🧑‍🔬 Liheng Xu

Liheng Xu works in natural language processing, specifically on getting structured knowledge out of text and representing it for machines. Their papers here address two connected problems in information extraction and knowledge graphs: filling gaps in large knowledge bases by predicting missing facts, and pulling events and their participants from raw sentences. In both, hand-designed features and external parsing tools give way to vector representations and neural networks learned directly from data.

The most cited of these papers is Knowledge Graph Embedding via Dynamic Mapping Matrix (2015). It introduces TransD, a translation-based model that represents each entity and each relation with two vectors, one for meaning and one for projection. Each entity–relation pair then gets its own mapping matrix, where earlier models made all entities of a relation share one. This captures how varied those entities can be and uses fewer parameters and less computation than TransR/CTransR. The second paper, Event Extraction via Dynamic Multi-Pooling Convolutional Neural Networks (2015), handles ACE event extraction with a convolutional network. The network splits each sentence at the trigger and argument positions and pools each segment separately, which keeps local evidence that a single max-pool would lose. Embeddings of each word's distance to the target words tell the model where to focus. Triggers are detected first, and a similar network then labels the arguments.

Both papers date from 2015 and share an approach: learned representations replace pipelines built on parsers and manual feature engineering. The current affiliation is China Jiliang University in Hangzhou. The work collected here is tied to the Institute of Automation, Chinese Academy of Sciences, in Beijing, where it was produced with Jun Zhao and Kang Liu and with Shizhu He, Daojian Zeng and Yubo Chen.

Liheng Xu has 2 papers in our library, in 2015, most of them on Information extraction, Knowledge graphs & ontologies; often with Jun Zhao, Kang Liu, Daojian Zeng.