基於追問問題建模的開放檢索對話式機器閱讀模型研究

Abstract

在當今充滿訊息的社會中，從龐大的數據中獲取所需要的訊息有重大挑戰性。人們期望能夠通過程式從廣泛的資料中檢索所需訊息並在推理判斷後提供所需答案。因此，我們在開放式檢索的框架下，建立了一個新的對話式機器閱讀理解資料集，名為WHITE-ShARC。同樣是對話式機器閱讀理解的設定下，與以往多數資料集不同，該資料集強調通過模型提問追問問題的方式逐步推理出用戶問題的答案，而不僅僅依賴代名詞指代以連接兩個不緊密關聯的問題及答案組。另外，相比OR-ShARC資料集，WHITE-ShARC在開放檢索和對話式機器閱讀理解的設定下擁有更廣泛的問題類型，並引入了無法回答的場景。

本論文將介紹WHITE-ShARC資料集的構建過程，並使用多種擴充方法，尤其是利用當前最火紅的大型語言模型ChatGPT。此外，我們使用retriever-reranker-reader的框架來處理對話式機器閱讀理解任務，並進行了全面的實驗和分析。特別的是，我們探索了ChatGPT在該任務中的應用，取得了一些正面的成果，同時也發現了一些潛在的改進空間。整體而言，通過建置這個新的資料集，我們試圖推進對話式機器閱讀理解領域的研究，促進其進一步的探索和進展。

In today's information-rich society, the ability to retrieve desired information from vast amounts of data is essential. There is a growing demand for systems that can effectively navigate complex information sources and provide accurate and relevant answers to user queries. In this thesis, we address this challenge by introducing a new conversational machine reading comprehension (CMRC) dataset, named WHITE-ShARC. Unlike previous datasets that rely on co-reference resolution to connect question-answer pairs, our dataset emphasizes the conversational aspect by requiring models to reason through follow-up questions to gradually arrive at the answers. WHITE-ShARC combines the characteristics of conversational dialogue and open-retrieval settings, enabling a broader range of question types and incorporating unanswerable instances.

The construction of WHITE-ShARC involves the use of various augmentation methods, including leveraging large language models like ChatGPT. We propose a retriever-reranker-reader framework to tackle the CMRC task and conduct comprehensive experiments and analyses, including exploring the capabilities of ChatGPT. The results demonstrate both positive contributions and potential areas for improvement. Overall, this research advances the field of CMRC by introducing a new dataset that encourages conversational reasoning and provides insights for future research in this area.