notesum.ai

Published at November 26

Push the Limit of Multi-modal Emotion Recognition by Prompting LLMs with Receptive-Field-Aware Attention Weighting

cs.CL

Released Date: November 26, 2024

Authors: Liyun Zhang¹, Dian Ding¹, Yu Lu¹, Yi-Chao Chen¹, Guangtao Xue¹

Aff.: ¹Shanghai Jiao Tong University

Arxiv: http://arxiv.org/abs/2411.17674v1

Refer to caption

Models	Acc. (%)	w-F1 (%)	Happy	Sad	Neutral	Angry	Excited	Frustrated
DialogueCRN	65.31	65.34	51.49	74.54	62.38	67.25	73.96	59.97
COGMEN	67.60	68.20	55.76	80.17	63.21	61.69	74.91	63.90
DialogueLLM	70.62	69.93	-	-	-	-	-	-
CORECT	69.93	70.02	59.30	80.53	66.94	69.59	72.69	68.50
SDT	73.95	74.08	72.71	79.51	76.33	71.88	76.79	67.14
CORECT (Multi-task)	70.18	70.28	57.64	86.11	68.00	64.42	79.11	64.43
CORECT + GPT-4	71.16	70.90	70.10	86.22	64.76	65.02	76.28	68.12
SDT (Multi-task)	73.94	74.14	60.56	86.61	71.65	65.45	85.56	71.07
SDT + GPT-4	74.43	74.63	61.11	86.55	76.49	65.43	85.02	68.37
SDT + Llama-405B	74.73	74.79	63.41	81.20	78.78	64.92	82.39	70.51