What is it about?
PKG is a novel system forintegrating the data of a user from different sources. We show how a user’s intention can be detected and how the personal data can be aligned and connected by the user behaviors. The constructed PKG allows the system makes reasonable and accurate recommendations for users by a “neural + symbolic” approach across different services.
Featured Image
Photo by Masakaze Kawakami on Unsplash
Why is it important?
Mobile internet users generate personal data on the devices all the time in this era. In this paper, we demonstrate a novel system for integrating the data of a user from different sources into a PKG. A user’s operations, i.e., intention entities centered around a user with links to other common entities, are captured and integrated into a PKG based on a pre-defined ontology. Then, given mobile using scenarios, the PKG is used to analyze the user complex intention which varies time scales and evolves over time. Finally, accurate recommendations are generated by a “neural + symbolic” approach based on the PKG across different services.
Perspectives
I hope this article will give people some new ideas about using knowledge graphs to store personal information for recommendation. Because this is an area worth exploring, there are limitations that need to be addressed and overcome, and I will continue to explore in this area. Most importantly, I hope you find this article thought-provoking.
Yu Yang
Southeast University
My work mainly involves the design and implementation of the "neural + symbol" information extraction model and related data set in the paper. In terms of datasets, we extracted user corpus information from several open source daily conversation data sets, selected available types of text and marked relevant entity locations through the method of rule matching + scoring. In terms of the model, we design an end-to-end multi-task joint information extraction model, which has two subtasks: intention type classification and entity extraction, in which entity information obtained from entity extraction subtask will be used together for intention type classification, improving the performance of this subtask. We also further discuss the differences between Pipeline model and Joint joint extraction model in training time, model space occupied and model performance. The joint extraction model with multi-task learning and additional entity information finally achieves the optimal performance.
Jiangxu Lin
Southeast University
Read the Original
This page is a summary of: PKG, July 2022, ACM (Association for Computing Machinery),
DOI: 10.1145/3477495.3531671.
You can read the full text:
Resources
Contributors
The following have contributed to this page
