[Paper Review] Global-to-local Memory Pointer Networks for Task-Oriented Dialogue

Chien-Sheng Wu, Richard Socher+1 more|arXiv (Cornell University)|Jan 15, 2019

Topic Modeling24 references34 citations

TL;DR

GLMP introduces a global memory encoder and a local memory decoder with shared external knowledge to copy relevant KB data, achieving state-of-the-art results on both simulated and human-human task-oriented dialogue datasets, including strong OOV handling.

ABSTRACT

End-to-end task-oriented dialogue is challenging since knowledge bases are usually large, dynamic and hard to incorporate into a learning framework. We propose the global-to-local memory pointer (GLMP) networks to address this issue. In our model, a global memory encoder and a local memory decoder are proposed to share external knowledge. The encoder encodes dialogue history, modifies global contextual representation, and generates a global memory pointer. The decoder first generates a sketch response with unfilled slots. Next, it passes the global memory pointer to filter the external knowledge for relevant information, then instantiates the slots via the local memory pointers. We empirically show that our model can improve copy accuracy and mitigate the common out-of-vocabulary problem. As a result, GLMP is able to improve over the previous state-of-the-art models in both simulated bAbI Dialogue dataset and human-human Stanford Multi-domain Dialogue dataset on automatic and human evaluation.

Motivation & Objective

• Motivate end-to-end task-oriented dialogue systems to better incorporate large, dynamic knowledge bases.
• Propose a global-to-local memory pointer architecture that shares external knowledge between encoder and decoder.
• Improve copying of relevant KB information and mitigate OOV issues in dialogue generation.
• Demonstrate state-of-the-art performance on both simulated (bAbI) and human-human (SMD) datasets.
• Provide analysis and visualization of memory pointers to illustrate the copying process.

Proposed method

• Propose a three-part model: a global memory encoder, shared external knowledge, and a local memory decoder.
• Use a context RNN to encode dialogue history and write hidden states into external knowledge.
• Compute a global memory pointer by reading external memory with an auxiliary multi-label loss.
• Use a sketch RNN to generate a slot-delexicalized response, then instantiate slots with local memory pointers from external knowledge.
• Query external knowledge at each decoding step to produce local memory pointers for copying objects into the final response.
• Train jointly with losses for global pointer, sketch generation, and local pointer supervision.

Experimental results

Research questions

• RQ1 Can a global memory pointer combined with a local memory pointer improve copying of KB entities in end-to-end task-oriented dialogue?
• RQ2 Does sharing external knowledge between encoder and decoder mitigate OOV issues and improve robustness across datasets?
• RQ3 How does GLMP perform in both simulated and human-human task-oriented dialogue settings compared to state-of-the-art baselines?

Key findings

• GLMP achieves up to 92.0% per-response accuracy in the bAbI OOV setting, outperforming baselines.
• On bAbI Task 5, GLMP with multiple hops (K=1,3,6) reaches high completion rates and reduces OOV performance drop.
• In Stanford Multi-domain Dialogue (SMD), GLMP with K=1/3/6 achieves the highest BLEU and strong entity F1 and human-evaluated scores.
• Ablation shows removing the global memory pointer or history-writing component degrades performance, confirming their contributions.
• GLMP demonstrates improved entity copy accuracy and robustness to unknown words, surpassing Mem2Seq and other baselines.