DOCUMENT CLASSIFICATION WITH SVM

1
DOCUMENT CLASSIFICATIONWITH SVM

• Studies in Computational Linguistics IIOpinion
  Mining and Sentiment Analysis
• Hwang Inbeom

2
Overview

• Considered only unigrams as features
• Score of each feature is evaluated using
  frequency of corresponding unigram
• Implemented with Ruby MySQL

3
Implementation

• Ruby
• http//www.ruby-lang.org/ko/
• Simple and productive
• SQL
• http//www.w3schools.com/sql/default.asp
• Ruby/MySQL connector
• http//www.tmtm.org/en/ruby/mysql/

4
Implementation ( contd.)

• Ruby examples

check if hello exists in table tf my
Mysqlnew(localhost, inbeom,
inbeom, inbeom) res
my.query(SELECT FROM tf
WHERE word hello) if
res.num_rows gt 0 puts hello exists in the
table! end
print all words in a file File.foreach(filename
) do l words l.split words.each do
w puts w end end
5
Dataset

• Bo Pangs polarity dataset v2.0
• 1000 positive and 1000 negative movie reviews
• Plain text format

films adapted from comic books have had plenty of
success , whether they're about superheroes (
batman , superman , spawn ) , or geared toward
kids ( casper ) or the arthouse crowd ( ghost
world ) , but there's never really been a comic
book like from hell before . for starters , it
was created by alan moore ( and eddie campbell )
, who brought the medium to a whole new level in
the mid '80s with a 12-part series called the
watchmen .
6
Stop Word Elimination

• List of stop words could be found on the web
• http//www.lextek.com/manuals/onix/stopwords1.html
  
• Unigrams in this list are excluded in feature
  evaluation process

7
Unigram Frequency

• Counted the number of appearance of each unigram
  in both positive and negative document sets
• Number of entries was over 45,000

8
Unigram Frequency Implementation

• Algorithm
• create an empty table in the previous slide
• while(there is a document d unprocessed)
• for every word w in d
• insert w into table if w is not inserted yet
• update a table row which word is w
• pos pos 1 if d is in positive set
• neg neg 1 if d is in negative set

9
Distribution of Unigram Frequency

• Excluded unigrams occurred less than 5 times,
  occurred only in positive or negative set, and
  total occurring count is more than 2000 times
• 12,830 unigrams used as features

10
Distribution of Unigram Frequency

• Unigrams occurred more in positive set
• Several name of someone or movie titles ranked
  very highly
• Mulan, Flynt, Lebowski, Jude, Winslet, Homer,
• Other positive words
• Hatred, whisperer, astounding, exotica,
  fascination,
• Unigrams occurred more in negative set
• Seagal, Jawbreaker, Jakob, Hudson, magoo,

11
Method 1 Unigram Presence

• Set the score(u,d) of unigram u as 1 if u is in
  the document d
• Implemented with hash table
• Algorithm for each document d,
• for every word w in document d
• If w is not a stop word
• hd, w 1
• print w1hw1, w2hw2,

12
Method 2 Unigram Frequency

• Presence score of each feature is multiplied by
  its number of occurrings in a document d
• hd, w occurringd, w

13
Method 3 Thresholding

• Cut off frequent and rare unigrams
• 2000 gt number of occurrings gt 5
• About 13,000 unigrams remained
• Applied to former two methods
• Thresholding presence
• Thresholding frequency

14
Method 4 Scoring

• Assigned base score to each unigram
• Feature score is evaluated by multiplying unigram
  frequency to base score

15
Base Scoring Method

• Base score function of a unigram

16
Feature Score

• Base score is multiplied by number of appearances
  of each unigram in a document
• Could be improved by applying a smoothing function

17
Evaluation Environment

• Training set
• 900 documents from both sets
• Unigram frequencies are counted in this set
• Test set
• 100 documents from both sets

18
Evaluation Results
19
????? )