A Probabilistic End-To-End Task-Oriented Dialog Model with Latent Belief States towards Semi-Supervised Learning

Abstract

Structured belief states are crucial for user goal tracking and database query in task-oriented dialog systems. However, training belief trackers often requires expensive turn-level annotations of every user utterance. In this paper we aim at alleviating the reliance on belief state labels in building end-to-end dialog systems, by leveraging unlabeled dialog data towards semi-supervised learning. We propose a probabilistic dialog model, called the LAtent BElief State (LABES) model, where belief states are represented as discrete latent variables and jointly modeled with system responses given user inputs. Such latent variable modeling enables us to develop semi-supervised learning under the principled variational learning framework. Furthermore, we introduce LABES-S2S, which is a copyaugmented Seq2Seq model instantiation of LABES 1 . In supervised experiments, LABES-S2S obtains strong results on three benchmark datasets of different scales. In utilizing unlabeled dialog data, semi-supervised LABES-S2S significantly outperforms both supervisedonly and semi-supervised baselines. Remarkably, we can reduce the annotation demands to 50% without performance loss on MultiWOZ.

AuGPT: Auxiliary Tasks and Data Augmentation for End-To-End Dialogue with Pre-Trained Language Models

“…Recently, large-scale multi-domain task-oriented datasets were proposed (Budzianowski et al, 2018;Byrne et al, 2019;Rastogi et al, 2020). To address multiple domains, Zhang et al (2020a) introduce the LABES-S2S model that -in addition to a two-stage seq2seq approach -models belief states as discrete latent variables. Zhang et al (2020b) present DAMD, a three-stage seq2seq architecture which explicitly decodes the system action.…”

Shades of BLEU, Flavours of Success: The Case of MultiWOZ

“…It uses a multi-action data augmentation and multiple GRU (Cho et al, 2014) decoders. Similarly, LABES (Zhang et al, 2020a) employs a few GRU-based decoders, but it represents the dialog state as a latent variable. DoTS (Jeon and Lee, 2021) also uses GRUs, but the model makes use of a BERT encoder (Devlin et al, 2019) to get a context representation.…”

End-to-End Task-Oriented Dialog Modeling with Semi-Structured Knowledge Management

“…To reduce deployment cost and error propagation, end-toend trainable networks [22] are introduced into TOD systems, which have continually been studied recently. Typical endto-end TOD systems include those with structured fusion networks [3], [23], and those with multi-stage sequence-tosequence framework [2], [4], [24], [25]. With the boom of Transformers [26] and its large-scale pretraining [11], [12], TOD systems based on auto-regressive language modeling have also been developed [5]- [7], which achieve strong TOD modeling performance.…”

“…1) E2E TOD Models and DST Models: We consider two light-weight baseline E2E TOD models with different types of structures: UniConv [3] uses a structured fusion [23] design, while LABES-S2S [4] is based on a multi-stage Seq2Seq [24] architecture. Besides, we consider two large-scale baseline E2E TOD models developed from pretrained GPT-2 [12]: SimpleTOD [5] directly finetunes GPT-2 to model TOD in the auto-regressive manner, while AuGPT [7] further pretrains GPT-2 on large TOD corpus before finetuning and applies training data augmentation based on back-translation [42]- [44].…”

“…We set the total number of training epoch as 15, with average training time 25 minutes per epoch using 1 GPU. In SeKnow-PLM, we follow AuGPT [7] to pretrain a GPT-2 [12] as our model backbone, using its suggested implementation settings 4 . We further finetune the GPT-2 on the training set of modified MultiWOZ 2.1, where we set the batch size as 8.…”