The School of ITEE is hosting the following PhD thesis review seminar:
Towards Practical Neural Entity Alignment
Speaker: Bing Liu
Host: Dr Mahsa Baktashmotlagh
Abstract: Knowledge Graphs (KGs) have been widely used to drive downstream applications but always suffer from incompleteness. Fusing different KGs is believed to be a promising direction for producing a more comprehensive KG. Entity Alignment (EA), which aims to identify equivalent entities across different KGs, is a primary step of KG fusion. Though many neural models have been explored, some critical problems still prevent the neural EA models from being deployed in practical applications: expensive annotation cost, poor scalability, and bottleneck in effectiveness.
In this thesis, we aim to go further towards practical EA. We found neural architectures have attracted intensive attention, whereas the annotation cost, scalability issue, and training guidance are still under-explored. Therefore, we fill these gaps to complement the neural EA studies. Our contributions include: (1) To reduce the annotation cost, we devise an Active Learning framework for EA, which aims to select the most informative entities to annotate. Our data sampling strategy can combine the uncertainty of the neural model and the graph structure of KG. Also, it can identify unmatchable entities by overcoming a sampling bias problem. (2) We propose a high-quality task division framework to adapt the existing neural EA models to large-scale KGs. The built subtasks can avoid raising memory issues and save time consumption by running in parallel. (3) To break the bottleneck in effectiveness, we seek help from other training guidance -- label-level compatibility -- apart from the labelled data. Our designed framework can bridge the gap between the reasoning-based and neural EA methods. (4) Furthermore, we pay attention to the pseudo-supervision produced by self-training. Specifically, we exploit the dependencies between entities to improve the quality of pseudo-labelled data.
Our research has expanded the landscape of neural EA and achieved remarkable improvement in the annotation cost, scalability, and effectiveness aspects.
Speaker Bio: Bing Liu obtained his Bachelor of Science and Master of Engineering degrees both from the Southeast University, China. He is currently a PhD candidate in the School of ITEE supervised by Dr. Wen Hua and A/Prof Guido Zuccon. His research topic is about the construction and application of Knowledge Graph.
About Data Science Seminar
This seminar series is hosted by EECS Data Science.
Venue
Online via Zoom: https://uqz.zoom.us/j/3084973358
Room:
Multimedia Lab 78-632