VITALAS at the University of Sunderland

1
VITALAS at the University of Sunderland

• Michael Oakes, Marco Palomino and Yan Xu

2
Example of an automatically-selected concept
vocabulary

• soccer (soccer, soccers, soccere) 27278.65
• pictur (picture, pictures, pictured, picturing,
  pictur, picturs, pictureds) 22565.06
• minist (minister, ministers, ministe, minist,
  ministeer, ministes, ministered) 21035.98
• team (team, teams, teamed, teaming) 18273.36
• cup (cup, cups, cupping, cupped) 18187.04
• citi (city, cities, citi, citys, citis, citie)
  17369.72
• leagu (league, leagues, leagu, leagus) 16693.42
• celebr (celebrates, celebrate, celebrations,
  celebration, celebrated, celebrating,
  celebrities, celebrity, celebrants, celebrer,
  celebre, celebrators, celebrateing, celebreated,
  celebreates, celebres, celebrational, celebral,
  celebratings, celebrant) 16545.52
• won (won, wons) 16436.22
• championship (championships, championship)
  16256.48

3
Statistical methods used

• Derived from 10,000 Belga captions in each case.
• Chi-squared test, with Europarl as the reference
  corpus.
• Information Radius, with Europarl as the
  reference corpus
• Raw frequency
• PageRank the most important words occur in the
  same captions as other important words.

4
Using search logs to make recommendations to
future users
Text query
Search Engine
Associated keywords
Images chosen by past users
Match
Search logs
User
Ranked images
5
Using TF.IDF to index the images in the search
logs

• .

• wkd is the weight reflecting the typicality of
  term k with respect to image d
• fkd is the frequency term k was used by searchers
  looking for image d
• NDoc is the total number of images downloaded in
  the search logs
• Dk is the number of images which which were
  searched for using term k at least once.
• The highest TF.IDF scores are given to those
  terms which were frequently used when searching
  for a given image, but were rarely used when
  searching for other images.