PUB-R：基於回合定位的端到端任務對話系統

Abstract

任務導向對話系統(Task-Oriented Dialog System)在無數行業中都有巨大的需求。可以大大降低客戶服務人員的管理費用並簡化人力資源流程。許多 TOD 系統皆使用 GPT-2 模型作為基底，但這些系統沒有考慮多輪對話的關鍵考慮因素－－對話回合的位置。本論文 PUB-R 為最近的端到端任務對話模型引入了一種新的嵌入輸入方法：“基於回合的定位嵌入方法”（TPEM）。實驗結果顯示：(1) 與之前的 SOTA 模型相比，PUB-R 可以在更短的訓練時間內獲得更好的訓練性能；(2) 可以透過減少訓練所需的 epoch 數而不影響性能來實現模型的快速收斂。我們的實驗成功地改進了以前的“基於回合的定位”的端到端對話系統，在端對端對話評估方法上取得更高的分數，訓練時間更短，並且不需要額外的手動註釋。

In high demand across countless industries, Task-Oriented Dialog (TOD) systems may greatly reduce the overhead costs of customer service personnel and simplify human resource processes. The GPT-2 model is used across a variety of TOD systems, but these systems do not take into account the turn number, a critical consideration for multi-turn dialogs. Our paper PUB-R introduces a new embedding input method for recent end-to-end task dialog models : the “Turn-Based Positioning Embedding Method” (TPEM). Our results show that (1) PUB-R can obtain better training performance in a shorter training time compared with previous SOTA models and (2) rapid model convergence can be achieved by reducing the number of epochs required without compromising performance. Our implementation successfully improves upon previous end-to-end dialog systems in evaluation score of the model with the "turn-based positioning" with shorter training times and without requiring additional manual annotation.