Probably you should weight the categories. Find all the matching words, and assign a value to all categories as follows:
- Add 3 for every word that is undoubtfully belongs to that category
- Add 1 for every word that may belong to more categories
It is a biased weighting (towards unique words), this way you can better decide where the pictures belong to.
Also, you can build a - continuously changing - weight-matrix, that which word is how relevant to a certain category. The frequent words bear less importance (because everybody is using them).
Also, based on the categorized texts, you can automatically extend the word-list, and automatically categorizing them. For example, if a new game name appears in the word-list (call it 'abc'), you will notice that 'abc' appears in a lot of texts in the hobby category, and nowhere else. So, you can tie this word to this category.
It's a very exciting area to build auto-learning systems!