• move keyword threshold from FFNN to Benchmark
• make FFNN return a list of probabilities instead of ones and zeros
• create a wrapper module ste-ml.js that can be initialised with only a modelType (svm or ffnn atm), a modelPath (path to trained exported .model) and a list of stemmedUniqueTerms and that behaves similar to ste-statistical.js

In this case, the name of the Spanish king's brother-in-law Iñaki Urdangarin was not matched with the image keyword Inaki Urdangarin because the original term was tokenized to I aki Urdangarin. Make sure the tokenizer's split heuristic is aware of such special chars.

Use RegexpTokenizer instead of WordTokenizer:

This system is designed for assigning news photography to articles, but Getty's database also includes illustration and infographics. Both go beyond the scope of this system and should therefore be excluded from search results.

See Node API:

https://github.com/gettyimages/gettyimages-api_nodejs/blob/82060b68003ab6be5f37fa12edb32321886514dc/lib/searchimages.js#L129-L132

and API Docs:

http://developers.gettyimages.com/docs/#operation/Search_GetImagesByPhrase

(&graphical_styles=illustration&graphical_styles_filter_type=exclude)

This function returns all the preprocessed terms in an article. Add an argument that specifies if it should strictly return all terms or if it should exclude stopwords.

E.g.

 Preprocessor.prototype.getStemmedTerms = function (sortFunction, excludeStopwords) { 
    ... 
         if (excludeStopwords) {
             terms = terms.filter(term => stopwords.indexOf(term) < 0)
         }
         if (sortFunction) { 
             return terms.sort(sortFunction) 
         } 
    ...
 }

SportsLeague should only be listed in "Event", not in "Product"