[Paper Review] Autoregressive Entity Generation for End-to-End Task-Oriented Dialog

Guanhuan Huang, Xiaojun Quan+1 more|arXiv (Cornell University)|Sep 19, 2022

Topic Modeling4 citations

TL;DR

This paper proposes an end-to-end task-oriented dialog system that first autoregressively generates entities from a knowledge base before generating responses, using a trie constraint to ensure entity consistency and a logit concatenation strategy to enable end-to-end training. Experiments on MultiWOZ 2.1 and CAMREST show significant improvements in response quality and entity consistency over strong baselines like GPT-KE.

ABSTRACT

Task-oriented dialog (TOD) systems often require interaction with an external knowledge base to retrieve necessary entity (e.g., restaurant) information to support the response generation. Most current end-to-end TOD systems either retrieve the KB information explicitly or embed it into model parameters for implicit access.~While the former approach demands scanning the KB at each turn of response generation, which is inefficient when the KB scales up, the latter approach shows higher flexibility and efficiency. In either approach, the systems may generate a response with conflicting entity information. To address this issue, we propose to generate the entity autoregressively first and leverage it to guide the response generation in an end-to-end system. To ensure entity consistency, we impose a trie constraint on entity generation. We also introduce a logit concatenation strategy to facilitate gradient backpropagation for end-to-end training. Experiments on MultiWOZ 2.1 single and CAMREST show that our system can generate more high-quality and entity-consistent responses.

Motivation & Objective

• To address entity inconsistency in end-to-end task-oriented dialog systems, where models generate conflicting or incorrect entity information.
• To eliminate reliance on external knowledge bases during inference by encoding KB information into model parameters through data augmentation.
• To improve response quality and entity consistency by generating the target entity first and using it as input for response generation.
• To enable end-to-end training despite the non-differentiable nature of integer-encoded entity tokens through a novel logit concatenation strategy.
• To ensure generated entities are valid KB entries by enforcing a trie constraint during decoding.

Proposed method

• Augment the training data by inserting KB entries into dialog contexts to embed knowledge into model parameters without scaling with KB size.
• Use an autoregressive decoder to generate the entity token sequence first, conditioned on the dialog context and user utterance.
• Enforce a trie constraint during decoding to filter out invalid entity tokens, ensuring only valid KB entities are generated.
• Introduce a logit concatenation mechanism that combines the logit vectors of the dialogue context, user utterance, and generated entity for gradient backpropagation.
• Train the entire system end-to-end using cross-entropy loss on the final response generation, with the generated entity serving as an input token sequence.
• Apply stop-gradient operations selectively to isolate the entity generation path during training, preserving gradient flow to the entity generator.

Experimental results

Research questions

• RQ1 Can autoregressive entity generation improve entity consistency in end-to-end task-oriented dialog systems?
• RQ2 How does enforcing a trie constraint during decoding affect the validity and consistency of generated entities?
• RQ3 To what extent does logit concatenation enable effective backpropagation through the entity generation module in an end-to-end training setup?
• RQ4 How does the gap between training (with LogitConcat) and inference (without) affect model performance and entity consistency?
• RQ5 How does the number of templates used for KB augmentation influence model performance on downstream metrics?

Key findings

• The proposed ECO system outperforms GPT-KE, a strong end-to-end baseline, by a large margin on MultiWOZ 2.1 and CAMREST, especially in entity consistency and success rate.
• ECO achieves a 5.31 absolute improvement in Inform score and 4.87 in Success rate over the variant without logit concatenation, demonstrating its effectiveness in end-to-end training.
• Removing the trie constraint reduces performance by 2.95 on Inform, 3.25 on Success, and 0.53 on Consistency, confirming its critical role in entity validity.
• Using only the original training set (ECO w/ tr) leads to a 16.08 drop in Inform and 20.5 drop in Success compared to full training, showing the importance of augmented data with entity labels.
• Applying LogitConcat during evaluation (ECO w/ LogitEval) causes significant performance drops, indicating that it disrupts the probability distribution and harms entity consistency.
• Performance improves with more templates for KB embedding, though fluctuations occur between 100 and 400 templates, suggesting diminishing returns and sensitivity to template selection.