Coreference annotation and resolution is a vital component of computational literary studies. However, it has previously been difficult to build high quality systems for fiction. Coreference requires complicated structured outputs, and literary text involves subtle inferences and highly varied language. New language-model-based seq2seq systems present the opportunity to solve both these problems by learning to directly generate a copy of an input sentence with markdown-like annotations. We create, evaluate, and release several trained models for coreference, as well as a workflow for training new models.

This article proposes a new family of LLM-based topic coherence metrics called Contextualized Topic Coherence (CTC) and inspired by standard human topic evaluation methods. CTC metrics simulate human-centered coherence evaluation while maintaining the efficiency of other automated methods. We compare the performance of our CTC metrics and five other baseline metrics on seven topic models and show that CTC metrics better reflect human judgment, particularly for topics extracted from short text collections by avoiding highly scored topics that are meaningless to humans.

Pretraining data design is critically under-documented and often guided by empirically unsupported intuitions. We pretrain models on data curated (1) at different collection times, (2) with varying toxicity and quality filters, and (3) with different domain compositions. First, we find that temporal shift between evaluation data and pretraining data leads to performance degradation, which is not overcome by finetuning. Second, we measure the effect of quality and toxicity filters, showing a trade-off between performance on standard benchmarks and risk of toxic generations. We also find that the effects of different types of filtering are not predictable from text domain characteristics. Third, we empirically validate that heterogeneous data sources, like books and web, are beneficial and warrant greater prioritization. To date, these experiments constitute the single largest publicly documented empirical study of the effects of pretraining data. Spanning 28 unique 1.5 billion parameter models pretrained from scratch, these findings validate, quantify, and expose many undocumented intuitions about text pretraining, which ultimately support more informed data-centric decisions in model development.

Cross-lingual transfer learning is an important property of multilingual large language models (LLMs). But how do LLMs represent relationships between languages? Every language model has an input layer that maps tokens to vectors. This ubiquitous layer of language models is often overlooked. We find that similarities between these input embeddings are highly interpretable and that the geometry of these embeddings differs between model families. In one case (XLM-RoBERTa), embeddings encode language: tokens in different writing systems can be linearly separated with an average of 99.2% accuracy. Another family (mT5) represents cross-lingual semantic similarity: the 50 nearest neighbors for any token represent an average of 7.61 writing systems, and are frequently translations. This result is surprising given that there is no explicit parallel cross-lingual training corpora and no explicit incentive for translations in pre-training objectives. Our research opens the door for investigations in 1) The effect of pre-training and model architectures on representations of languages and 2) The applications of cross-lingual representations embedded in language models.

Generative language models (LMs) are increasingly used for document class-prediction tasks and promise enormous improvements in cost and efficiency. Existing research often examines simple classification tasks, but the capability of LMs to classify on complex or specialized tasks is less well understood. We consider a highly complex task that is challenging even for humans: the classification of legal reasoning according to jurisprudential philosophy. Using a novel dataset of historical United States Supreme Court opinions annotated by a team of domain experts, we systematically test the performance of a variety of LMs. We find that generative models perform poorly when given instructions (i.e. prompts) equal to the instructions presented to human annotators through our codebook. Our strongest results derive from fine-tuning models on the annotated dataset; the best performing model is an in-domain model, LEGAL-BERT. We apply predictions from this fine-tuned model to study historical trends in jurisprudence, an exercise that both aligns with prominent qualitative historical accounts and points to areas of possible refinement in those accounts. Our findings generally sound a note of caution in the use of generative LMs on complex tasks without fine-tuning and point to the continued relevance of human annotation-intensive classification methods.

Large language models achieve high performance on many but not all downstream tasks. The interaction between pretraining data and task data is commonly assumed to determine this variance: a task with data that is more similar to a model’s pretraining data is assumed to be easier for that model. We test whether distributional and example-specific similarity measures (embedding-, token- and model-based) correlate with language model performance through a large-scale comparison of the Pile and C4 pretraining datasets with downstream benchmarks. Similarity correlates with performance for multilingual datasets, but in other benchmarks, we surprisingly find that similarity metrics are not correlated with accuracy or even each other. This suggests that the relationship between pretraining data and downstream tasks is more complex than often assumed.

A common factor in bias measurement methods is the use of hand-curated seed lexicons, but there remains little guidance for their selection. We gather seeds used in prior work, documenting their common sources and rationales, and in case studies of three English-language corpora, we enumerate the different types of social biases and linguistic features that, once encoded in the seeds, can affect subsequent bias measurements. Seeds developed in one context are often re-used in other contexts, but documentation and evaluation remain necessary precursors to relying on seeds for sensitive measurements.

We explore Boccaccio’s Decameron to see how digital humanities tools can be used for tasks that have limited data in a language no longer in contemporary use: medieval Italian. We focus our analysis on the question: Do the different storytellers in the text exhibit distinct personalities? To answer this question, we curate and release a dataset based on the authoritative edition of the text. We use supervised classification methods to predict storytellers based on the stories they tell, confirming the difficulty of the task, and demonstrate that topic modeling can extract thematic storyteller “profiles.”

Much of the progress in contemporary NLP has come from learning representations, such as masked language model (MLM) contextual embeddings, that turn challenging problems into simple classification tasks. But how do we quantify and explain this effect? We adapt general tools from computational learning theory to fit the specific characteristics of text datasets and present a method to evaluate the compatibility between representations and tasks. Even though many tasks can be easily solved with simple bag-of-words (BOW) representations, BOW does poorly on hard natural language inference tasks. For one such task we find that BOW cannot distinguish between real and randomized labelings, while pre-trained MLM representations show 72x greater distinction between real and random labelings than BOW. This method provides a calibrated, quantitative measure of the difficulty of a classification-based NLP task, enabling comparisons between representations without requiring empirical evaluations that may be sensitive to initializations and hyperparameters. The method provides a fresh perspective on the patterns in a dataset and the alignment of those patterns with specific labels.

Images can give us insights into the contextual meanings of words, but current image-text grounding approaches require detailed annotations. Such granular annotation is rare, expensive, and unavailable in most domain-specific contexts. In contrast, unlabeled multi-image, multi-sentence documents are abundant. Can lexical grounding be learned from such documents, even though they have significant lexical and visual overlap? Working with a case study dataset of real estate listings, we demonstrate the challenge of distinguishing highly correlated grounded terms, such as “kitchen” and “bedroom”, and introduce metrics to assess this document similarity. We present a simple unsupervised clustering-based method that increases precision and recall beyond object detection and image tagging baselines when evaluated on labeled subsets of the dataset. The proposed method is particularly effective for local contextual meanings of a word, for example associating “granite” with countertops in the real estate dataset and with rocky landscapes in a Wikipedia dataset.

Images and text co-occur constantly on the web, but explicit links between images and sentences (or other intra-document textual units) are often not present. We present algorithms that discover image-sentence relationships without relying on explicit multimodal annotation in training. We experiment on seven datasets of varying difficulty, ranging from documents consisting of groups of images captioned post hoc by crowdworkers to naturally-occurring user-generated multimodal documents. We find that a structured training objective based on identifying whether collections of images and sentences co-occur in documents can suffice to predict links between specific sentences and specific images within the same document at test time.

Despite great scalability on large data and their ability to understand correlations between topics, spectral topic models have not been widely used due to the absence of reliability in real data and lack of practical implementations. This paper aims to solidify the foundations of spectral topic inference and provide a practical implementation for anchor-based topic modeling. Beginning with vocabulary curation, we scrutinize every single inference step with other viable options. We also evaluate our matrix-based approach against popular alternatives including a tensor-based spectral method as well as probabilistic algorithms. Our quantitative and qualitative experiments demonstrate the power of Rectified Anchor Word algorithm in various real datasets, providing a complete guide to practical correlated topic modeling.

Multimodal machine learning algorithms aim to learn visual-textual correspondences. Previous work suggests that concepts with concrete visual manifestations may be easier to learn than concepts with abstract ones. We give an algorithm for automatically computing the visual concreteness of words and topics within multimodal datasets. We apply the approach in four settings, ranging from image captions to images/text scraped from historical books. In addition to enabling explorations of concepts in multimodal datasets, our concreteness scores predict the capacity of machine learning algorithms to learn textual/visual relationships. We find that 1) concrete concepts are indeed easier to learn; 2) the large number of algorithms we consider have similar failure cases; 3) the precise positive relationship between concreteness and performance varies between datasets. We conclude with recommendations for using concreteness scores to facilitate future multimodal research.

Most previous work in unsupervised semantic modeling in the presence of metadata has assumed that our goal is to make latent dimensions more correlated with metadata, but in practice the exact opposite is often true. Some users want topic models that highlight differences between, for example, authors, but others seek more subtle connections across authors. We introduce three metrics for identifying topics that are highly correlated with metadata, and demonstrate that this problem affects between 30 and 50% of the topics in models trained on two real-world collections, regardless of the size of the model. We find that we can predict which words cause this phenomenon and that by selectively subsampling these words we dramatically reduce topic-metadata correlation, improve topic stability, and maintain or even improve model quality.

Word embeddings are increasingly being used as a tool to study word associations in specific corpora. However, it is unclear whether such embeddings reflect enduring properties of language or if they are sensitive to inconsequential variations in the source documents. We find that nearest-neighbor distances are highly sensitive to small changes in the training corpus for a variety of algorithms. For all methods, including specific documents in the training set can result in substantial variations. We show that these effects are more prominent for smaller training corpora. We recommend that users never rely on single embedding models for distance calculations, but rather average over multiple bootstrap samples, especially for small corpora.

Duplicate documents are a pervasive problem in text datasets and can have a strong effect on unsupervised models. Methods to remove duplicate texts are typically heuristic or very expensive, so it is vital to know when and why they are needed. We measure the sensitivity of two latent semantic methods to the presence of different levels of document repetition. By artificially creating different forms of duplicate text we confirm several hypotheses about how repeated text impacts models. While a small amount of duplication is tolerable, substantial over-representation of subsets of the text may overwhelm meaningful topical patterns.

Despite their ubiquity, word embeddings trained with skip-gram negative sampling (SGNS) remain poorly understood. We find that vector positions are not simply determined by semantic similarity, but rather occupy a narrow cone, diametrically opposed to the context vectors. We show that this geometric concentration depends on the ratio of positive to negative examples, and that it is neither theoretically nor empirically inherent in related embedding algorithms.

It is often assumed that topic models benefit from the use of a manually curated stopword list. Constructing this list is time-consuming and often subject to user judgments about what kinds of words are important to the model and the application. Although stopword removal clearly affects which word types appear as most probable terms in topics, we argue that this improvement is superficial, and that topic inference benefits little from the practice of removing stopwords beyond very frequent terms. Removing corpus-specific stopwords after model inference is more transparent and produces similar results to removing those words prior to inference.

Rule-based stemmers such as the Porter stemmer are frequently used to preprocess English corpora for topic modeling. In this work, we train and evaluate topic models on a variety of corpora using several different stemming algorithms. We examine several different quantitative measures of the resulting models, including likelihood, coherence, model stability, and entropy. Despite their frequent use in topic modeling, we find that stemmers produce no meaningful improvement in likelihood and coherence and in fact can degrade topic stability.