Dialogue Structure Parsing on Multi-Floor Dialogue Based on Multi-Task Learning

End-to-end dialogue structure parsing on multi-floor dialogue based on multi-task learning

“…Utterances in a multi-floor dialogue can be classified into three types: starting a new TU (Start), and continuing the currently open TU (Continue), resuming another TU that is already open (Other). Table 7 shows the prediction performance of the antecedent prediction corresponding to such three statuses of TUs ( Kawano et al., 2021 ). The results showed that the model suffered in the case of predicting the utterance of another TU already open as the antecedent compared to other cases.…”

“…Thus, parsing of multi-floor dialogues is inherently different from typical multiparty dialogue parsing tasks that do not take such aspects into account (e.g., conversation disentanglement task). In this study, we generalize and extend previous works ( Zhang et al., 2017 ; Shi and Huang, 2019 ; Kawano et al., 2021 ) to fit the problem of multi-floor dialogue structure parsing. Additionally, we propose an auxiliary objective function to enhance the consistency of the predicted multi-floor dialogue structure.…”

“…One solution for resolving an entangled conversation is to formulate it as a topic-tracking task, such that each message in an entangled conversation determines whether it starts a new conversation or belongs to an existing conversation ( Wang et al., 2020 ). Another solution is to identify the referential and non-referential relationships within two utterances that determine whether an utterance is a response to another utterance ( Kawano et al., 2021 ). Supervised and unsupervised conversation disentanglement methods have been proposed for both of these approaches, and in particular, the development of neural conversation disentanglement models based on the powerful performance of deep neural networks is noteworthy ( Kummerfeld et al., 2019 ; Yu and Joty, 2020 ; Liu et al., 2021 ).…”

“…This paper is an extended version of our conference paper that was accepted at ROBOTDIAL 2020 (Kawano et al, 2021)…”

“…We can view parsing the dialogue structure on multi-floor dialogue as a problem that extracts directed graphs corresponding to a span of transaction units within the dialogue. However, there is still a lack of previous work on automatic dialogue structure parsing for multi-floor dialogue ( Kawano et al., 2021 ).…”