题目： Relation Adversarial Network for Low Resource Knowledge Graph Completion
摘要： 知识图谱补全(Knowledge Graph Completion, KGC)是一种通过链接预测或关系提取来填充缺少的链接来改进知识图谱的方法。KGC的主要困难之一是资源不足。之前的方法假设有足够训练的三元组来学习实体和关系的通用向量，或者假设有足够数量的标签句子来训练一个合格的关系提取模型。然而，在KGs中，少资源关系非常普遍，这些新增加的关系往往没有很多已知的样本去进行训练。在这项工作中，我们的目标是在具有挑战性的环境下只有有限可用的训练实例预测新的事实。我们提出了一个加权关系对抗性网络的通用框架，它利用对抗性过程来帮助将从多资源关系中学习到的知识/特征调整为不同但相关的少资源关系。具体地说，该框架利用了一个关系鉴别器来区分样本和不同的关系，帮助学习更容易从源关系转移到目标关系的关系不变量特征。实验结果表明，该方法在少资源设置下的链路预测和关系提取都优于以往的方法。
Reader reviews of literary fiction on social media, especially those in persistent, dedicated forums, create and are in turn driven by underlying narrative frameworks. In their comments about a novel, readers generally include only a subset of characters and their relationships, thus offering a limited perspective on that work. Yet in aggregate, these reviews capture an underlying narrative framework comprised of different actants (people, places, things), their roles, and interactions that we label the "consensus narrative framework". We represent this framework in the form of an actant-relationship story graph. Extracting this graph is a challenging computational problem, which we pose as a latent graphical model estimation problem. Posts and reviews are viewed as samples of sub graphs/networks of the hidden narrative framework. Inspired by the qualitative narrative theory of Greimas, we formulate a graphical generative Machine Learning (ML) model where nodes represent actants, and multi-edges and self-loops among nodes capture context-specific relationships. We develop a pipeline of interlocking automated methods to extract key actants and their relationships, and apply it to thousands of reviews and comments posted on Goodreads.com. We manually derive the ground truth narrative framework from SparkNotes, and then use word embedding tools to compare relationships in ground truth networks with our extracted networks. We find that our automated methodology generates highly accurate consensus narrative frameworks: for our four target novels, with approximately 2900 reviews per novel, we report average coverage/recall of important relationships of > 80% and an average edge detection rate of >89\%. These extracted narrative frameworks can generate insight into how people (or classes of people) read and how they recount what they have read to others.