language ranging from syntactic up to semantic features. Beyond these features, these architectures remain almost entirely unex- plored. This task poses an attempt to explore the limits of the prevailing approach, in particular, in- vestigating Transformers ability to capture prag- matic features. 2 Background The central focus of this assignment is manipu- lative persuasion. The related task is to identify and evaluate propaganda and persuasion as found in social media. To identify and characterize ma- nipulative persuasion, the context can be stretched arbitrarily far across aspects of epistemology, logic, intent estimates, psychological biases, knowledge of pre-existing narratives, and even physical con- text. However, to potentially solve this problem in an automated fashion, the prevailing method is to frame the given task as a classification prob- lem. The different propaganda methods are under- stood as distinguishing criteria. Documents or even subsections of documents are annotated based on these features and thus form the input for training machine learning models, which should then be able to automatically recognize and classify corre- sponding sections. The undeniable success of this approach …

Content truncated to 500 words