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

Yichi Zhang, Zhijian Ou+2 more|arXiv (Cornell University)|Sep 17, 2020

Topic Modeling40 references4 citations

TL;DR

This paper proposes LABES, a probabilistic end-to-end task-oriented dialog model that treats belief states as latent variables, enabling semi-supervised learning using unlabeled dialog data. By leveraging variational inference and a copy-augmented Seq2Seq architecture (LABES-S2S), the model achieves state-of-the-art performance on benchmark datasets and reduces annotation requirements by 50% on MultiWOZ without performance loss.

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 copy-augmented Seq2Seq model instantiation of LABES. 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 supervised-only and semi-supervised baselines. Remarkably, we can reduce the annotation demands to 50% without performance loss on MultiWOZ.

Motivation & Objective

• To reduce reliance on expensive turn-level belief state annotations in end-to-end task-oriented dialog systems.
• To enable effective semi-supervised learning by leveraging unlabeled dialog data through latent variable modeling.
• To develop a probabilistic generative model that jointly infers belief states and generates system responses.
• To demonstrate that unlabeled dialog data contain exploitable cues for belief tracking, even without explicit annotations.
• To achieve strong performance with significantly reduced annotation requirements, especially in low-resource settings.

Proposed method

• Models dialog systems as a conditional generative model: pθ(b1:T, r1:T | u1:T), where belief states b1:T are latent variables.
• Employs a variational inference framework to optimize the model on mixed labeled and unlabeled data.
• Introduces LABES-S2S, a copy-augmented sequence-to-sequence instantiation for end-to-end response generation and belief tracking.
• Uses attention mechanisms and copy mechanisms to preserve slot values and improve response fluency.
• Leverages implicit cues in user inputs and system responses—such as word repetition or query results—to inform belief state inference.
• Applies posterior regularization and autoencoding objectives within the variational learning framework to improve generalization.

Experimental results

Research questions

• RQ1 Can belief states be effectively modeled as latent variables to reduce dependency on annotated training data?
• RQ2 To what extent can unlabeled dialog data improve belief tracking and response generation in end-to-end dialog systems?
• RQ3 How does the proposed semi-supervised learning framework compare to supervised-only and prior semi-supervised baselines in low-resource settings?
• RQ4 Can the model maintain or improve performance when training data is reduced by 50%, particularly on large-scale benchmarks like MultiWOZ?
• RQ5 What role do implicit cues in user inputs and system responses play in enhancing belief state inference without explicit annotations?

Key findings

• LABES-S2S achieves state-of-the-art joint goal accuracy on CamRest676 and In-Car, and outperforms non-pretrained models on MultiWOZ 2.1.
• Semi-supervised LABES-S2S significantly outperforms both supervised-only and prior semi-supervised baselines across all three benchmark datasets.
• On MultiWOZ, the model maintains a joint goal accuracy of 49.47% with only 50% labeled data, compared to 50.05% with 100% labeled data—indicating no performance loss at 50% annotation cost.
• The model reduces annotation requirements by approximately 30,000 belief state annotations on MultiWOZ without performance degradation.
• The semi-supervised variant achieves larger performance gains when labeled data is scarce, with the largest improvements observed at 10% labeling proportion.
• Case studies show that semi-supervised learning enables the model to capture implicit user goals and improve response quality by learning cue words like 'British' and avoiding error propagation.