Understanding and explaining affective traits in english and code-mixed conversations

Abstract

In the past decade, Natural Language Processing has undergone a transformative journey, marked by profound changes. The realm of conversational discourse, in particular, has witnessed remarkable advancements, with contemporary systems exhibiting significant potential. The ubiquitous integration of conversational agents into our daily lives often obscures the intricate computations underpinning their functionality. Yet, instances of non-empathetic responses or a failure to grasp nuances like humour or sarcasm serve as stark reminders that our interactions extend to the realm of machines. Addressing this limitation forms the core of our research, which revolves around refining a specific facet of conversational understanding – the nuanced focus on affects. Affects in conversation encapsulate a myriad of discourse attributes, including emotions, sarcasm, humour, and speaker profiles, all playing a pivotal role in comprehending the comprehensive meaning inherent in a spoken statement. Our dedicated efforts unfold in the unraveling of these intricate characteristics, aimed at enhancing the interpretative capabilities of dialogue agents. Moreover, we posit that the mere identification of these affective cues inadequately captures the profound essence embedded within the uttered statement. Consequently, our inquiry extends beyond identification to elucidate these affective dimensions, fostering a more profound understanding of conversational discourse. Throughout this thesis, we address multiple novel problem statements, curate innovative datasets, and develop cutting-edge methods tailored to solve each task. Specifically, our focus encompasses the tasks of emotion recognition in conversation, emotion flip reasoning, humour identification, sarcasm detection, sarcasm explanation, and speaker profiling. This thesis, therefore, seeks to establish a foundation for dedicated research in the domain of affects in conversation.