Question --> computing relevance VS reconstruction error
drwi opened this issue · comments
Dear Sir,
This really is not an issue, but rather a technical question / to better understand the scientific choices behind the code you proposed.
I wonder whether there is a reason why the whole document reconstruction error is computed as [the sum of residuals between words and reconstructed words]
whereas the relevance is computed as
[the average cosine similarity between summary words and reconstructed summary words].
Is it mainly to ease numerically solving for topic extraction (computing the best H) whereas relevance should not depend on the norm of {word vectors} (or should it )? Would you have any clue that would enable to underpin this choice?
Have a great day,
Yours sincerely,
Drwi