III: Small: Multi-field Hierarchical Discovery and Tracking (mf-HDT) of Emerging Topics

该项目的目标是解决基于不同数据中的组合证据在不同粒度级别对新兴主题的多字段分层发现和跟踪(MF-HDT)的开放挑战。这些技术方法包括一个新的贝叶斯框架和强大的推理算法，即多字段层次相关主题建模，用于发现潜在主题的多字段层次，捕获主题间和跨层次的相关性，并允许在马尔科夫层次链上进行查询驱动的主题线程。这些技术创新和能力超越了现有的主题检测和跟踪(TDT)方法和用于表示主题、引文等之间关系的图形模型。预计在有效性和可扩展性方面都将比现有方法有显著改进，特别是在检测新出现的主题和跟踪时间敏感影响方面。拟议的方法将在范围广泛的科学文献数据(物理、数学、计算机科学、数量生物学、数量金融和统计学)以及新闻故事的四个大数据集上进行评估，并使用人为产生的查询和相关性判断以及人为分配的主题标签，以支持面向任务的评估。研究人员、教育从业者和学生、支持研究和行业的政府机构的生产率在很大程度上取决于许多领域内和跨领域科学涌现和共同涌现的最新大图景的可用性，以及新技术影响的证据和研究或开发资金。拟议的技术如果成功，将提供原则性和有效的解决方案，在上述和更广泛的应用中具有广泛的未来影响。网站(http://nyc.lti.cs.cmu.edu/mfhdt/)将提供开放源码软件、数据集、结果和出版物，以便能够进行比较评价和相关研究界的进一步研究。参与该项目的学生受益于在现实世界应用程序中使用和评估尖端IT技术的直接经验。这是对课堂教学的补充，在课堂教学中，学生可以第一手观察选择各种分类、主动学习和分布式计算策略的直接含义。

The goal of this project addresses the open challenge of Multi-field Hierarchical Discovery and Tracking (mf-HDT) of emerging topics at different granularity levels based on combined evidence in heterogeneous data. The technical approaches consist of a new Bayesian framework with powerful inference algorithms, namely the multi-field Hierarchical Correlated Topic Modeling, for discovering multi-field hierarchies of latent topics, capturing inter-topic and cross-hierarchy correlations, and enabling query-driven threading of topics over a Markov chain of hierarchies. These technical innovations and capabilities go beyond existing Topic Detection and Tracking (TDT) methods and graphical models used to represent relationships between topics, citations, etc. Significant improvements are expected in both effectiveness and scalability over the existing methods, especially in terms of detecting newly emerging topics and tacking time-sensitive impact. The proposed approach will be evaluated on a four large datasets of scientific literature data in a broad range (Physics, Mathematics, Computer Science, Quantitative Biology, Quantitative Finance and Statistics) as well as news stories, with human-produced queries and relevance judgments and human-assigned topic labels to support task-oriented evaluations. Productivity of researchers, educational practitioners and students, government agencies supporting research and industries highly depends on the availability of up-to-date big pictures about scientific emergence and co-emergence within and across many fields, along with evidence of the impact of new technologies, and research or development funding. The proposed techniques, if successful, will provide principled and effective solutions with a broad future impact in the applications above and beyond. Web site (http://nyc.lti.cs.cmu.edu/mfhdt/) will provide access to open-source software, of datasets, results and publication in order to enable comparative evaluations and further studies by related research communities. The students involved in the project benefit from direct experience with using and evaluating cutting-edge IT technologies in real-world applications. This is complementary to classroom teaching where the students can observe first-hand the direct implication of choosing various strategies for categorization, active learning and distributed computing.