Emotion Contextual Fusion Network: a Simple yet Versatile Approach for Emotion Recognition in Textual Conversations
Nicola DE LUCA, Daniela Gîfu, Diana Trandabăţ · Procedia Computer Science · 2024
Emotion detection in conversational settings holds significant importance across various domains, such as customer service, mental health support, and virtual assistants. In this paper, we introduce the Emotion Contextual Fusion Network (ECFN), a novel model architecture designed to discern emotional nuances within conversations, leveraging Language-Agnostic Sentence Representations (LASER) [1], alongside categorical metadata indicating speakers and numerical sentiment scores characterizing emotional content. Through a deliberate focus on relationship capture, ECFN employs attention mechanisms to gather information from both immediate and historical context. We perform tests on three different task-specific datasets using both paid and free resources. Experimental results on standard datasets show that our model managed to match and even surpass state of art models, all while being trainable within a reasonable timeframe.