FramingFreaks at SemEval-2023 Task 3: Detecting the Category and the Framing of Texts as Subword Units with Traditional Machine Learning

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 922–926 July 13-14, 2023 ©2023 Association for Computational Linguistics FramingFreaks at SemEval-2023 Task 3: Detecting the Category and the Framing of Texts as Subword Units with Traditional Machine Learning Rosina Baumann University of Tübingen rosina.baumann@student.uni-tuebingen.de Sabrina Deisenhofer University of Tübingen sabrina.deisenhofer@student.uni-tuebingen.de Abstract This paper describes our participation as team FramingFreaks in the SemEval-2023 task 3 “Category and Framing Predictions in online news in a multi-lingual setup.” We partici- pated in subtasks 1 and 2. Our approach was to classify texts by splitting them into subwords to reduce the feature set size and then using these tokens as input in Support Vector Ma- chine (SVM) or logistic regression classifiers. Our results are similar to the baseline results. 1 Introduction Fake news and false arguments are everywhere. In Twitter posts or news articles, people use persua- sion techniques to influence the opinion of other people. For this reason, it is important to detect the framing and persuasion techniques in texts automat- ically. This was the aim of the third shared task of SemEval 2023 (Piskorski et al., 2023), in which we took part as team FramingFreaks. This task focuses on the detection of non-neutral texts (first subtask), evaluating the frame of the argumentation on the ar- ticle level (second subtask), and finding out which techniques were used on the sentence level (third subtask). The data consists of online news and Twitter posts in English, French, German, Italian, Polish, and Russian. For the test set the organizers of the task proposed a few-shot learning task where our systems had to generalize to sources of other languages, namely Spanish, Greek, and Georgian. We participated in the first and second subtasks in all languages. We did not tackle the third subtask. The focus of our contribution is on computationally simple and efficient models. We use ‘traditional’ machine learning models, like SVM. However, un- like word or character-based features that are com- monly used with these models, we use subword units to further reduce the feature set size. Besides their computational efficiency, the tra- ditional classifiers based on bag-of-words (or n- grams) features have several other interesting prop- erties. For example, unlike recent deep-learning models, they can process sequences of arbitrary length, the computational cost comes mainly comes from the size of the feature set. Convex optimiza- tion methods used in (some of) these models make training and tuning these models more straightfor- ward. Furthermore, in some cases, they may work as well or even better than deep learning models, especially when information from pre-training is not available (see, for example, Çöltekin and Rama, 2018; Piskorski and Jacquet, 2020). However, the bag-of-words features may typi- cally result in sparse features. This may particularly be bad for morphologically complex languages as many words will be observed only a few times, and the relation between the morphologically-related words cannot be used by the classifier. A common alternative is to use character n-grams (e.g., Ifrim et al., …

Content truncated to 500 words