We adapt BLiMP (Benchmark of Linguistic Minimal Pairs) language model evaluation framework to the context of poetry, introducing the first of a series of tasks titled Benchmark of Poetic Minimal Pairs (BPoMP). The tasks presented herein use one genre of English-language poetry, the limerick (five-lines, rhyme scheme AABBA). Following the BLiMP schema, the BPoMP tasks use 10,000 minimal pairs of limerick/corrupted limerick. The latter is created by (1) shuffling two rhyming end-of-the-line words, (2) shuffling two rhyming lines, (3) replacing end-of-the-line word by a non-rhyming synonym. Our general task is detection of the original limerick, which we believe tests a language model’s capacity to utilize “end rhymes”, a common feature of poetry. We evaluate Transformer-based models by checking if they assign a higher probability to the non-corrupted limerick in each minimal pair. We find that the models identify the original limerick at rates better than chance, but with a nontrivial gap relative to human accuracy (average of 98.3% across tasks). The publicly available curated set of limericks accompanying this paper is an additional contribution. In general, we see this as a first step to create a community of NLP activity around the rigorous computational study of poetry.

The style transfer task (here style is used in a broad “authorial” sense with many aspects including register, sentence structure, and vocabulary choice) takes text input and rewrites it in a specified target style preserving the meaning, but altering the style of the source text to match that of the target. Much of the existing research on this task depends on the use of parallel datasets. In this work we employ recent results in unsupervised cross-lingual language modeling (XLM) and machine translation to effect style transfer while treating the input data as unaligned. First, we show that adding “content embeddings” to the XLM which capture human-specified groupings of subject matter can improve performance over the baseline model. Evaluation of style transfer has often relied on metrics designed for machine translation which have received criticism of their suitability for this task. As a second contribution, we propose the use of a suite of classical stylometrics as a useful complement for evaluation. We select a few such measures and include these in the analysis of our results.

In this paper we take up the problem of “limerick detection” and describe a system to identify five-line poems as limericks or not. This turns out to be a surprisingly difficult challenge with many subtleties. More precisely, we produce an algorithm which focuses on the structural aspects of the limerick – rhyme scheme and rhythm (i.e., stress patterns) – and when tested on a a culled data set of 98,454 publicly available limericks, our “limerick filter” accepts 67% as limericks. The primary failure of our filter is on the detection of “non-standard” rhymes, which we highlight as an outstanding challenge in computational poetics. Our accent detection algorithm proves to be very robust. Our main contributions are (1) a novel rhyme detection algorithm that works on English words including rare proper nouns and made-up words (and thus, words not in the widely used CMUDict database); (2) a novel rhythm-identifying heuristic that is robust to language noise at moderate levels and comparable in accuracy to state-of-the-art scansion algorithms. As a third significant contribution (3) we make publicly available a large corpus of limericks that includes tags of “limerick” or “not-limerick” as determined by our identification software, thereby providing a benchmark for the community. The poetic tasks that we have identified as challenges for machines suggest that the limerick is a useful “model organism” for the study of machine capabilities in poetry and more broadly literature and language. We include a list of open challenges as well. Generally, we anticipate that this work will provide useful material and benchmarks for future explorations in the field.