I Forged Emily Dickinson’s Poetry with an RNN

DATA SCIENCE | MACHINE LEARNING

The end result is the fourth worst poetry in the universe!

Feb 23 ·3min read

I utilized a Recurrent Neural Network trained on the collective works of Emily Dickinson to generate new poetry. It’s not quite as bad as Vogon poetry , but it’s not quite as good as the real deal either.

iQBFbi2.jpg!web

Profound.

In order to make this happen, I repurposed a demonstration available here on Google Colaboratory and added a few tweaks of my own.

The interesting thing to note is that the network predicts individual characters (including punctuation and return characters) rather than words. The model wasn’t trained on the concept of words, sentences, paragraphs, etc. but appears to have picked up on it.

After training the model on a massive list of Emily Dickinson’s poems, I asked it to predict the next few characters based on the input of \n (a return character). The example poems below are what the model considered to be the most probable characters in the most probable order.

I used The Complete Project Gutenberg’s Poems to fuel the model’s training data. I fed the model a large single file, but a better approach would be to break each poem out separately before training.

Sample input poem

DESIRE.

Who never wanted, — maddest joy
   Remains to him unknown:
The banquet of abstemiousness
   Surpasses that of wine.

Within its hope, though yet ungrasped
   Desire's perfect goal,
No nearer, lest reality
   Should disenthrall thy soul.

Highlights from the generated poems

Untitled (first attempt)

Smale got wot' lost the blood.Sweec that from though moundered
And God, and-beath just are
Concomate I lest you,
    How Itemity.Who lust it glow is tere.The ridet his wait,
In privilener through till.Whose badnity what it too!
That fluen the seal that?
"Revice the atte!I failed of ancont, aganor!
Winds thou all take uthingful
Of inlit bolnow, —
But shall yet in ear,
To his these mistrancelf near,
In compentecant my little mound
Which landly guight be;I the

OH

OHPungred so smides and hold!
I doon st is can to chair, abseab,
Cancelt to perches of the ais,
And you shall not have to take
And ofter donther hell
    On thou the slope would be;If I should disappoint the eyes
That hunted take thee?
Then dinnar' revel,
Of the is lone becad.The grass die wou nettled monnetsmy
At Lavid for your theen
Berest the heaven haughty.
It was not night that way!
Sovered solemn when the East
With such a pitely children on my father's will.

not aftaing!

not aftaing!Each neighbor zoulled hours,
Araugh of surprise,
One little host in blaze below
That mast and pranor of wine.An oper for meray'mand
    And anl yet to-day, —
Could kingdows hannot a charge
    To hel she is failed syefest
Upon a thought my dairy.Why need insprites the strite;
'T was with I ever canchead,"
We usselves in her going;Of, fas it Just as that deathous
    When burness to rest.The say, I shall despaired thread, —
    A hocome to reat.Caultions

THE GAI HE MAL

THE GAI HE MAL.The old brighatedices chass Carved ther away
    Serappicabed to be a field,
And yet that ore the decture sea,
Some natire of lately row,
The plead is brigrion
Then the hurred strings the one
Than head I sailod that ang I
South through the streets came a fut
    Bereath the wing was so nore,
As if the room from through poats!

Key things to notice:

The model appropriately places punctuation at the end of lines.
The model is adding return characters and indentations to match Emily Dickinson’s style.
The model capitalizes the first letter on each line.
The model capitalizes the word “God”.
The model uses dashes and semicolons heavily.

These are all consequences of Emily Dickinson’s writing style as well as the format of the text file that I provided. There’s a related note from the transcriber:

“As is well documented, Emily Dickinson’s poems were edited in these early editions by her friends, better to fit the conventions of the times. In particular, her dashes, often small enough to appear as dots, became commas and semi-colons.”

I Forged Emily Dickinson’s Poetry with an RNN

DATA SCIENCE | MACHINE LEARNING