wfeely / hw5-team05

For class project homework 5
0 stars 0 forks source link

Improve Candidate Selection #17

Closed wfeely closed 10 years ago

wfeely commented 10 years ago

Keerthi should also be assigned.

Improve selection of candidate answers. Erzhuo will improve the candidate selection and test this improvement on his branch. Keerthi will run the system for Erzhuo's improvement.

ernestwang commented 10 years ago

Guys, do we meet at class today?

On Mon, Nov 11, 2013 at 5:09 PM, Wes Feely notifications@github.com wrote:

Keerthi should also be assigned.

Improve selection of candidate answers.

— Reply to this email directly or view it on GitHubhttps://github.com/wfeely/hw5-team05/issues/17 .

Erzhuo(Ernie) WANG University of Washington, Seattle Department of BioEngineering Email address: erzhuo.wang88@gmail.com Phone: (206)673-6527

汪尔卓 华盛顿大学,西雅图 生物工程系

ernestwang commented 10 years ago

Keerthi and all, I took a look at the searching results and the raw material this afternoon. I noticed that the current candidate answer selection algorithm can hardly find the real relevant sentences. Partial reason is that it simply tries to match the question to the sentences, regardless of the option answers. For example, text 1 question 5: "Which CLU protein isoforms in the brain have been characterized?" and the correct answer is "rs11136000". Whereas the questions retrieved don't mention this protein at all. So I suggest that we try to search combinations of questions and answers. The drawback is that noise will be introduced, the candidate answers are certain to match each of the options. Thus new similarity metric is to be used. Any ideas?

wfeely commented 10 years ago

I've re-assigned this to Siping, because he's mentioned that he has ideas on how to improve candidate selection. We will need to get our Verb Phrase matching working first, and the use those to improve candidate selection.

keerthi166 commented 10 years ago

Moved to Milestone 3

spininertia commented 10 years ago

I look at the wiki page and found there are something wrong about the configuration. The only component in the pipeline that uses the background corpus as a source is PMI scorer. I'll update the wiki page tomorrow before class.

2013/11/24 Keerthiram Murugesan notifications@github.com

Moved to Milestone 3

— Reply to this email directly or view it on GitHubhttps://github.com/wfeely/hw5-team05/issues/17#issuecomment-29177232 .

wfeely commented 10 years ago

Erzhuo has investigated improving our candidate selection and found that adding more candidates doesn't help. Closing this for now.

wfeely commented 10 years ago

Nevermind, Erzhuo has another experiment he wants to do on his branch.