Knowledge triple mining via multi-task learning

Recent years have witnessed the rapid development of knowledge bases (KBs) such as WordNet, Yago and DBpedia, which are useful resources in AI-related applications. However, most of the existing KBs are suffering from incompleteness and manually adding knowledge into KBs is inefficient. Therefore, automatically mining knowledge becomes a critical issue. To this end, in this paper, we propose to develop a model (S² AMT) to extract knowledge triples, such as <Barack Obama, wife, Michelle Obama> from the Internet and add them to KBs to support many downstream applications. Particularly, because the seed instances¹ for every relation is difficult to obtain, our model is capable of mining knowledge triples with limited available seed instances. To be more specific, we treat the knowledge triple mining task for each relation as a single task and use multi-task learning (MTL) algorithms to solve the problem, because MTL algorithms can often get better results than single-task learning (STL) ones with limited training data. Moreover, since finding proper task groups is a fatal problem in MTL which can directly influences the final results, we adopt a clustering algorithm to find proper task groups to further improve the performance. Finally, we conduct extensive experiments on real-world data sets and the experimental results clearly validate the performance of our MTL algorithms against STL ones.