@inproceedings{,
    author = {Pritzkau, Albert; Waldmüller, Julia; Blanc, Olivier; Geierhos, Michaela; Schade, Ulrich},
    title = {Current language models' poor performance on pragmatic aspects of natural language},
    editor = {Ghosh, Kripabandhu; Mandl, Thomas; Majumder, Prasenjit; Mitra, Mandar},
    booktitle = {Working Notes of FIRE 2023 - Forum for Information Retrieval Evaluation (FIRE-WN 2023)},
    series = {CEUR-WS.org},
    journal = {},
    address = {},
    publisher = {},
    edition = {},
    year = {2024},
    isbn = {},
    volume = {3681},
    number =  {},
    pages = {159-169},
    url = {https://ceur-ws.org/Vol-3681/T3-2.pdf},
    doi = {},
    keywords = {Pragmatics ; Information Extraction ; Text Classification ; RoBERTa},
    abstract = {With the following system description, we present our approach for claim detection in tweets. We address both Subtask A, a binary sequence classification task, and Subtask B, a token classification task. For the first of the two subtasks, each input chunk - in this case, each tweet - was given a class label. For the second subtask, a label was assigned to each individual token in an input sequence. In order to match each utterance with the appropriate class label, we used pre-trained RoBERTa (A Robustly Optimized BERT Pretraining Approach) language models for sequence classification. Using the provided data and annotations as training data, we fine-tuned a model for each of the two classification tasks. Though the resulting models serve as adequate baseline models, the exploratory data analysis suggests fundamental problems in the structure of the training data. We argue that such tasks cannot be fully solved if pragmatic aspects of language are ignored. This type of information, often contextual and thus not explicitly stated in written language, is insufficiently represented in the current models. For this reason, we posit that the provided training data is under-specified and imperfectly suited to these classification tasks.},
    note = {},
    institution = {Universität der Bundeswehr München, Fakultät für Informatik, INF 7 - Institut für Datensicherheit, Professur: Geierhos, Michaela},
    
    
}