Sundong Kim The personal website of Sundong Kim

Working on BBox-2015

Things have done

I have been modified B-Box throughout this week.

  • Migration from Old, Messy B-Box2015, I used branch 'BBoxAugust2015' and Jaesung used branch 'Complete' before.
  • Add code to compare inferred type to original type (If exist in DBPedia:ko), if original type result is firm enough, then we don't spend time to load triple and infer its type.
  • 2nd year Schema-based (Which domain should we take?)
    • DBO-Prop (English-defined prop) Domain: Just crawled related domain defined in DBPedia2014.
    • If Prop:ko-DBO mapping is available (By SWRC), we used that handcrafted information, and used pre-defined domain information explained above.
    • For Prop:ko which doesn't have DBO mapping information, and its domain: We referred PROPDI result. Now, we can use inferred domain of domain-unknown prop:ko.
  • Integrate my code & Jaesung's code (Hardest part). Previously, we put all out code into one class, so I separated task by task and repackaged stuff.
  • Package in maven so that everyone can use it easily.

Things to do

  • Another Migration to Official repo
  • Slightly more optimization (Not comparing if original answer is good)
  • Check with WP6 pipeline and I/O (Which isn't well decided)
  • Predicate equivalence will not be applied in our current pipeline, so new B-Box repo only contains type-inferencing part.


  • 가장 중요한 점은 다음과 같다. Instance의 evidence 트리플이 1-2개밖에 없는 상황에서 타입 추론을 하라는 무리수가 많음... L2K 데이터만 봐도, subject 하나에 관련된 트리플이 1-2개밖에 없기 때문에, 우리가 그 트리플을 보아도 완벽한 타입을 추론하기 어렵다. 그런데 그걸 바탕으로 평가를 해서, B-Box 타입추론기의 성능을 평가하는 건 무리가 있다고 생각된다.

Some results러시아 2nd: {,,,} 3rd: {,,,,,,,} Aggregated: {,,,,,,,} BBox result: {} Existing best type:

comments powered by Disqus