In-context learning (ICL) is a new learning paradigm that has gained popularity along with the development of large language models. In this work, we adapt a recently proposed hardness metric, pointwise 𝒱-usable information (PVI), to an in-context version (in-context PVI). Compared to the original PVI, in-context PVI is more efficient in that it requires only a few exemplars and does not require fine-tuning. We conducted a comprehensive empirical analysis to evaluate the reliability of in-context PVI. Our findings indicate that in-context PVI estimates exhibit similar characteristics to the original PVI. Specific to the in-context setting, we show that in-context PVI estimates remain consistent across different exemplar selections and numbers of shots. The variance of in-context PVI estimates across different exemplar selections is insignificant, which suggests that in-context PVI estimates are stable. Furthermore, we demonstrate how in-context PVI can be employed to identify challenging instances. Our work highlights the potential of in-context PVI and provides new insights into the capabilities of ICL.

The bias-variance tradeoff is the idea that learning methods need to balance model complexity with data size to minimize both under-fitting and over-fitting. Recent empirical work and theoretical analysis with over-parameterized neural networks challenges the classic bias-variance trade-off notion suggesting that no such trade-off holds: as the width of the network grows, bias monotonically decreases while variance initially increases followed by a decrease. In this work, we first provide a variance decomposition-based justification criteria to examine whether large pretrained neural models in a fine-tuning setting are generalizable enough to have low bias and variance. We then perform theoretical and empirical analysis using ensemble methods explicitly designed to decrease variance due to optimization. This results in essentially a two-stage fine-tuning algorithm that first ratchets down bias and variance iteratively, and then uses a selected fixed-bias model to further reduce variance due to optimization by ensembling. We also analyze the nature of variance change with the ensemble size in low- and high-resource classes. Empirical results show that this two-stage method obtains strong results on SuperGLUE tasks and clinical information extraction tasks. Code and settings are available: https://github.com/christa60/bias-var-fine-tuning-plms.git

This study evaluated ChatGPT’s ability to understand causal language in science papers and news by testing its accuracy in a task of labeling the strength of a claim as causal, conditional causal, correlational, or no relationship. The results show that ChatGPT is still behind the existing fine-tuned BERT models by a large margin. ChatGPT also had difficulty understanding conditional causal claims mitigated by hedges. However, its weakness may be utilized to improve the clarity of human annotation guideline. Chain-of-Thoughts were faithful and helpful for improving prompt performance, but finding the optimal prompt is difficult with inconsistent results and the lack of effective method to establish cause-effect between prompts and outcomes, suggesting caution when generalizing prompt engineering results across tasks or models.

Prior studies have raised concerns over specificity issues in clinical advice. Lacking specificity — explicitly discussed detailed information — may affect the quality and implementation of clinical advice in medical practice. In this study, we developed and validated a fine-grained annotation schema to describe different aspects of specificity in clinical advice extracted from medical research literature. We also presented our initial annotation effort and discussed future directions towards an NLP-based specificity analysis tool for summarizing and verifying the details in clinical advice.

Health and medical researchers often give clinical and policy recommendations to inform health practice and public health policy. However, no current health information system supports the direct retrieval of health advice. This study fills the gap by developing and validating an NLP-based prediction model for identifying health advice in research publications. We annotated a corpus of 6,000 sentences extracted from structured abstracts in PubMed publications as ‘“strong advice”, “weak advice”, or “no advice”, and developed a BERT-based model that can predict, with a macro-averaged F1-score of 0.93, whether a sentence gives strong advice, weak advice, or not. The prediction model generalized well to sentences in both unstructured abstracts and discussion sections, where health advice normally appears. We also conducted a case study that applied this prediction model to retrieve specific health advice on COVID-19 treatments from LitCovid, a large COVID research literature portal, demonstrating the usefulness of retrieving health advice sentences as an advanced research literature navigation function for health researchers and the general public.

Press releases have an increasingly strong influence on media coverage of health research; however, they have been found to contain seriously exaggerated claims that can misinform the public and undermine public trust in science. In this study we propose an NLP approach to identify exaggerated causal claims made in health press releases that report on observational studies, which are designed to establish correlational findings, but are often exaggerated as causal. We developed a new corpus and trained models that can identify causal claims in the main statements in a press release. By comparing the claims made in a press release with the corresponding claims in the original research paper, we found that 22% of press releases made exaggerated causal claims from correlational findings in observational studies. Furthermore, universities exaggerated more often than journal publishers by a ratio of 1.5 to 1. Encouragingly, the exaggeration rate has slightly decreased over the past 10 years, despite the increase of the total number of press releases. More research is needed to understand the cause of the decreasing pattern.

Causal interpretation of correlational findings from observational studies has been a major type of misinformation in science communication. Prior studies on identifying inappropriate use of causal language relied on manual content analysis, which is not scalable for examining a large volume of science publications. In this study, we first annotated a corpus of over 3,000 PubMed research conclusion sentences, then developed a BERT-based prediction model that classifies conclusion sentences into “no relationship”, “correlational”, “conditional causal”, and “direct causal” categories, achieving an accuracy of 0.90 and a macro-F1 of 0.88. We then applied the prediction model to measure the causal language use in the research conclusions of about 38,000 observational studies in PubMed. The prediction result shows that 21.7% studies used direct causal language exclusively in their conclusions, and 32.4% used some direct causal language. We also found that the ratio of causal language use differs among authors from different countries, challenging the notion of a shared consensus on causal language use in the global science community. Our prediction model could also be used to help identify the inappropriate use of causal language in science publications.

The discrepancy between science and media has been affecting the effectiveness of science communication. Original findings from science publications may be distorted with altered claim strength when reported to the public, causing misinformation spread. This study conducts an NLP analysis of exaggerated claims in science news, and then constructed prediction models for identifying claim strength levels in science reporting. The results demonstrate different writing styles journal articles and news/press releases use for reporting scientific findings. Preliminary prediction models reached promising result with room for further improvement.