Research

Publication

“Ensuring Readability and Data-fidelity Using Head-modifier Templates in Deep Type Description Generation”. Jiangjie Chen, Ao Wang, Haiyun Jiang, Suo Feng, Chenguang Li and Yanghua Xiao. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), 2036-2046.
“CN-Probase: A Data-driven Approach for Large-scale Chinese Taxonomy Construction”. Jindong Chen, Ao Wang, Jiangjie Chen, Yanghua Xiao, Zhendong Chu, Jingping Liu, Jiaqing Liang, and Wei Wang. In IEEE 35th International Conference on Data Engineering (ICDE), 1706-1709.

Objective: Based on the hypernym-hyponym relations extracted from multiple sources, we built CN-Probase, a Chinese taxonomy that contains 17 million entities, 270 thousand concepts and 33 million isA relations with a precision over 95%. It is widely used by many renowned enterprises for different applications.
Core Contents:
- Carried out extraction of isA relations from multiple sources along with the effective integration
- Added conflict detection and noise canceling module by using statistic methods
- Joined in the construction of CN-Probase and the composition of the paper as second author

Objective: Inspired by the representational capacity of Markov logic network (MLN), the aims are to figure out an effective approach to learn the contributions of features towards concept in categorization, so as to make interpretable and convincing categorization for entities in ontological knowledge bases.
Core Contents:
- Proposed multiple methods for first-order Horn clause mining
- Designd a heuristic sampling algorithm to solve the scalability problem of MLN
- Applied Markov logic network to learn weights of rules to acquire interpretability
- Demonstrated better performance in categorization than other classical and state-of-the-art models

Objective: Different from conventional methods of building a complete taxonomy for commonsense reasoning, we brought about the idea of online commonsense proposition determination by building a commonsense oracle based on information from knowledge graph and search engine.
Core Contents:
- Carried out online commonsense inference by extended feature dimensions
- Applied graph-based supporting facts from knowledge graph
- Applied search engine to acquire web-based facts with pre-defined lexical patterns and templates