Our hypothesis is that graph retrieval should take into account the relevance of
the structural and message components of an infographic to the requirements
conveyed by the user's query.
We consider three mixture models which respectively capture structural relevance, message relevance, and both structural and message relevance. Since
the results of query processing are not always correct, we add to each model a
back-off relevance measurement R(Qt;Gt) which measures the relevance of all
the words in the query to all the words in a candidate infographic. In addition,
we include a baseline model that treats the words in the graphic and the words
in the query as two bags of words and measures their relevance to one another.