| 摘要: | 隨著智慧家庭市場經濟規模的擴大，許多專屬的封閉生態系統逐漸形成，
例如 Apple 的 HomeKit、Google 的 Nest 和 Amazon 的 Alexa 等。這種封閉
的生態系統不僅限制了用戶的選擇，還迫使用戶學習不同系統的操作方式，難
以全面體驗智慧家庭應有的便利性。
本文旨在透過建構智慧家庭任務導向式對話系統，降低用戶的學習曲線
並突破封閉生態的問題。在建立任務導向式對話系統之前，需要先收集對話
語料庫。相較於傳統的人對人或人對機器語料收集方式，本文參考了 SchemaGuided Dialogue (SGD) 方法，並建構了 SmartHomeSGD 對話模擬器。透過
環境感知的機率方法決策用戶代理與助理代理的對話行為，使用戶代理人和助
理代理人能更有效地模擬人與系統的交互方式。
在對話行為設計方面，我們針對用戶代理和助理代理設計了專屬的對話
行為。此外，為了實現任務導向式對話系統與外部智慧家庭服務（傳統裝置、
多媒體播放器和暖通空調服務）的串接，我們加入了助理代理的 EXECUTE
對話行為，藉由此對話行為，發送 HTTP 請求，串接外部的智慧家庭服務。
在傳統機器對機器的語料生成中，機器生成的對話大綱往往需要大量人
力進行改寫，以提升語料的多樣性與自然性。為了解決此問題，本文設計了對
話改寫提示，以此引導大型語言模型執行對話改寫任務，有效降低人力成本。
最後，本文使用 mT5 (Multilingual Text-to-Text Transfer Transformer) 預
訓練模型作為基礎，並基於 SmartHomeSGD 語料庫進行微調，成功建構了中
文智慧家庭任務導向式對話系統。
;With the expansion of the smart home market, many proprietary closed
ecosystems have gradually emerged, such as Apple’s HomeKit, Google’s Nest,
and Amazon’s Alexa. These closed ecosystems not only limit users’choices
but also require them to learn different system operation methods, making it
difficult to fully experience the convenience that smart homes should offer.
This paper aims to address the learning curve and the issue of closed
ecosystems by constructing a task-oriented dialogue system for smart homes.
Before building the task-oriented dialogue system, a dialogue corpus must first
be collected. Compared to traditional methods of collecting human-to-human or
human-to-machine dialogues, this paper refers to the Schema-Guided Dialogue
(SGD) approach and constructs the SmartHomeSGD dialogue simulator. Using
a context-aware probabilistic method, the dialogue actions of the user agent and
assistant agent are decided, allowing the agents to more effectively simulate the
interactions between humans and systems.
In the design of dialogue actions, we created specific actions for both the
user agent and the assistant agent. Additionally, to integrate the task-oriented
dialogue system with external smart home services (such as traditional devices,
media players, and HVAC services), we added the EXECUTE dialogue action
for the assistant agent. Through this action, HTTP requests are sent to connect
to external smart home services.
In traditional machine-to-machine dialogue corpus generation, the dialogue
outlines generated by machines often require a significant amount of manual effort
to revise in order to increase diversity and naturalness. To address this issue,
this paper designs dialogue rewriting prompts to guide large language models in
performing dialogue rewriting tasks, effectively reducing human labor costs.
Finally, this paper uses the mT5 (Multilingual Text-to-Text Transfer Transformer) pre-trained model and fine-tunes it based on the SmartHomeSGD corpus
to successfully construct a chinese task-oriented dialogue system for smart homes |
