RoMa at SemEval-2021 Task 7: A Transformer-based Approach for Detecting and Rating Humor and Offense

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 297–305 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 297 RoMa at SemEval-2021 Task 7: A Transformer-based Approach for Detecting and Rating Humor and Offense Roberto Labadie Tamayo Universidad de Oriente Cuba rlabadiet@gmail.com Mariano J. Rodriguez Cisneros Universidad de Oriente Cuba mjasoncuba@gmail.com Reynier Ortega-Bueno PRHLT Research Center Universitat Polit`ecnica de Val`encia Valencia Spain rortega@prhlt.upv.es Paolo Rosso PRHLT Research Center Universitat Polit`ecnica de Val`encia Valencia Spain prosso@dsic.upv.es Abstract In this paper we describe the systems used by the RoMa team in the shared task on Detecting and Rating Humor and Offense (HaHackathon) at SemEval 2021. Our systems rely on data rep- resentations learned through fine-tuned neural language models. Particularly, we explore two distinct architectures. The first one is based on a Siamese Neural Network (SNN) combined with a graph-based clustering method. The SNN model is used for learning a latent space where instances of humor and non-humor can be distinguished. The clustering method is ap- plied to build prototypes of both classes which are used for training and classifying new mes- sages. The second one combines neural lan- guage model representations with a linear re- gression model which makes the final ratings. Our systems achieved the best results for hu- mor classification using model one, whereas for offensive and humor rating the second model obtained better performance. In the case of the controversial humor prediction, the most sig- nificant improvement was achieved by a fine- tuning of the neural language model. In general, the results achieved are encouraging and give us a starting point for further improvements. 1 Introduction Detecting humor has become a popular research field at the same time that the bad phenomenon of offensiveness spreading exaggeratedly grows in social media. In this scenario it is very frequent to find out alarming volumes of heterogeneous data such as textual messages, images, advertisements, etc., that harm some age groups, ethnicity, sexual gender or other demographic characteristics ( Betul Keles and Niall McCrae and Annmarie Grealish , 2020). Most of these harmful contents are often masquerade as innocent jokes or simply as a funny content. Therefore, it is crucial to shed light on the commonalities and differences between both phenomena in order to properly addressing the chal- lenge of computationally distinguishing humorous messages from aggressive or offensive ones. Recognizing humorous and offensive utterances on written messages is a very difficult task for hu- man beings and even more for computers (Waseem, 2016). These difficulties increase when the textual messages are isolated from the context in which they are produced. Additional knowledge from gestures, prosody features, visual content, situ- ational environment and sociocultural rules play an important role in how humans properly under- stand the real meaning behind funny and hateful contents. All this makes humor recognition and offensiveness detection challenging tasks within Natural Language Processing (NLP) and Human- Computer Interaction (HCI). On this line, the Task 7, HaHackathon: Detecting and Rating Humor and Offense at SemEval-2021 aims …

Content truncated to 500 words