Sonnet finder

Finds snippets in iambic pentameter in English-language text and tries to combine them to a rhyming sonnet.


This is a Python script that should run without a GPU or any other special hardware requirements.

  1. Install the required packages, e.g. via: pip install -r requirements.txt

  2. Prepare a plain text file, say input.txt, with text you want to make a sonnet out of (sonnet-ize? sonnet-ify?). It can have multiple sentences on the same line, but a sentence should not be split across multiple lines.

    For example, I used pandoc --to=plain --wrap=none to generate a text file from my LaTeX papers. You could also start by grabbing some text files from Project Gutenberg.

  3. Run sonnet finder: python input.txt -o output.tsv

    Using -o will save a list of all extracted candidate phrases, sorted by rhyme pattern, so you can generate new sonnets more quickly (see below) or browse and cherry-pick from the candidates to make your own sonnet out of these lines.

    Either way, the script will output a full example sonnet to STDOUT (provided enough rhyming pairs in iambic pentameter were found).

  4. If you’ve saved an output.tsv file before, you can quickly generate new sonnets via python output.tsv. Since the stress and pronunciation prediction can be slow on larger files, this is much better than re-running if you want more automatically generated suggestions.


This is a sonnet (with cherry-picked lines) made out of my PhD thesis:

the application of existing tools
describe a mapping to a modern form
applying similar replacement rules
the base ensembles slightly outperform

hungarian, icelandic, portuguese
perform a similar evaluation
contemporary lexemes or morphemes
a single dataset in isolation

historical and modern language stages
the weighted combination of encoder
the german dative ending -e in phrases
predictions fed into the next decoder

in this example from the innsbruck letter
machine translation still remains the better

These stanzas are compiled from a couple of automatically-generated suggestions based on the abstracts of all papers published in 2021 in the ACL Anthology:

effective algorithm that enables
improvements on a wide variety
and training with adjudicated labels
anxiety and test anxiety

obtain remarkable improvements on
decoder architecture, which equips
associated with the lexicon
surprising personal relationships

the impact of the anaphoric one
complexity prediction competition
developed for a laboratory run
existing parsers typically condition

examples, while in practice, most unseen
evaluate translation tasks between

Here’s the same using Moby Dick:

among the marble senate of the dead
offensive matters consequent upon
a crawling reptile of the land, instead
fifteen, eighteen, and twenty hours on

the lakeman now patrolled the barricade
egyptian tablets, whose antiquity
the waters seemed a golden finger laid
maintains a permanent obliquity

the pequod with the little negro pippin
and with a frightful roll and vomit, he
increased, besides perhaps improving it in
transparent air into the summer sea

the traces of a simple honest heart
the fishery, and not the thousandth part

(The emjambment in the third stanza here is a lucky coincidence; the script currently doesn’t do any kind of syntactic analysis or attempt coherence between lines.)

How it works

This script relies on the grapheme-to-phoneme library g2p_en by Park & Kim to convert the English input text to phoneme sequences (i.e., how the text would be pronounced). I chose this because it’s a pip-installable Python library that fulfills two important criteria:

  1. it’s not restricted to looking up pronunciations in a dictionary, but can handle arbitrary words through the use of a neural model (although, obviously, this will not always be accurate);

  2. it provides stress information for each vowel (i.e., whether any given vowel should be stressed or unstressed, which is important for determining the poetic meter).

The script then scans the g2p output for occurrences of iambic pentameter, i.e. a 0101010101(0) pattern, additionally checking if they coincide with word boundaries.

For finding snippets that rhyme, I rely mostly on Ghazvininejad et al. (2016), particularly §3 (relaxing the iambic pentameter a bit by allowing words that end in 100) and §5.2 (giving an operational definition of “slant rhyme” that I mostly try to follow).

QNA (Questions Nobody Asked)

  • Why does the script sometimes output lines that don’t rhyme or don’t fit the iambic meter? This script can only be as good as the grapheme-to-phoneme algorithm that’s used. It frequently fails on words it doesn’t know (for example, it tries to rhyme datasets with Portuguese?!) and also usually fails on abbreviations. Maybe there’s a better g2p library that could be used, or the existing g2p_en could be modified to accept a custom dictionary, so you could manually define pronunciations for commonly used words.

  • Could this script also generate other types of poems? Sure. You could start by changing the regex iambic_pentameter to something else; maybe a sequence of dactyls? There are some further hardcoded assumptions in the code about iambic pentameter in the function get_stress_and_boundaries() that might have to be modified.

  • Could this script generate poems in languages other than English? This would require a suitable replacement for g2p_en that predicts pronunciations and stress patterns for the desired language, as well as re-writing the code that determines whether two phrases can rhyme; see the comments in the script for details. In particular, the code for English uses ARPABET notation for the pronunciation, which won’t be suitable for other languages.

  • Can this script generate completely novel phrases in the style of an input text? This script does not “hallucinate” any text or generate anything that wasn’t already there in the input; if you want to do that, take a look at Deep-speare maybe.


Written by Marcel Bollmann, inspired by a tweet, licensed under the .

I’m not the first one to write a script like this, but it was a fun exercise!