观点型机器阅读理解是自然语言处理和文本挖掘领域的重要研究问题。现有方法未区别事实型问题和观点型问题，缺乏对观点特征进行针对性刻画，难以对“问题-文档-观点”之间的语义关系进行统一建模，导致模型性能受限。因此本项目拟开展面向观点型机器阅读理解的关键技术研究。具体包括：通过对文档和问题进行基于网络节点嵌入的交互学习，研究融合观点特征的“问题-文档”交互表示；利用异步注意力编码器方法进行观点证据的提取，研究基于问题匹配的观点证据选择；通过观点合并网络对答案生成过程注入观点证据，研究基于观点驱动的答案生成；应用“自举法”和回译方法进行数据扩展，研究大规模高质量数据集的构建。在此基础上，构建面向金融证券领域的观点型阅读理解示范应用。本研究将推动观点型阅读理解的基础研究，也为阅读理解相关应用提供借鉴。

Opinion reading comprehension is a significant topic in the field of natural language processing and text mining. Existing methods neglect to distinguish between fact-based question and opinion-based question. These previous models insufficiently consider opinion-based features, and fail to learn the unified semantic representation among question, passage and opinion, which leads to an inferior performance. Therefore, we plan to study key technologies on opinion reading comprehension in this proposal. Firstly, we will apply interactive learning based on network node embedding to passage and question, then obtain the opinion-aware representation of "passage-question" by integrating opinion features. Secondly, we will utilize asynchronous attention encoder to extract opinion evidences in order to select opinion evidences based on question matching. Thirdly, in answer reasoning, we plan to inject opinion evidences into the answer generation process by the opinion merging network to achieve opinion-driven answer generation. Finally, we will combine bootstrapping and back-translation methods to implement data augmentation for building a large-scale and high-quality opinion reading comprehension data set. Benefiting from these methods, the result of our work will improve the research of opinion reading comprehension, and it will be also implemented in practical applications such as financial prediction.