Philgoo Han writeup of Bunescu and Mooney
From Cohen Courses
Jump to navigationJump to searchThis is a review of Bunescu_2007_learning_to_extract_relations_from_the_web_using_minimal_supervision by user:Ironfoot.
- Relation extraction with minimal supervision, Tested on general data (web searched + noisy data)
- Based on a very simple but powerful observation.
- "Although not all of the sentences for positive pairs will state the desired relationship, many of them will. Presumably, none of the sentences for negative pairs state the targeted relation"
- MILs tranform to a standard supervised learning problem(Ray and Craven, 2005)
- A positive bag garuntees at least one positive instance -> At most all but one may be false positive label ... does this really work?
- MIL and two types of bias
- Type1: words correlated with the argument itself but not with the relation between the arguments.
- Overcome with adjusted weight in subsequent kernels
- Type2: words describing only the relation which doest not contain any information about the two arguments.
- Author still working on
- Type1: words correlated with the argument itself but not with the relation between the arguments.
- Competative results of MIL with type1 bias reduced
- MIL and SIL were trained and tested on different constraints which makes the descussion weak
- SSK-SIL should have penalty of requiring great effort on labeling data. Anyway to measure this?
