在口語問答系統(Spoken Question Answering, SQA)中，一個簡單且直覺的作法，是先將一段音訊透過自動語音辨識(Automatic Speech Recognition, ASR)轉換成一連串的辨識文字結果，再輸入給現有各式基於文字的問答系統模型來完成任務需求。然而，這樣的做法通常會遭遇自動語音辨識錯誤(Recognition Errors)的影響，導致問答系統模型的效果不如預期。為了解決此一問題，本論文提出一種基於輸入特徵粒度的訓練策略，其目標是改善自動語音辨識錯誤所造成的效能損失，而且不需要額外模型的需求即可完成。我們將本論文所提出之訓練策略運用於中文口語機器閱讀理解(Machine Reading Comprehension, MRC)任務之中，驗證此一方法對於自動語音辨識錯誤的影響與改善。

In spoken question answering, a segment of audio is usually converted into a textual representation through an automatic speech recognition (ASR) system, and then input to a text-based question answering model to generate the answer. However, based on the ASR transcriptions, which usually contain lots of recognition errors, text-based question answering system may produce imperfect results. In order to mitigate the performance gap, in this study, a featured-granularity training strategy is proposed. Accordingly, we evaluate the proposed training strategy on spoken Chinese machine reading comprehension task, which not only demonstrates the capability and ability of the proposed strategy, but several valuable observations can be drawn from the experimental results.