Citations are a fundamental and indispensable part of research writing. They provide support and lend credibility to research findings. Recent GPT-fueled interest in large language models (LLMs) has shone a spotlight on the capabilities and limitations of these models when generating relevant citations for a document. Recent work has focused largely on title and author accuracy. We underline this effort and expand on it with a preliminary exploration in relevance of model-recommended citations. We define three citation-recommendation tasks. We also collect and annotate a dataset of model-recommended citations for those tasks. We find that GPT-4 largely outperforms earlier models on both author and title accuracy in two markedly different CS venues, but may not recommend references that are more relevant than those recommended by the earlier models. The two venues we compare are CHI and EMNLP. All models appear to perform better at recommending EMNLP papers than CHI papers.

Cross-references link source passages of text to other passages that elucidate the source passage in some way and can deepen human understanding. Despite their usefulness, however, good cross-references are hard to find, and extensive sets of cross-references only exist for the few most highly studied books such as the Bible, for which scholars have been collecting cross-references for hundreds of years. Therefore, we propose a new task: generate cross-references for user-selected text on demand. We define a metric, coverage, to evaluate task performance. We adapt several models to generate cross references, including an Anchor Words topic model, SBERT SentenceTransformers, and ChatGPT, and evaluate their coverage in both English and German on existing cross-reference datasets. While ChatGPT outperforms other models on these datasets, this is likely due to data contamination. We hand-evaluate performance on the well-known works of Jane Austen and a less-known science fiction series Sons of the Starfarers by Joe Vasicek, finding that ChatGPT does not perform as well on these works; sentence embeddings perform best. We experiment with newer LLMs and large context windows, and suggest that future work should focus on deploying cross-references on-demand with readers to determine their effectiveness in the wild.

Topic models are typically evaluated with respect to the global topic distributions that they generate, using metrics such as coherence, but without regard to local (token-level) topic assignments. Token-level assignments are important for downstream tasks such as classification. Even recent models, which aim to improve the quality of these token-level topic assignments, have been evaluated only with respect to global metrics. We propose a task designed to elicit human judgments of token-level topic assignments. We use a variety of topic model types and parameters and discover that global metrics agree poorly with human assignments. Since human evaluation is expensive we propose a variety of automated metrics to evaluate topic models at a local level. Finally, we correlate our proposed metrics with human judgments from the task on several datasets. We show that an evaluation based on the percent of topic switches correlates most strongly with human judgment of local topic quality. We suggest that this new metric, which we call consistency, be adopted alongside global metrics such as topic coherence when evaluating new topic models.