报告题目：Computing Problems from the Humanities
报告人：Associate Prof. Steve Cassidy, Macquarie University, Sydney, Australia
Collaborating with researchers in the Digital Humanities is the source of a number of interesting problems in Computer Science. This talk looks at some examples from my past work and some emerging problems from more recent work.
Linguists and other language researchers make use of large collections of language data in the form of text, audio and video recordings. They like to annotate this data to add information to the base signal. This can be textual annotation such as grammatical structure or the location of named entities or structure in spoken language such as speaker turns or the location of words and phonetic segments. My work with this kind of data has looked at how best to represent these annotations and how to work with very large collections of annotations. Interesting problems include the development of a graph-based annotation model, the use of query languages to find interesting annotations and the problem of version control for annotation data stored as a graph.
More recently we have begun working with Oral History researchers who record stories from the community as the source material for their research. Our goal is to make use of speech and language technology to help the researchers make the most of the data that they collect. One of the interesting problems here is about segmenting a long interview recording into useful chunks and assigning those to different speakers. This is known as speaker diarization and is a well understood task but there are still some outstanding problems that happen to turn up when looking at these interviews. Once we have a transcript we can also begin to look at how to apply some Natural Language Processing techniques to this data to make it more useful. I’ll present some of the research directions we are looking at as we explore this area.
Steve Cassidy is a computer scientist and has working various areas relating to speech and language technology over the last 30 years after completing a PhD in Cognitive Science.
With Jonathan Harrington, he developed the Emu Speech Database System to support corpus-based research in speech and acoustic phonetics. Emu supports a flexible hierarchical annotation system and provides a query language and analysis environment based on the R Statistical environment. Emu is widely used to support research on small and large-scale speech corpora and includes tools to support every stage of the corpus collection and analysis lifecycle. Emu is now maintained by a team of developers in Munich.
He was recently involved in the development and collection of an audio-visual corpus of Australian English from around 1000 speakers around Australia. He built the software for data capture and a server-based system for data upload and publishing.
His most recent work has been on the Alveo Virtual Laboratory which is both a repository for language resources and a platform to support tools for exploration and analysis of language data. Alveo currently holds around 30 collections including audio, video and text resources and is working on new acquisitions of data and tools.
报告题目：Exploring Features for Complicated Objects: Cross-View Feature Selection for Multi-Instance Learning
In traditional multi-instance learning (MIL), instances are typically represented by using a single feature view. As MIL becoming popular in domain specific learning tasks, aggregating multiple feature views to represent multi-instance bags has recently shown promising results, mainly because multiple views provide extra information for MIL tasks. Nevertheless, multiple views also increase the risk of involving redundant views and irrelevant features for learning. To this end, we formulate a new cross-view feature selection problem that aims to identify the most representative features across all feature views for MIL. To achieve the goal, we design a new optimization problem by integrating both Multiview representation and multi-instance bag constraints. The solution to the objective function will ensure that the identified top-m features are the most informative ones across all feature views. Experiments on two real-world applications demonstrate the performance of the cross-view feature selection for content-based image retrieval and social media content recommendation.
Jia Wu（吴佳）：国际数据挖掘顶级期刊ACM Transactions on Knowledge Discovery Data（TKDD）副主编。澳大利亚麦考瑞大学计算机学院讲师，博士、IEEE会员。主要研究领域为数据挖掘、机器学习、人工智能，及其在商业、工业、生物信息学、医疗信息学等领域的应用。迄今，在国际学术期刊和会议上共发表论文100多篇, 包括IEEE Transactions on Knowledge and Data Engineering、IEEE Transactions on Neural Networks and Learning Systems、IEEE Transactions on Cybernetics、ACM Transactions on Knowledge Discovery Data、IEEE Transactions on Systems, Man, and Cybernetics: Systems、Pattern Recognition、IJCAI、AAAI、ICDM、SDM、CIKM等。指导学生曾获得2017顶级国际神经网络大会International Joint Conference on Neural Networks (IJCNN) 最佳学生论文奖、2014顶级国际数据挖掘会议International Conference on Data Mining的最佳论文提名奖。现任SCI、JCR一区期刊Journal of Network and Computer Applications副主编和Complexity Journal (SCI: 3.514) 客座主编。担任国际顶级神经网络大会2016、2017、2018 International Joint Conference on Neural Networks的专题分会主席 (Special Session Chair)、2018 International Conference on Applications and Techniques in Information Security 的程序委员会主席(Program Committee Chair)、顶级人工智能国际会议International Joint Conference on Artificial Intelligence, IJCAI 2017和2018的高级程序委员 (Senior Program Committee)，顶级国际学术会议的程序委员 (Program Committee), 包括IJCAI、AAAI、KDD、ICDM、SDM、CIKM等。