GenPADS: Reinforcing politeness in an end-to-end dialogue system

• DOI: 10.1371/journal.pone.0278323
• Publication Date: 2023-01-06T18:32:41Z
• Authors: Kshitij Mishra; Mauajama Firdaus; Asif Ekbal
• Keywords: End-to-end principle

In a task-oriented dialogue setting, user’s mood and demands can change in an ongoing dialogue, which may lead to non-informative conversation or result drop-off. To rectify such scenarios, conversational agent should be able learn the behaviour online, form informative, empathetic interactive responses. incorporate these three aspects, we propose novel end-to-end system GenPADS . First, build train two models, viz. politeness classifier extract polite information present agent’s utterances generation model (G) generate varying but semantically correct We then both of models reinforcement learning (RL) setting using different oriented reward algorithms adapt our classifier, annotate recently released Taskmaster dataset into four fine-grained classes depicting impoliteness. Further, generator model, prepare GenDD same dataset. Lastly, perform automatic human evaluation by building seven user simulators. Detailed analysis reveals that performs better than considered baselines, transformer based seq2seq for utterance retrieval adaptive (PADS).