For SemEval-2023 Task 5, we have submitted three DeBERTaV3[LARGE] models to tackle the first subtask, classifying spoiler types (passage, phrase, multi) of clickbait web articles. The choice of basic parameters like sequence length with BERT[BASE] uncased and further approaches were then tested with DeBERTaV3[BASE] only moving the most promising ones to DeBERTaV3[LARGE]. Our research showed that information-placement on webpages is often optimized regarding e.g. ad-placement Those informations are usually described within the webpages markup which is why we conducted an approach that takes this into account. Overall we could not manage to beat the baseline, which we lead down to three reasons: First we only crawled markup for Huffington Post articles, extracting only p- and a-tags which will not cover enough aspects of a webpages design. Second Huffington Post articles are overrepresented in the given dataset, which, third, shows an imbalance towards the spoiler tags. We highly suggest re-annotating the given dataset to use markup-optimized models like MarkupLM or TIE and to clear it from embedded articles like “Yahoo” or archives like “archive.is” or “web.archive” to avoid noise. Also, the imbalance should be tackled by adding articles from sources other than Huffington Post, considering that also multi-tagged entries should be balanced towards passage- and phrase-tagged ones.

This paper describes our approach (ur-iw-hnt) for the Shared Task of GermEval2021 to identify toxic, engaging, and fact-claiming comments. We submitted three runs using an ensembling strategy by majority (hard) voting with multiple different BERT models of three different types: German-based, Twitter-based, and multilingual models. All ensemble models outperform single models, while BERTweet is the winner of all individual models in every subtask. Twitter-based models perform better than GermanBERT models, and multilingual models perform worse but by a small margin.