Enhancing Cognitive Models of Emotions with Representation Learning

https://aclanthology.org/2021.cmcl-1.18/

Transcript

1. Enhancing Cognitive Models of Emotions with Representation Learning Yuting Guo , Jinho Choi Department of Computer Science, Emory University In this work, we investigated how to learn computational representations for human emotions by pre-trained deep learning models. Our contributions are: • Develop a deep probing model that allows us to interpret the process of representing learning on emotion classification. • Achieve the SOTA result on the Empathetic Dialogue dataset for the classification of 32 emotions. • Generate emotion representations that can derive an emotion graph and an emotion wheel. INTRODUCTION AND METHODOLOGY LAYER-WISE ANALYSIS We trained a logistic regression model per layer on the concatenated hidden states of each head to predict emotions. By analyzing the classification results for each layer, we derived an emotion graph shown in Figure 2. Figure 3: Generated emotion wheel. GENERATION OF EMOTION WHEEL We generated emotion embeddings and derived an emotion wheel by representing each complex emotion as a weighted sum of two basic emotions shown in Figure 3. AUGMENTATION OF PAD MODEL We applied the emotion embeddings to augment the PAD model visualized in Figure 4. prepared apprehensive annoyed terrified caring anxious ashamed devastated proud grateful hopeful confident faithful furious impressed lonely sentimental disappointed anticipating angry afraid trusting disgusted surprised joyful sad excited content embarrassed prepared caring grateful surprised impressed proud hopeful confident anticipating excited sad devastated sentimental nostalgic lonely ashamed jealous disappointed embarrassed guilty terrified afraid anxious apprehensive faithful trusting disgusted annoyed furious angry E N C O D E R PH11 PH1k PHℓ1 ⋯ ⋯ ⋯ ⋯ e 0 e 1k e 11 ⋯ ⋯ PHℓk e ℓ1 e ℓk ⊕ ⊕ ⋯ N O R M L
 I N E A R o w 1 w 2 w n ⋯ Figure 4: The 2D plot from the PAD values of 32 emotions predicted by our regression models. Emotions in red are predicted emotions. Figure 2: The emotion graph shows Most emotion pairs point from coarse- grained emotions to fine-grained emotions. Figure 1: Model Architecture.