Microsoft Concept Graph: Mining Semantic Concepts for Short Text Understanding

Four tables and 23 figures of this paper. Table 1 shows the concept space comparison of existing taxonomies. Table 2 presents Hearst pattern examples. Table 3 shows labeling guideline for conceptualization. Table 4 presents precision of short text understanding. Figure 1 shows the framework overviews. Figure 2 is local taxonomy construction. Figure 3 shows horizontal merging. Figure 4 shows vertical merging: single sense alignment. Figure 5 shows vertical merging: multiple sense alignment. Figure 6 is a subgraph of heterogeneous semantic network around watch. Figure 7 is the compression procedure of typed-term co-occurrence network. Figure 8 presents an example of short text understanding. Figure 9 present examples of Chain model and Pairwise model. Figure 10 is a snapshot of the Probase browser. Figure 11 is a snapshot of single instance conceptualization.Figure 12 is a snapshot of context-aware single instance conceptualization. Figure 13 shows an example of short text conceptualization. Figure 14 is the framework of topic search. Figure 15 is a snapshot of the Web tables. Figure 16 shows query recommendation snapshot. Figure 17 shows the correlation of CTR with ads relevance score. Figure 18 presents the distribution of concepts in Microsoft Concept Graph. Figure 19 shows concept coverage of different taxonomies. Figure 20 shows precision of extracted isA pairs on 40 concepts.Figure 21 is precision of isA pairs after each iteration. Figure 22 shows the number of discovered concepts and isA pairs after each iteration. Figure 23 shows precision and nDCG comparison.

本文共包含四张表格和二十三幅图表。表1展示了现有分类法的概念空间比较。表2呈现了Hearst模式示例。表3展示了概念化的标注指南。表4展示了短文本理解的精确度。图1展示了框架概述。图2为局部分类法构建。图3展示了水平合并。图4展示了垂直合并：单一意义对齐。图5展示了垂直合并：多重意义对齐。图6是围绕手表的异构语义网络子图。图7展示了类型词共现网络的压缩过程。图8呈现了短文本理解的示例。图9展示了链式模型和成对模型示例。图10是Probase浏览器的快照。图11是单一实例概念化的快照。图12是上下文感知单一实例概念化的快照。图13展示了短文本概念化的示例。图14是主题搜索的框架。图15是Web表格的快照。图16展示了查询推荐快照。图17展示了点击率（CTR）与广告相关性评分的相关性。图18展示了Microsoft概念图中概念的分布。图19展示了不同分类法的概念覆盖范围。图20展示了在40个概念上提取的isA对的精确度。图21展示了每次迭代后isA对的精确度。图22展示了每次迭代后发现的概念和isA对的数量。图23展示了精确度和nDCG的比较。