There are some very interesting techniques in the linked paper, particularly the moving average alternative to bag of words and the use of RankSVM.<p>But this generated sample (verbatim) isn't rap, it's just rhyming words:<p><pre><code> You've been a survivor and we love you so much
You don't know me - I kill everyone who show love
I I know you all love me to be sad and shit but
Will lead you to peace to love to unity and havin' fun
And say I love you momma and you are not alone
To say I love you! I love you! Forever so!
I say I love you you say it don't show
Tell me you love me I love you baby oh oh
Oh baby I love you
Can I love you baby?
What we about to do
''HOW COULD YOU JUST LEAVE ME AND LOVE HIM OUT THE BLUE?!''
But your on it too so you tell me you love me too
Let your love home you Let me move you tonight Baby
I know you love it when I shine I'm waitin' ..
And I think I love him I love him just like I raised him
I love rap no matter how much I say I hate it
</code></pre>
Contrary to popular belief, rap lyrics do contain insight and capture cultural moments. The gap there relates to the question of whether or not computer programs can be 'creative'. This line from the linked paper also sheds light on the quality of these lyrics:<p>"In terms of quantitative rhyme density, the produced lyrics outperform best human rappers by 21%"<p>I'm not sure if sheer rhyme density is how rap or any other form of music should be judged, but if it is then human rappers are only 79% as good as the best machine MCs.
The Actual Thesis: <a href="http://arxiv.org/pdf/1505.04771v1.pdf" rel="nofollow">http://arxiv.org/pdf/1505.04771v1.pdf</a>