Learning Table Similarity Measures

表格是在非结构化文本(如报告、出版物或网页)中嵌入结构化信息片段的一种有效且流行的方法。但是，表的特殊属性(二维结构、标题、列或行中的语义同质性，...)在典型的检索方法中被忽略。另一方面，直接查找与给定搜索条件匹配的表将提供对大量结构化信息的快速访问。实现这种功能的一种方法是表相似性搜索：给定一个查询表，在给定表语料库中找到最相似的表。在本项目中，我们将研究学习高质量表相似性度量的方法，作为表相似性搜索方法的基础，也可以用于其他应用，如表信息提取、表聚类或表融合。特别是，我们将研究设计有监督的表格相似性度量的深度学习方法，其目标是1)自动识别表格方向，2)在多个抽象级别学习合适的表格表示，以及3)将这些表示合并到单个表格相似性分数中。所有方法都将在黄金标准注释表语料库上进行评估，并与不同的最先进方法进行比较。所有语料库和软件都将在许可的开放获取许可证下发布。

Tables are an efficient and popular mean to embed structured pieces of information in unstructured texts, such as reports, publications, or web pages. However, the particular properties of tables (two-dimensional structure, headers, semantic homogeneity in columns or rows, ...) are disregarded in typical retrieval methods. On the other hand, directly finding tables matching a given search criterion would offer fast access to a wealth of structured information. One way to achieve such functionality is table similarity search: Given a query table, find the most similar tables in a given table corpus. In this project, we will research methods to learn high-quality tablesimilarity measures as fundamental pieces of table similarity search methods, but also for other applications such as table information extraction, table clustering, or table fusion. In particular, we will study deep learning methods for designing supervised table similarity measures with the objectives of 1) automatic identification of table orientation, 2) learning appropriate table representations at multiple levels of abstraction, and 3) merging these representations into a single table similarity score. All methods will be evaluated on a gold standard annotated table corpus and compared to different state-of-the-art methods. All corpora and software will be published under a permissive open access license.