12M legal training/eval examples for 17 global jurisdictions & 24 languages; suggestions that models' legal and commonsense reasoning interfere with each other.

The first application of structured outputs and constrained decoding to large language models; 74% error reduction in speech translation performance (relative to an oracle) across languages.

A framework for integrative research in text analysis by connecting findings to social science theories.

Transfer learning of grapheme-to-phoneme conversion in truly low-resource settings.

Large language models extract useful representations for alignment, annotation projection, and downstream tasks like POS tagging and NER.

In recent years, subpar time allocation by the field has been a major obstacle for NLP research. We propose remedies to improve the status quo.

A multi-faceted distributional analysis of the underlying word-formation processes that can aid in OOV words' compositional translation, tagging, parsing, language modeling, and other NLP tasks.

Construction of a useful and timely corpus: the Hong Kong Protest News Dataset, along with investigation of patterns of journalistic practice.

A push towards inclusion of derivational morphology in UniMorph, expanding on UniMorph 3.0 (McCarthy et al., 2020).

Computational and social science can benefit from greater interaction, considering the questions of big data versus theory, prediction versus explanation, and praxis versus substance.

We learn a language's morphology from raw text alone, without even external word lists.

Considering five dimensions of training data and domain, we show that mBART and mT5 cannot learn to translate languages phylogenically unrelated to their pre-training languages.

Energy-based sequence models backed by expressive parametric families can result in uncomputable and inapproximable partition functions. Statistical procedures with asymptotic guarantees and sheer amounts of compute are not the only things that make sequence modeling work.

A multi-task generalized linear model with BERT features enables new item difficulty estimates without piloting them first, which in turn limits item exposure and thus enhances test item security. Deployed in the Duolingo English Test, cutting pilot testing time 10×.

NLP discovers different media portrayals of the Hong Kong protests between the Hong Kong and Western presses, and analysis interprets it.

The goal of the shared task is to cluster word types from a raw text corpus into paradigms, and the best performing system is based on adaptor grammars.

We apply statistical techniques from natural language processing to a collection of Western and Hong Kong–based English-language newspaper articles spanning the years 1998–2020.

We design a probabilistic approach to synthesize new word forms, allowing it to operate in settings where correct translations have not been observed in text.

We craft novel information theoretic metrics for measuring similarity between gender systems, then use these to test hypotheses about Indo-European phylogeny.

Fine-tuning strong machine translation models with in-domain data and filtering n-best lists yields top performance in the shared task.

Design and findings of the SIGMORPHON 2020 shared task on multilingual grapheme-to-phoneme conversion.

A simple and effective approach to address posterior collapse in conditional variational autoencoders (CVAEs). Yields improved translation with noisy or monolingual data.

Discovery of morphological paradigms from raw text, with the potential to improve tools for low-resource languages or to assist linguistic annotators.

An information-theoretic method supporting the finding that form and meaning are relevant to grouping nouns into declension classes, with variation across languages.

A novel task in the field of inflectional morphology, where none of the submitted systems was able to improve over the baseline on average over all 9 test languages.

Kilolanguage generation, lemmatization, and fine-grained tagging, supplemented and improved through ensembling and dictionary-based reranking, for 1000+ languages.

Findings from the creation of a massively parallel corpus in over 1600 languages, the Johns Hopkins University Bible Corpus (JHUBC).

Adding more languages to a training set is often better, but too many harms performance—the best number depends on the source language.

We introduce WikiPron, an open-source command-line tool for extracting pronunciation data from Wiktionary, a collaborative multilingual online dictionary.

More than 2x the number of languages, as well as new parts of speech and validation tools, in the latest UniMorph morphological inventory release.

Autoencoding speaker conversion is combined with augmentation by machine-translated transcripts to obtain a competitive end-to-end AST model that outperforms a very strong cascade model.

By using the correct random model for the problem domain, this work provides a stronger, exact No Free Lunch theorem for community detection and generalizes to other set-partitioning tasks including core/periphery separation, k-clustering, and graph partitioning.

Normalized mutual information, the standard metric for community detection, exaggerates performance. We show why and recommend one-sided AMI under the 𝕄ₐₗₗ random model (all partitions of 𝑛 nodes) instead.

Tests the Berlin and Kay color-term hypotheses on thousands of languages at once, using massively parallel text as weak supervision.

An error taxonomy for morphological generation, applied by hand to the errors of the top two CoNLL–SIGMORPHON shared task systems across twelve languages.

A variety of approaches like data augmentation by translating ASR transcripts proves most effective on the English–French augmented LibriSpeech dataset, closing the performance gap from 8.2 to 1.4 BLEU, compared to a very strong cascade that could directly utilize copious ASR and MT data.

A holistic quantification of the systematicity of the sign using mutual information and recurrent neural networks.

Cross-lingual transfer and contextual analysis in morphology examined transfer learning of inflection between 100 language pairs, as well as contextual lemmatization and morphosyntactic description in 66 languages.

A precursor to modern parameter-efficient finetuning approaches: performance is surprisingly good when a single component is adapted while holding the rest of the model fixed.

A deterministic mapping from Universal Dependencies v2 features into the UniMorph schema is presented to ease interoperability and a critical evaluation of the foundations, strengths, and weaknesses of the two annotation projects is presented.

Morphological generation in 103 languages and inflecting words in sentential context, similar to a cloze task.

A project hosted at the Johns Hopkins University. Advances made to the schema, tooling, and dissemination of project resources since the UniMorph 1.0 release described at LREC 2016 are detailed.

Interdisciplinary analyses of protest coverage in print media that will "appeal to a wide range of readers with interests in computational social science, public policy, political sciences as well as policy-makers, think tanks, and practitioners who focus on the China-Hong Kong nexus".