Research projects
Event Extraction : Can computers learn to read news articles and understand real-world events?
I developed a joint inference approach for extracting semantic representations of events from text. It can effectively identifies not only the anchors of events but also the related entities (participants, times, and places) and their semantic roles. By making joint predictions about these elements across a document context, my approach significantly outperforms the state-of-the-art event extraction approaches [NAACL2016].
Figure 1: Examples of the system extracted event anchors (or triggers), entities, and the participant roles of the entities. The text spans in green indicate the head words of the extracted event triggers; the text spans in blue indicate the head words of the extracted entity mentions, and the labeled arrows indicate the participant roles of the entities.
Event Coreference : Can computers learn to aggregate event information across multiple sources?
I developed a novel Bayesian clustering model for event coreference resolution both within a document and across multiple documents. It successfully combines linguistic intuitions with the advances of Bayesian statistics, resulting in significant improvement over the existing clustering approaches [TACL2015].
Figure 2(a): Distance-dependent Bayesian Clustering.
Figure 2(b) : Examples of the system extracted coreferential events.
Knowledge Representation and Reasoning : How to represent entities and relations in large-scale knowledge bases in order to support efficient and effective inference and reasoning?
I developed a scalable neural network model for learning continuous representations of entities and relations in large knowledge bases like Freebase. I showed that the learned representations lead to accurate prediction of unseen relations [LS2014]. Moreover, I showed a novel use of the learned representations in mining horn clauses (e.g., BornInCity(a,b) and CityInCountry(b,c) => Nationality(a,c)) from data [ICLR2015].
Figure 3(a): A neural network framework for multi-relational learning. The input data is in the form of tripples (entity_1, relation, entity_2). The model outputs are low-dimensional continuous representations for entities and relations.
Figure 3(b): Horn clauses inferred by conducting simple multiplicative operations on the learned representations.
Opinion Extraction : Can computers learn to recognize and interpret opinions, beliefs, and sentiment from text?
I developed structured learning and inference approaches for extracting semantic representations of opinions from text. Unlike existing work, they allow joint interpretation of opinion attributes (i.e., polarity and intensity), the holders (who is giving the opinion), and the targets (what the opinion is about), and produce state-of-the-art performance on the fine-grained opinion extraction tasks [ACL2013, EMNLP2012, TACL2014].
Figure 4: Examples of the system extracted opinions. The highlighted text spans indicate the extracted opinion expressions, opinion holders and opinion targets, and the labeled arrows indicate their association relations. The green color indicates positive polarity, red indicates negative polarity, orange indicates neutrality, and ★ indicates intensity (more stars means higher intensity).
Mining Massive Structured Data : What can we learn from large-scale mobile user networks? Can we predict when customers are gonna churn? Can we customize mobile resouce allocations based on traffic patterns?
I was the lead developer (from 2007-2010) of MobileMiner, a data mining system for data analytics in mobile communication. It includes a wide range of data mining algorithms and supports applications such as social community analysis, churn prediction, and traffic pattern mining [SIGMOD2009, PAKDD2008]. I implemented the distributed versions of the algorithms in Hadoop and deployed them in China Mobile's Cloud computing platform to processing multi-terabyte data.
Figure 5: User clustering based on social interactions (left) and user clustering based on moving patterns (right).