RRCGAN:對抗式學習強化機器閱讀理解模型可靠性

Abstract

儘管隨著深度學習模型技術的研發以及硬體計算能力的提升讓機器閱讀理解系統獲得快速的發展，在現實的條件中往往有許多模型可靠性的考量，一個缺乏可靠性的閱讀理解模型仍然欠缺實用性。在這篇研究中我們提出兩項閱讀理解任務的可靠性問題: 訓練資料稀少以及缺乏無法回答的問句的訓練資料，前者會造成模型無法充分發揮並有過度擬合的問題；後者則造成模型遇到無法回答的問句時會亂猜答案。 我們利用問句生成的技術希望能夠透過現有的資料產生出更多的問句進而有更多的閱讀理解訓練資料，同時，結合現有資料與生成資料時賦予每一筆資料一個來自問句評斷器的權重，來平衡品質好壞的問句對模型所造成的影響。對於缺少無法回答問句的問題，我們進一步利用對抗式生成網路的結構來進一步調整預訓練好的問句生成器，這樣的技術能夠解決預訓練時最大似然估計法所產生的問題，並產生出更接近真實問句的結果，另外，再將現有的可回答問句透過我們提出的無法回答問句替換規則轉化為一個偽無法回答問句，即可以用來訓練閱讀理解模型讓模型學習一些無法回答問句的辨識方法。 在許多實驗中也證明了我們提出的資料擴增方法是有效的，跟現有方法比較能夠一定程度提升閱讀理解模型的回答可靠性，同時也進一步分析實驗結果並討論我們的方法的優點與限制所在。

Despite the popularity of deep learning techniques applied in machine reading comprehension (MRC) systems, the robustness issues of the systems may slow down their deployment in real-world scenarios. We describe two of the robustness issues as data-limited condition MRC, which may constrain the capacity of the resultant model, and MRC without unanswerable questions, which makes unreliable guesses on unanswerable questions. In this research, we exploit the question generation (QG) technique aiming to expand the existing training triplets and loss weighting by a question discriminator to balance the influence of different quality questions. Generative adversarial net is further incorporated into the QG learning to alleviate the exposure bias caused by maximum likelihood estimation training. We also propose unanswerable question perturbation rules that convert an answerable question to a pseudo unanswerable one, which can be used to teach the MRC model what they do not know. Extensive experiments are conducted on these two tasks and demonstrate significant improvements over the baselines. We also analyze the experiment results and discuss the pros and cons of our proposed methods.