基於失智長者語料之主題導向對話系統

Abstract

近年失智長者已成為迫切需要幫助與陪伴的一群。研究發現: 如果照顧者能與失智長者持續交談，可以相當程度穩定其心理和情緒狀態。因此本研究的目標為實現與失智症長者持續對話的軟體系統。根據在台北市立聯合醫院仁愛院區失智長者日間照護中心仁鶴軒收集到的語料，我們發現失智長者的對話多半侷限在少數主題，經常反復述說而不自知，恰好符合電腦系統不怕單調反復的優點。因此，我們不必採用複雜的自然語言處理系統，而只需要較簡單的規則庫方法，就能發展一套對話雛形系統。對話模式分為常規對話與主題式對話，系統依照當下狀況，判定使用何種模式來鼓勵失智症患者繼續進行對話。系統初步完成後，實際測試時，發現缺少自動語音識別以及語音合成模組，無法即時直接與失智症患者進行交談。因此後續只好改以修訂綠野仙蹤法，由照護中心人員參考雛型系統中的主題分類與對話規則，挑選適當主題，與真實失智症患者進行對話實驗。實驗結果顯示：我們由語料整理出來的少量主題，就可以涵蓋大部分談話主題，而且根據這些主題的對談，持續對話時間顯著延長。雖然我們沒能完成自動化程度更高的系統，但也發現這樣的研究方向似乎很合理，希望之後有機會完成完整系統，可以即時與失智症患者持續對話。

Patients of Alzheimer's disease have been in urgent need of help recently. Studies found that if caregivers maintain a conversation with the elderly with dementia, their mental and emotional state could be stabilized. This study aims to realize a software system for continuous dialogue with the elderly with dementia. From the corpus we collected for the Demented Elderly in the Taipei City Hospital Renai Branch, we found that conversations of the elderly with dementia are mostly limited to a few topics, often repeated un-noticeably. Therefore, a more straightforward rule base approach has been sufficient to develop a prototype dialogue system. The system will decide what sentences to respond to and encourage the patient to continue the conversation. However, the system could not communicate with dementia patients directly in real time due to the lack of automatic speech recognition and speech synthesis modules. Therefore, we use the modified Wizard of Oz method. The nursing center staff refer to the prototype system's topic classification and dialogue rules and select appropriate topics to conduct dialogue experiments. The experimental results show that the small number of topics we sorted out from the corpus can cover most of the topics, and the conversation duration based on these topics is significantly prolonged. Although the prototype system has a limited degree of automation, such a research direction seems reasonable. We hope that there will be an opportunity to complete a system that can continuously conduct real-time conversations for people with dementia in the future.