Institute of Information Science, Academia Sinica

Conventional approaches to relevance ranking typically optimize ranking models by each query separately. However, we find that for those queries related to a certain topic, web pages of some websites would be consistently ranked higher than web pages of the other websites. For example, there is a great consistency between web pages on ESPN and IMDb for queries about different athletes, i.e., ESPN>IMDb. In addition to web search, we also find that semantic tagging of mathematical expressions (STME) also has the problem of "tagging consistency," which is the major difference between STME and POS tagging, e.g., a variable usually denotes the same semantic tag in a math expression while a word might have different POS tags in a sentence. In the first part of this talk, we will address the issue of consistency and present several methods to enhance web search and STME in both of supervised and unsupervised manners.

Knowledge graph (KG) benefits various valuable IR applications. To model multi-relational types such as many-to-many relationships in KG, conventional approaches are usually devoted to mapping two entity embeddings into a low-dimensional hyperspace, and then differentiating their relation through a linear translation. Most of them assume relations are independent from each other. In the second part of this talk, we will talk about how to learn to effectively select the most influential neighbors with an attention mechanism and dynamically incorporate these neighbors for better relation inference. We will further make use of contextual information to learn unsupervised document representation and apply it to the problem of fine-grained aspect-based sentiment analysis.

Pu-Jen Cheng received his Ph.D. in Computer Science at National Chiao-Tung University in 2001. He went to the Institute of Information Science, Academia Sinica as a Postdoctoral Fellow for more than four years. Starting from August 2006, he joined the department of Computer Science and Information Engineering faculty at the National Taiwan University, and also jointly appointed as faculty of the Graduate Institute of Networking and Multimedia, National Taiwan University. Prof. Cheng's research interests include information retrieval, machine learning and text mining. He has authorized many papers published on premier journals and conference proceedings such as ACM SIGIR, WWW, CIKM and JASIST. He usually served as a committee member in IR and NLP related international conferences such as a PC member of SIGIR, ACL, AAAI and the conference chair of 2010 Asia Information Retrieval Society Conference (AIRS)and the conference chair of 2017 Conference on Technologies and Applications of Artificial Intelligence (TAAI). He was a recipient of Google research award 2007, Microsoft research awards 2008 and 2016, and ACM-ICPC coach award 2012.