Fig 1.
An example of generative politeness adaptive dialogue system.
Light orange (impolite or noisy) and light red (partial information) utterance boxes depict the user’s dissatisfaction with the ongoing dialogue. GenPADS adapts the dialogue system towards generating polite and diverse responses as per user’s and agent’s politeness feedback.
Fig 2.
Distribution of different politeness classes in the Taskmaster dataset.
Table 1.
PADD show the number of instances in the gold-standard polite annotated data.
Cleaned shows the statistics of cleaned Taskmaster data; DG-Dataset and G-Dataset shows the statistics of dialogue dataset used to train the two generation models.
Table 2.
Statistics of user simulator for each of the seven domains.
Table 3.
Hyperparameters information.
Table 4.
GenPADS polite classifier (PC), generation module (G) evaluation results.
Table 5.
Automatic evaluation results of GenPADS, RetrievalPADS and Dialogue Generator for all the domains with Baseline (BL) and proposed PRRP reward algorithms.
Fig 3.
Average politeness factor with different reward algorithms for (a) flights (b) food-ordering (c) hotels (d) movies and (e) music (f) restaurant-search (g) sports domain.
Fig 4.
Average convergent success rate with different reward algorithms for (a) flights (b) food-ordering (c) hotels (d) movies and (e) music (f) restaurant-search (g) sports domain.
Table 6.
Results of manual evaluation.
Fig 5.
A conversation example of flight search using GenPADS, RetrievalPADS and Dialogue Generator.