Neural Question Generation for Reading Comprehension
Deep Learning for Fine-grained Opinion Extraction
Research intern at Cornell NLP group, supervised by Prof. Claire Cardie.
Proposed to evaluate labeling sequences using sentence-level log-likelihood (SLL) at output layer of deep recurrent neural networks. Empirical results showed around 4% improvement on F-measure of opinion target extraction and more accurate n-best ranking for labeling sequences.
Proposed to use deep recurrent neural networks to produce the n-best labeling sequences which can be fed into integer linear programming (ILP) system for joint inference.
Designed heuristic rules using dependency parse tree to eliminate inappropriate opinion candidates during inference. Empirical results showed higher precision and higher F-measure.
Building an ensemble system using the above algorithm to participate in Belief and Sentiment Evaluation 2015 (BeSt2015).
Online Auction Mechanism Design with Time-varying Value
Research assistant at Advanced Network Laboratory, Shanghai Jiao Tong University
Proposed the new scenario in online auction mechanism design where agent’s value may vary over time.
Extended the classic payment determination algorithm (Myerson) to fit the new model. Proposed mechansim ensured strategy-proofness and achieved constant competitive ratio.