Help with determining window size and min count

[vondiplo]

I'm trying train a model that would include certain topics.
Relying on the default parameters somehow keeps the topic out of the model.
I was thinking of changing the window size to 5 and the min count to 5 to get more granular results. However, I don't seem to actually know what would be the effect of changing these parameters. Could someone please shed some light regarding the impact ?

[dav009]

• Have you tried stemming the vocabulary as Phillip suggested?
• Are you sure the given topic has enough links on wikipedia ? You can probably check dbpedia NLP datasets to check how many times it appears

[vondiplo]

• I did try stemming the vocabulary but there are times were automatically getting other descriptive concepts aren't obvious
• I'm not sure how popular the topic is on wikipedia, but I do know it was returned by dbpedia-spotlight (https://github.com/dbpedia-spotlight/dbpedia-spotlight/). If they could recognise the topic using their model why wouldn't my trained model be able to do so ?

[dav009]

can you add some info on how many annotations do the topics you are looking for have?
I think the stats on annotations is available here: http://spotlight.sztaki.hu/downloads/latest_data/

presumably if it is returnable by spotlight it means it has enough counts

[vondiplo]

@dav009 - It doesn't appear there at all (I've downloaded the english tar, then performed a grap on the unzipped folder with my searched concept, 'visual_cortex'). Yet on the other hand neither does Barak_Obama, but it surely does appear in both spotlight's annotations and in wik2vec vectors.

[vondiplo]

Hi @dav009, is there any update regarding this?

[dav009]

Sorry for the late reply,

• Barack_Obama(note your typo) is for sure there.
• if you get the URI counts out of that 1G tar.gz, I would take a look, Im currently under not so good internet.