October 30 , 2009

elbowcheepΤεχνίτη Νοημοσύνη και Ρομποτική

15 Οκτ 2013 (πριν από 4 χρόνια και 28 μέρες)

86 εμφανίσεις

October 30
th
, 2009


Sandra:


1)
What’s the timeline?

2)
Any existing tools you can use to implement EM
-
NB and S3VM?

Elin:


1 )
What’s your justifications for the strategies you designed for creating co
-
training
classifiers?(put it in the literature review

to show that they are not ad hoc, but have been
investigated somehow, maybe in other domains).

2)
Sentence labeling by yourself and anyone else? (do not need to go through human subject
committee for extra judge) (Kiduk: There are already some implicit e
valuations in it since
the sentences to label is not from some random sample, but from blog training data, which
are already reviewed by two TREC judges at the document level)


Ying:


1)
W
hy NB, SVM, why not others? SVM kernel selection? Transformed SVM? C5
.4? Latent
Dirichlet?

Kiduk:

1)


I
f the experiments didn’t yield good results, where to look into

to improve the

performance? Domain adaptation(the most interesting question to me)
;

2)

Main different between lexicon based and machine learning based approach is
that the
former is essentially rule
-
based while the latter is based on some model.

3)

How to improve opinion detection in blogosphere
-
> SL doesn’t work, period. Why
-
> noisy
data, short of labeled data(a big problem)
-
> use unlabeled data


The dissertation s
hould be have a tailored introduction which focus on the challenge of the lack
of opinion labeled data and the significant of it. Literature review should cover SSL and SL and
compare their algorithms to provide background of why I choose certain methods f
or my
experiments.