GPT-3 Generated Haikus
Ive been facinated by this art form since reading The Heart of Haiku by Jane Hirshfield a number of years ago.
Given i’ve been learning the basics of GPT-3 the past couple of days, I thought that generating short poems would be firmly within its wheelhouse, and also a good learning experience on how to ‘prompt’ the system.
Method:
GPT-3 was prompted using the following haikus (in English):
1. The Old Pond by Matsuo Bashō,
2. A World of Dew by Kobayashi Issa,
3. Lighting One Candle by Yosa Buson,
4. A Poppy Blooms by Katsushika Hokusai,
5. Over the Wintry by Natsume Sōseki
I ‘asked’ GPT-3 to produce both the haiku title as well as the body of text. The engine used was davinchi, a Temp of 0.7 and a Top P of 1.
Limitations:
- We are training this on Japanese poems translated into English, and therefore the limitations on poem structure not conforming to the standard 5–7–5 syllables.
- Limited ‘prompt’ material.
- Only a small batch of poems generated: the below is pretty much an unfiltered first set of results.
Results:
Below are the results of the first batch of GPT-3 AI generated Haikus (in order of my subjective preference!):
By Cleaving Wood
By cleaving wood
I learn more of Truth
Than by reading booksA Child’s Dream
A child’s dream
Runs into a wall
Of reality.Starlight
On a clear October night
The fireflies glitter
Over the mountains.Years Are Just Hours
Years are just hours
Though dream-dimmed hours
of sleep.Riding on the Clouds
Riding on the clouds
The soul is lifted up
Above the woes of the world.A Winter Moon
A winter moon
In the silent sky —
All is frozen.Flowers of the season
Flowers of the season
All scatter their petals,
One by one, in the windAt the Foot of Autumn
The foot of autumn:
Infinite sadness and silence —
Yet bursts of inspiring energy!The Wild Birds
Dropping down to the ground
The wild birds choose a branch
To stayWinter Daydreaming
Winter daydreaming;
The heron under the deep snow;
High tide smacks as it retreats.
Initial Observations
- From what I can tell from a cursory review there is little to no plagiarism here, it has not simple repeated Haikus in its training material, or the prompt I have provided it.
- Most of the titles are simply the first line of the Haiku, this is likely given that a similar trend occurs within the training material. However, this is not universal. It is very interesting that “Starlight’ has no mention of stars but there is a reference to fireflies glittering over mountains.
- The structure of the Haikus are not very consistent, but this is likely due to training material of Japanese poems translated to English, therefore a next step is to train on native English versions.
- I do not see much evidence yet of juxtaposition or kireji ‘cutting word’, but again this may improve using native english haikus.
Overall i’m impressed with the outputs, which I believe could be improved by refining the prompts using a native english set — more to follow.
Thanks to Hamilton Keats @ Azoras.co.uk for GPT-3 access