notesum.ai

Published at November 25

Learning from Relevant Subgoals in Successful Dialogs using Iterative Training for Task-oriented Dialog Systems

cs.CL

cs.AI

Released Date: November 25, 2024

Authors: Magdalena Kaiser¹, Patrick Ernst², György Szarvas

Aff.: ¹Max Planck Institute for Informatics, Saarland Informatics Campus, Germany; ²Amazon, Berlin, Germany

Arxiv: http://arxiv.org/abs/2411.16305v1

Refer to caption

Model	BLEU	INFORM	SUCCESS	COMBINED	#samples
Mars Sun et al. (2023)	$19.90$	$88.9$	$78.0$	$103.4$	-
KRLS Yu et al. (2023)	$19.00$	$89.2$	$80.3$	$103.8$	-
DiactTOD Wu et al. (2023)	$17.50$	$89.5$	$84.2$	$104.4$	-
$\textsc{Suit}_{0}$ (initial SFT)	$\boldsymbol{19.94}$	$80.4$	$72.5$	$96.39$	-
+ all subgoals (SFT)	$19.50$	$87.0$	$79.4$	$102.70$	$31586$
+ all subgoals (DPO)	$17.79$	$86.9$	$80.6$	$101.54$	$31586$
$\textsc{Suit}_{1}$ (SFT)	$17.75$	$89.8$	$84.0$	$104.65$	$4838$
$\textsc{Suit}_{1}$ (DPO)	$17.44$	$88.5$	$82.7$	$103.04$	$4838$
$\textsc{Suit}_{2}$ (SFT-SFT)	$15.11$	$89.7$	$85.9$	$102.91$	$2493$
$\textsc{Suit}_{2}$ (SFT-DPO)	$17.17$	$89.5$	$84.4$	$104.12$	$2493$
$\textsc{Suit}_{2}$ (DPO-SFT)	$16.47$	$\boldsymbol{90.0}$	$\boldsymbol{87.1}$	$\boldsymbol{105.02}$	$2166$
$\textsc{Suit}_{2}$ (DPO-DPO)	$16.92$	$88.8$	$84.4$	$103.52$	$2166$