Institutional Repository
Technical University of Crete
EN  |  EL



My Space

Belief state space representation for statistical dialogue managers using deep autoencoders

Lygerakis Fotios

Full record

Year 2019
Type of Item Diploma Work
Bibliographic Citation Fotios Lygerakis, "Belief state space representation for statistical dialogue managers using deep autoencoders", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2019
Appears in Collections


Statistical Dialogue Systems (SDS) have proved their humongous potential over the past few years. However, the lack of efficient and robust representations of the belief state (BS) space refrains them from revealing their full potential. Furthermore, there is a great need for automatic BS representations, which will replace the old hand-crafted, variable-length ones. To tackle those problems, we introduce a novel use of the Autoencoder (AE). Our goal is to obtain a low-dimensional and compact, yet robust representation of the BS space. We investigate the use of dense AE, Denoising AE (DAE), Sparse Denoising AE (SDAE) and Variational Denoising AE (VDAE). Denoising approaches are particularly useful when corrupted BS vectors exist in environments with high uncertainty. We also explore the capabilities of Domain Independent Parameterization (DIP), a domain-independent alternative to the BS representation, and we combine them with different types of the AE, in order to obtain a more compact and robust representation of it. The BS space representation obtained from the AE is then used by two state-of-the-art Reinforcement Learning (RL) algorithms to learn the dialogue policies from simulated users in the PyDial toolkit; the non-parametric GP-SARSA and the parametric LSPI. In this framework, the BS is normally represented in a relatively compact, but still redundant summary space which is obtained through a heuristic mapping of the original master space. We show that the proposed AE-based representations consistently outperform the summary BS representation. Especially, as the Semantic Error Rate (SER) increases, the DAE/VDAE-based representations obtain state-of-the-art and sample efficient performance.

Available Files