Abstract
Despite impressive progress in Natural Language Processing (NLP) applications, Natural Language Understanding (NLU), i.e., extracting semantic and discourse representations from text, remains an elusive task, leading to errors. One of the issues current NLP methods have not been able to solve yet involves implicit language use. When dealing with implicitness, i.e., conveying meaning without explicit expression, in which the intention of the speaker is deduced from indirect cues, Machine Learning (ML) models used in NLP are required to look beyond superficial textual patterns and rely on more profound world knowledge and reasoning abilities. To tackle such complex tasks, researchers have invested in incorporating common sense (CS) in ML models. Incorporating CS, while being a challenging task, is essential to interpret implicit language. However, there is no consensus in existing literature to what extent this knowledge is present in large language models. Therefore, our main goal is to evaluate and improve the incorporation of CS in these models. As a practical application of this framework, we focus on sarcasm detection, which currently suffers from lack of CS reasoning abilities. Moreover, since sarcasm also has implications for hate speech detection, we will explore using CS for sarcasm in hate speech, which has not been researched yet. Lastly, we will extrapolate our method to Dutch, to verify that our approach of CS reasoning can be generalized over languages.
Researcher(s)
Research team(s)
Project type(s)