@inproceedings{christodoulou-2023-nlp-christine,
    title = "{NLP}{\_}{CHRISTINE}@{LT}-{EDI}-2023: {R}o{BERT}a {\&} {D}e{BERT}a Fine-tuning for Detecting Signs of Depression from Social Media Text",
    author = "Christodoulou, Christina",
    editor = "Chakravarthi, Bharathi R.  and
      Bharathi, B.  and
      Griffith, Joephine  and
      Bali, Kalika  and
      Buitelaar, Paul",
    booktitle = "Proceedings of the Third Workshop on Language Technology for Equality, Diversity and Inclusion",
    month = sep,
    year = "2023",
    address = "Varna, Bulgaria",
    publisher = "INCOMA Ltd., Shoumen, Bulgaria",
    url = "https://preview.aclanthology.org/fix-sig-urls/2023.ltedi-1.16/",
    pages = "109--116",
    abstract = "The paper describes the system for the 4th Shared task on ``Detecting Signs of Depression from Social Media Text'' at LT-EDI@RANLP 2023, which aimed to identify signs of depression on English social media texts. The solution comprised data cleaning and pre-processing, the use of additional data, a method to deal with data imbalance as well as fine-tuning of two transformer-based pre-trained language models, RoBERTa-Large and DeBERTa-V3-Large. Four model architectures were developed by leveraging different word embedding pooling methods, namely a RoBERTa-Large bidirectional GRU model using GRU pooling and three DeBERTa models using CLS pooling, mean pooling and max pooling, respectively. Although ensemble learning of DeBERTa{'}s pooling methods through majority voting was employed for better performance, the RoBERTa bidirectional GRU model managed to receive the 8th place out of 31 submissions with 0.42 Macro-F1 score."
}