Automatic Sexism Detection with Multilingual Transformer Models

• DOI: 10.48550/arxiv.2106.04908
• Publication Date: 2021-01-01
• Authors: Mina Schütz; Jaqueline Boeck; Daria Liakhovets; Djordje Slijepčević; Armin Kirchknopf; Manuel Hecht; Johannes Bogenspe...; Sven Schlarb; Alexander Schindler; Matthias Zeppelzauer
• Keywords: Binary classification; Benchmark (surveying); Macro; F1 score

Sexism has become an increasingly major problem on social networks during the last years. The first shared task sEXism Identification in Social neTworks (EXIST) at IberLEF 2021 is international competition field of Natural Language Processing (NLP) with aim to automatically identify sexism media content by applying machine learning methods. Thereby detection formulated as a coarse (binary) classification and fine-grained that distinguishes multiple types sexist (e.g., dominance, stereotyping, objectification). This paper presents contribution AIT_FHSTP team EXIST2021 benchmark for both tasks. To solve tasks we applied two multilingual transformer models, one based BERT XLM-R. Our approach uses different strategies adapt transformers content: first, unsupervised pre-training additional data second, supervised fine-tuning augmented data. For our best model XLM-R EXIST datasets provided dataset. run binary (task 1) achieves macro F1-score 0.7752 scores 5th rank benchmark; multiclass 2) submission 6th 0.5589.