Barnard, K. & Johnson, M. (2005). Word sense disambiguation with pictures. Artificial Intelligence, 167, 13-30.

@Article{BarnardJohnson2005,
  author = 	 {Barnard, Kobus and Johnson, Matthew},
  title = 	 {Word sense disambiguation with pictures},
  journal = 	 {Artificial Intelligence},
  year = 	 {2005},
  key = 	 {},
  volume = 	 {167},
  pages = 	 {13--30},
  note = 	 {},
  annote = 	 {}  
}

Author of the summary: Jeanne-Marie Musca, 2009, jmusca@gmail.com

Cite this paper for:

• Using visual information is more effective than using textual information for word sense disambiguation. [26-27]
• Using both sources of information is better than using either one alone [26-27]

Word disambiguation is essential to natural language processing. Context gives the information necessary for disambiguation.
The typical approach to word meaning disambiguation consists of looking at surrounding words. [p.14]

Proposed alternative: use image information in order to disambiguate. [15]

2. Disambiguating using textual context

Several methods have been used to disambiguate by looking at surrounding text:

• Using semantic networks which include sets of synonyms and "is a" relationships between these sets.
• Looking at usage frequency
• Incorporating world knowledge from multiple sources [16]

An image is segmented into regions.
The model predicts, using probabilities, which words will be associated with a region based on several attributes such as colour and shape. [17-19]

4. Using word prediction for sense disambiguation

The vocabulary used in the model is disambiguated (bank: bank_1, bank_2,...). So when the model predicts the word associated with a particular region, this word is disambiguated.[20]

This method and the traditional method (which uses textual context) can be combined for better results.[20- 23]

5. ImCor

ImCor links Corel images with SemCor disambiguated text.[23]
This results in a corpus with 20,153 image/text pairings which can be used in experiments [24].

6. Experiments

Only documents with at least one ambiguous term were used to calculate performance.
Combining both images and textual information for disambiguating word sense gave promising results.[25]

7. Conclusion

Using images to disambiguate word sense gives better results than using textual context.
Combining both gives even better results.
Imcor is a useful data set for further exploring this area.[28]

Summary author's notes:

• none