中国科学院自动化研究所机构知识库

随着互联网的迅猛发展，社会媒体已成为人们发表观点、交换看法的一种重要渠道。社会媒体用户通过发布原创内容、参与公共讨论等方式表达自身对于争议对象（例如，实体、短命题、待查事实等）的立场倾向（支持、反对与中立）以及支持/反对意见。面向社会媒体的立场识别与观点摘要生成旨在挖掘海量社会媒体数据中蕴含的公众观点，准确判断出社会媒体文本对于特定对象的立场，并对各个立场持有方（即，支持方、反对方）所表达的核心观点进行总结概括。该研究课题不但在社会媒体分析和文本挖掘等研究领域具有重要意义，并且在国家公共安全和商业智能等应用领域具有良好价值。一方面，已有的立场识别与观点摘要生成方法大多为有监督机器学习模型，模型性能较大程度上依赖于标注数据的质量和数量。由于人工标注大量数据的代价昂贵、费时费力，实际应用中通常需要面对缺少标注数据的低资源情形，因此现有方法的性能往往受到了严重制约。另一方面，已有的立场识别方法在学习文本表示时较少考虑文本中的隐式表达现象以及文本在会话中的结构上下文信息，观点摘要生成方法大多缺乏对于摘要内容的覆盖度与冗余度控制。本论文聚焦于低资源情形下的立场识别与观点摘要生成问题，深入研究如何设计弱监督学习、迁移学习等方法进行有效的立场识别与观点摘要生成，并力求解决文本立场识别中的隐式表达和结构上下文建模问题，以及观点摘要生成中的覆盖度和冗余度控制等难点问题。 

本论文的主要贡献与创新点归纳如下：

1. 以往的弱监督立场识别工作通常仅将对象信息融合到文本表示中，使得隐式地表达立场的文本难以被有效建模，并且使用自动标注的含噪数据集训练模型时往往没有考虑错误立场标签带来的负面影响。针对上述问题，提出了一种基于多视角表示建模与强化学习的弱监督立场识别方法。该方法利用与对象相关的话题信息来建模多视角的文本表示，捕捉文本中的隐式表达；进一步，基于策略梯度算法来学习能够移除噪声数据的去噪策略，缓解错误立场标签对于模型训练的干扰，并提升样本使用率来加速去噪过程。基准数据集上的实验结果验证了所提出方法的有效性。

2. 以往的跨对象立场识别工作通常仅依靠源对象的有标注数据进行学习，忽略了目标对象的大量无标注数据的潜在价值，难以捕捉到可以跨对象迁移的知识。针对上述问题，提出了一种基于可迁移话题建模与对抗学习的跨对象立场识别方法。该方法利用源对象与目标对象之间的共享隐话题作为可迁移知识，从无标注数据中进行话题知识获取，并将其融入文本表示学习过程中以提升立场分类器的跨对象适应性；同时，通过一个统一的对抗迁移学习框架来端到端地学习话题知识获取和立场分类器这两个部分。基准数据集上的实验结果验证了所提出方法的有效性。

3. 以往面向待查事实的立场识别和事实查验工作往往忽略了文本在会话中的结构上下文信息以及对立场时序动态性的利用，难以学习到有效的特征表示并且跨事件通用性较差。针对上述问题，提出了一种基于会话结构与立场时序建模的层次化多任务学习方法用于立场识别和事实查验。该方法建模完整的会话结构，聚合关键的会话上下文信息来对文本进行立场识别，并通过进一步建模立场随事件发展所展现出的时序演化模式来进行事实查验；在此基础上，两个呈现级联关系的任务通过层次化的方式来联合学习。基准数据集上的实验结果验证了所提出方法的有效性。

4. 以往的观点摘要生成工作较少面向成对数据稀少的低资源情形，并且在生成观点摘要时大多缺乏对信息覆盖度和冗余度的控制。针对上述问题，提出了一种基于预训练词表示与边缘相关度引导的观点摘要生成方法。该方法从大规模预训练模型中获取词表示，利用其充分捕捉了海量外部数据中蕴含的世界知识的特点来辅助低资源情形下的模型学习；为了使生成的摘要内容能够覆盖到分散在不同观点句中的重要信息，利用结合了显著性和冗余性的边缘相关度来引导解码过程，通过避免生成的摘要内容仅关注到部分观点句来降低冗余度、提升覆盖度。基准数据集上的实验结果验证了所提出方法的有效性。

With the rapid development of the Internet, social media platforms have become the important channels for people to express opinions and exchange viewpoints. Users in social media publish original posts and public comments to express their stance (favor, against, or neutral) towards controversial targets (e.g., entities, short claims, and rumors) and corresponding reasons. Stance detection and opinionated text summarization for social media texts can mine public opinions expressed in large-scale social media data, detecting stances towards targets and summarizing the core opinions of each stance holder (i.e., supporter and opposition). This is a core research topic in social media analytics and text mining, and is of great application value in national security and business intelligence. On one hand, current approaches for stance detection and opinionated text summarization are mainly based on supervised machine learning models, whose performances largely rely on the amount of labeled data. Because collecting sufficient labeled data is time-consuming and labor-intensive, in real world applications we are usually faced with the low-resource scenarios (i.e., the lack of labeled data), and the performances of current approaches are quite limited. On the other hand, existing stance detection approaches usually omit the implicit expression and structural context modeling issues, and opinionated text summarization approaches lack the controllability on informativeness and redundancy of summaries. This dissertation focuses on low-resource stance detection and opinionated text summarization, which takes an in-depth look at the solutions that how to exploit weakly-supervised learning and transfer learning for detecting stance and summarizing opinionated texts effectively, and aims at addressing the issues of implicit expression and context modeling in stance detection, as well as informativeness and redundancy control in opinionated text summarization. 

The main contributions of this dissertation are summarized as follows:

1) Previous methods for targeted stance detection only consider target information during learning text representations, which results in ineffective representation when a text implicitly expresses its stance towards the target. Besides, when they are trained with automatically-annotated noisy data, they cannot handle the negative impact produced by noisy stance labeling. To address the above issues, we propose a multi-view reinforced model for weakly-supervised stance detection. The model utilizes target-related topic information to learn multi-view text representations, capturing implicit expression in texts. Further, it learns a denoising policy to eliminate noisy stance labeled data with policy gradient algorithms, reducing the negative impact of noisy data. And it accelerates the denoising via improving sample-efficiency. Experimental results on benchmark dataset verify the effectiveness of the proposed method.

2) Previous methods for cross-target stance detection are only trained with labeled data from source target, which ignores the potential value of unlabeled data from destination target and usually cannot capture transferable knowledge across targets. To address the above issues, we propose a transferable topic based adversarial model for cross-target stance detection. The model utilizes shared latent topics as transferable knowledge between source target and destination target, which acquires topic knowledge from unlabeled data, and incorporates such knowledge in text representation learning to improve the transfer ability across targets. Topic knowledge acquisition and stance classification are integrated into an adversarial transfer learning framework, which performs end-to-end optimization. Experimental results on benchmark datasets verify the effectiveness of the proposed method. 

3) Previous methods for rumor stance detection omit structural context information in conversation and fail to modeling the temporal dynamics of stance, thus their transfer ability across events is usually limited. To address the above issues, we propose a hierarchical multi-task learning method for rumor stance detection and rumor verification via modeling conversation structure and stance evolution. The method models entire conversation structures to learn stance features effectively for texts in the conversation via aggregating structural context information. Further, it models the temporal dynamics of stance evolution for rumor verification. Two cascaded tasks are jointly learned with a hierarchical fashion. Experimental results on benchmark datasets verify the effectiveness of the proposed method.

4) Previous methods for opinionated text summarization focus less on the low-resource scenario where only a few paired data is available. Besides, they usually lack the controllability on the informativeness and redundancy of generated summaries. To address the above issues, we propose an opinionated text summarization method with pretrained word representations and marginal relevance guidance. It exploits contextualized word representations from large pretraining models to utilize its nature of capturing world knowledge that helps model training with small paired data. Further, to capture important opinions that are scattered over multiple different opinionated texts, it adopts marginal relevance that considers both saliency and non-redundancy to guide the summary generation process, refraining from paying attention to partial contents for reducing redundancy and improving informativeness. Experimental results on benchmark datasets verify the effectiveness of the proposed method.