a quantitative history of metaphors in english poetry jean.land/metaphors Ever since I was a kid I've wanted to map every metaphor in English poetry. Now thanks to a very cool Cohere grant and Muse Contributor Tier pricing, I mapped 139k metaphors from 75k poems out of a 136k poem corpus. This is but a small undertaking in investigating the evolution of figurative language, an interesting problem in a larger slice of my larger recreational undertaking, the computational history of consciousness. Every time I work on this project, I have felt an acute, stabbing pain in my balls. I have tried a billion iterations of this project and never achieved a result good enough to publish until AI got good enough. But now I think it's good enough to share! my problem there are no corpus for figurative language in poetry. the closest thing is the VU metaphor corpus but that one lowkey has no poetry. There are some hand-tagged metaphor libraries and prior studies (1, 2,3,4,5) on this subject but all the problems are 1) the corpus seemed too small 2) not poetry specific (mapping metaphor project for example is for all language). Past studies have tested too few poems. To put it shortly, the problem is 1) skill issue and 2) no one gaf. I wanted a large dataset of canonically accepted poems and a temporal network of "A is like B" type shit in all of English poetry. To do this, I read a lot of papers over course of multiple years when I'm not working, and I finally have a good dataset. I want it to be representative for the past, and selectively comprehensive for the present. For problems like this, it is more important to know what to exclude than what to include. To this end I've brought together teaching corpora and contemporary magazines. Every poem here belongs here. my method LLMs have supposedly performed poorly in detecting metaphors. "Prompted LLMs implementing MIPVU performed substantially below supervised metaphor detectors in Reimann and Scheffler's 2025 study, whereas Fuoli et al.'s 2025 full-text experiments obtained a median F1 of .79 after fine-tuning and .76 with their best chain-of-thought configuration." I wondered if this was a skill issue. Models have came far; and also no one gaf. Since that study was for film reviews, not poetry, there was lowkey no reason I would redo this study. But because I had no benchmarks I thought I had to do that anyways. my dataset I have perhaps the awesomest poetry corpus ever. jean.land/metaphors/reader.h… I built this painstakingly over the past 5 or so years. These are things people read, or should read, and definitely take serious. I took the poets.org poems; poetryfoundation, stitched together a few open source datsets, and then from ~425 literary magazines. I deduped common ones. In total I have 163k poems. The timing/slant of this poetry corpus is deliberate: I'm not really interested in a more complete sample of older English poetry. I think listing on the Internet means that those poems are continuing to being read, taught, and studied (by however few people) and thereby influence language of future poets. What is around in this current dataset also roughly proxies influence. This can very easily be done for other languages. Despite all this, my dataset still lowkey ass due to the following reasons: there are some undated poems; there may be some metadata issues, and the quality is too uneven in periods. Whats more, there are period biases, and the contemporary corpus may be a little ass and not comprehensive. It's definitely representative enough within the confines of what I, independent jean, can get without violating copyright. some stats below: If I had more money I could do it better. Right now I do a cheap pass where I go "Is this a poem?" and model goes yeah/nah and if its a poem its added to corpus. This is all extremely cheap btw like I assume they are doing it at the Universities. I would love to be pointed towards better corpora made by people with more money and more human attention and more time than me. the metaphor extraction step is simple. I have Muse Spark contributor tier (I give facebook my own money. It was cheapest option for outcome. Thank you Sophia for suggestion) and I have Cohere's very generous grant (~$3k total, ty!). Running all 163k poems, I just kept metaphors that both Muse Spark 1.2 contributor and Cohere Command A found. EZ🫛Z. The results really do look pretty good to me. Since there are no evals for this, if you think AI missed anything, please let me know. image is love in 15th century -> love in 21st century