The Associated Press owns over 30 million historical and contemporary photographs, which makes searching for the ideal image a challenge for both internal and external editors. At the time of their request, editors had to manually input subjective criteria for image search. The objective of our capstone project is to replace the manual process with an automated image recommendation system.
In building the recommendation system, we explored 2 overall approaches:
1. Tag-to-Tag Recommendation: leveraging AP’s existing tagging system, we ranked tag frequency and relevancy based on article content.
2. Text-to-Text Recommendation: giving article text as raw input, we mapped word embeddings of the input text to those of images.
We tested our models with both provided and internet-sourced data, and added a third approach, which is an ensemble of all models explored in the scheme of above approaches. We featured 2 Text-to-Text Recommendation models investigated in this article. To learn more about our project, check out our poster below:
Karina Huang 
You must be logged in to post a comment.