[Paper Review] Global-Locally Self-Attentive Dialogue State Tracker

Victor W. Zhong, Caiming Xiong+1 more|arXiv (Cornell University)|May 19, 2018

Speech and dialogue systems35 references70 citations

TL;DR

GLAD introduces a global-local self-attentive encoder for DST that shares parameters across slots while learning slot-specific features, achieving state-of-the-art results on WoZ and DSTC2. It improves rare slot-value pair tracking via global-local parameter sharing and self-attention.

ABSTRACT

Dialogue state tracking, which estimates user goals and requests given the dialogue context, is an essential part of task-oriented dialogue systems. In this paper, we propose the Global-Locally Self-Attentive Dialogue State Tracker (GLAD), which learns representations of the user utterance and previous system actions with global-local modules. Our model uses global modules to share parameters between estimators for different types (called slots) of dialogue states, and uses local modules to learn slot-specific features. We show that this significantly improves tracking of rare states and achieves state-of-the-art performance on the WoZ and DSTC2 state tracking tasks. GLAD obtains 88.1% joint goal accuracy and 97.1% request accuracy on WoZ, outperforming prior work by 3.7% and 5.5%. On DSTC2, our model obtains 74.5% joint goal accuracy and 97.5% request accuracy, outperforming prior work by 1.1% and 1.0%.

Motivation & Objective

• Motivate improving dialogue state tracking by addressing rare slot-value pairs in DST datasets.
• Propose a global-locally self-attentive encoder to share parameters across slots while learning slot-specific features.
• Demonstrate state-of-the-art performance on WoZ and DSTC2 through end-to-end training without hand-crafted lexicons.

Proposed method

• Use a global bidirectional LSTM to capture general features across all slots (global module).
• Use a local bidirectional LSTM conditioned on the specific slot (local module).
• Combine global and local encodings with a slot-specific weighting to form the final encoder representation (beta^s).
• Apply global and local self-attention to compute context vectors (c^g and c^s) and form a global-local attention context (c).
• Encode user utterance, previous system actions, and slot-value pair with the GLAD encoder, then score slot-value pair predictions using two scorers (one from utterance, one from actions) and a learned weighting to produce the final score.

Experimental results

Research questions

• RQ1 Can sharing parameters across slots while maintaining slot-specific representations improve DST, particularly for rare slot-value pairs?
• RQ2 Do global and local self-attentive components provide complementary benefits for accurate turn-level and joint goal tracking?
• RQ3 How does GLAD perform relative to prior state-of-the-art on DSTC2 and WoZ datasets under noisy ASR conditions (DSTC2) and clean utterances (WoZ)?

Key findings

• GLAD achieves state-of-the-art results: 88.1% joint goal accuracy and 97.1% request accuracy on WoZ.
• GLAD achieves 74.5% joint goal accuracy and 97.5% request accuracy on DSTC2.
• Ablation studies show global-local sharing and self-attention contribute to robustness for rare slot-value pairs, outperforming variants that remove either global sharing or slot-specific components.
• On WoZ, GLAD outperforms prior best by 3.7% in joint goal accuracy and 5.5% in turn-level request accuracy.
• On DSTC2, GLAD outperforms prior best by 1.1% in joint goal accuracy and 1.0% in turn-level request accuracy.