在實際情況中,話語由ASR(自動語音識別)系統轉錄,其通常提出多個候選轉錄(假設)。大多數時候,第一個假設通常是最好和最常用的假設。但是,在嘈雜的環境中,ASR的第一個假設經常會錯漏一些對LU(語言理解)而言很重要的詞,而這些詞經常可以在其他假設中找到。但總的來說,第一個ASR假設明顯優於其他的ASR假設。如果我們放棄第一個ASR假設,就因為它缺少一些單詞,這並不是最好的選擇。如果我們可以參考第2個ASR假設來修改第1個ASR假設的缺失的或冗餘的詞,我們可以使話語更接近使用者的真實意圖。在這篇論文中,我們提出了一種通過強化學習模型自動校正第1個ASR假設的方法。它可以通過地2假設逐字逐句糾正第一個假設。我們的方法將第1次ASR假設的得分從70.18提高到76.74。;In real situations, utterances are transcribed by ASR(Automatic Speech Recognition) systems, which usually propose multiple candidate transcriptions(hypothesis). Most of the time, the first hypothesis is the best and most commonly used. But the first hypothesis of ASR in a noisy environment often misses some words that are important to the LU(Language Understanding), and these words can be found among second hypothesis. But on the whole, the first ASR hypothesis is significantly better than the second ASR hypothesis. It is not the best choice if we abandon the first ASR hypothesis because it lacks some words. If we can refer to the 2th ASR hypothesis to modify the missing or redundant words of the first ASR hypothesis, we can get utterances closer to the user′s true intentions. In this paper we propose a method to automatically correct the 1th ASR hypothesis by the reinforcement learning model. It can correct the first hypothesis word by word by other hypothesis. Our method raises the bleu score of 1th ASR hypothesis from 70.18 to 76.74.