An Empirical Study on Multiple Knowledge from ChatGPT for Emotion Recognition in Conversations
Geng Tu, Bin Ming Liang, Bing Qin, Kam‐Fai Wong, Ruifeng Xu · 2023
Multiple knowledge (e.g., co-reference, topics, emotional causes, etc) has been demonstrated effective for emotion detection.However, exploring this knowledge in Emotion Recognition in Conversations (ERC) is currently a blank slate due to the lack of annotated data and the high cost involved in obtaining such knowledge.Fortunately, the emergence of Large Language Models (LLMs) holds promise in filling this void.Therefore, we propose a Multiple Knowledge Fusion Model (MKFM) to effectively integrate such knowledge generated by LLMs for ERC and empirically study its impact on the model.Experimental results on three public datasets have demonstrated the effectiveness of multiple knowledge for ERC.Furthermore, we conduct a detailed analysis of the contribution and complementarity of this knowledge 1 .* Corresponding authors. 1 The code is available at https://github.com/ TuGengs/MKFM.(#1) Speaker 0: Ok, I know this is gonna sound really stupid, but I feel that if I can do this, you know, if I can actually do my own laundry, there isn't anything I can't do.(Scared) (#2) Speaker 1: That does not sound stupid to me.You know, it's like the first time I had to make dinner for myself, after Carol left me?I'm sorry, that's all the time we have.Next on Ross... Uh-oh.(Peaceful) (#3) Speaker 0: What uh-oh?(Peaceful) (#4) Speaker 1: Uh-oh, uh-oh, the laundry's done.It's, uh, it's a song.The laundry song that we sing.Uh-oh the laundry's done, uh-oh, uh-oh.(Peaceful) (#5) Speaker 0: Ross, what's the matter?(Peaceful) (#6) Speaker 1: Nothing, nothing.Lee-lo, the laundry's done.(Mad) (#7) Speaker 0: Come on, show me.(Peaceful) (#8) Speaker 1: All right, all right, it's just that you left a red sock in with all your whites, and now, everything's kinda pink.(Sad) (#9) Speaker 0: Oh, everything's pink.(Sad) (#10) Speaker 1: Yeah, uh, except for the red sock, which is still red.I'm sorry, please don't be upset, it could happen to anyone.(Sad) (#11) Speaker 0: Except it didn't.It happened to me.Oh, god, I'm gonna look like a big marshmallow peep.What am I doing?What am I doing?My father's right.I can't live on my own!I can't even do laundry!(Sad) Knowledge 1: ALK Knowledge 2: AUK Knowledge 3: ACK TP: [0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2] SC: [1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1] MP: [1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1] EC: As shown in (#10, and #11) CS: As shown in (#10, and #11) ACS: As shown in (#10, and #11)EC of #10: Apologizes and reassures Speaker 0 that it's a common mistake and can happen to anyone.EC of #11: Expresses self-doubt and anxiety, feeling incapable of living on her own or even doing laundry.CS of #10: Speaker 1 tries to console Speaker 0, acknowledging the mistake and reassuring them that it could happen to anyone.The emotion behind this statement is empathy and reassurance.CS of #11: Speaker 0 becomes overwhelmed with self-doubt, questioning their ability to live independently and even do simple tasks like laundry.The emotion behind this statement is frustration and self-deprecation.ACS of #10: Attempt to console and normalize -Ross tries to console Speaker 0 by saying that such mishaps can happen to anyone, implying it is a common mistake.ACS of #11: Self-criticism and doubt -Speaker 0 expresses self-doubt and criticizes their ability to live independently based on the laundry mishap.EC2 of #10 and #11: [9], [10] CT of #10 and #11: [1, 4, 8, 9], [1, 9, 10] CR of #10 and #11: [8], [9] Intra-and Inter-speaker Intra-and Inter-CT Classifier Utterance-level Encoder (AUK) Graph Context Encoder (ACK) Contrastive Learning (ALK)RoBerta Encoder A A A C 0 X i c j V H L S s N A F D 2 N r 1 p f V Z d u g q 3 g q i Q F H 8 u C m y 4 r 2 g e 0 t S T p t A 3 N i 8 l E K K U g b v 0 B t / p T 4 h / o X 3 h n T E E t o h O S n D n 3 n j N z 7 7 U j z 4 2 F Y b x m t K X l l d W 1 7 H p u Y 3 N r e y e / u 9 e I w 4 Q 7 r O 6 E X s h b t h U z z w 1 Y X b j C Y 6 2 I M 8 u 3 P d a 0 x x c y 3 r x l P H b D 4 F p