A key focus of this work was to recognize implicit emotions in blog posts.
Different from past approaches using surface features, we utilize syntax and semantic knowledges and learn meaningful patterns to detect a user's intended emotion.
Extract Patterns from LJ40K
Construct Lexicon
Pattern Scoring
Document Scoring
naive approache
machine learning
include
feature extraction
position feature (position_feature.py)
[options]
-b: percentage of beginning section
-m: percentage of middle section
-e: percentage of ending section
-c: counting unit for document segmentation
0: number of words
1: number of sentences (not implemented yet)
-f: feature value computation
0: pattern scores (patscore_p2_s0)
1: accumulated threshold by 0.68 (1 std) using pattern scores
2: accumulated threshold by 0.68 (1 std) using pattern count
3: [type 2] & min_count=4
4: [type 2] & remove the pattern occurrence counted from oneself (for ldocID 0-799)
5: [type 3] & remove the pattern occurrence counted from oneself (for ldocID 0-799)
-v, --verbose: show debug message
pattern feature (pattern_feature.py)
[options]
-l, --limit: minimum occurrence of a pattern
0: (default) collect all patterns
n: at least occurs < n > times for each pattern
-v, --verbose: show debug message
[options]
-k: keyword set in WordNetAffect
0: basic
1: extend
--lemma: use word lemma when looking for keywords
-v, --verbose: show debug message
training
testing
Evaluation
naive document scoring
machine learning
git clone git@github.com:AcademiaSinicaNLPLab/emotion-detection-modules.git
// check (un)tracked, (un)staged files
git st
// add files for this commit
git add <files>
git ci -m "add a new line blah..."
// push commit to remote
git push -u origin master // first time using push
git push
// if cannot push to remote (fast forward), pull first
git pull -u origin master // first time using pull
git pull