Abstract: Emotional cause span extraction in conversations aims to extract causal spans that induce target emotion expression from conversational history,which plays a pivotal role in emotional conversation systems.However,causal spans extracted by exi-sting methods still have problems to be solved urgently,such as utterance position errors and boundary recognition errors.To this end,this paper proposes a multi-level semantic structure enhanced emotional cause span extraction method in conversations.The discourse-level coreferential structure is used to enhance the positioning of utterances where causal spans are located.The sentence-level syntactic structure is used to enhance the recognition of causal span boundaries.Firstly,according to preprocessed semantic structures and conversational content feature representations,the graph attention network is utilized to construct comprehensive graphs and model conversations at token level and utterance level,respectively.Meanwhile,the biaffine mechanism is utilized to promote interactions and integrations between two-level graphs,and structure-enhanced semantic comprehensive representations are obtained.Then,the linear layer is applied to extract causal spans.Experimental results on the two public datasets show that compared with the benchmark model,the F1 value and EM_pos value are improved by 2.42% and 2.26%,respectively.The proposed model also outperforms other baseline models in both F1_pos and EM_pos metrics,and can also be effectively compatible withutterance-level emotion cause entailment.

Key words: Natural language processing, Structure enhancement, Conversational emotion, Causal span extraction, Graph attention networks