Natural Language Processing research has become increasingly concerned with understanding data quality and complexity at the instance level. Instance-level complexity scores can be used for tasks such as filtering out noisy observations and subsampling informative examples. However, there exists a diverse taxonomy of complexity metrics that can be used for a classification task, making metric selection itself a difficult task. We empirically examine the relationship between these metrics and find that simply storing training loss provides similar complexity rankings as other more computationally intensive techniques. Metric similarity allows us to subsample data with higher aggregate complexity along several metrics using a single a priori available meta-feature. Further, this choice of complexity metric does not impact demographic fairness, even in downstream predictions. Researchers should consider metric availability and similarity, as using the wrong metric or sampling strategy may hurt performance.

Transformer-based pretrained large language models (PLM) such as BERT and GPT have achieved remarkable success in NLP tasks. However, PLMs are prone to encoding stereotypical biases. Although a burgeoning literature has emerged on stereotypical bias mitigation in PLMs, such as work on debiasing gender and racial stereotyping, how such biases manifest and behave internally within PLMs remains largely unknown. Understanding the internal stereotyping mechanisms may allow better assessment of model fairness and guide the development of effective mitigation strategies. In this work, we focus on attention heads, a major component of the Transformer architecture, and propose a bias analysis framework to explore and identify a small set of biased heads that are found to contribute to a PLM’s stereotypical bias. We conduct extensive experiments to validate the existence of these biased heads and to better understand how they behave. We investigate gender and racial bias in the English language in two types of Transformer-based PLMs: the encoder-based BERT model and the decoder-based autoregressive GPT model, LLaMA-2 (7B), and LLaMA-2-Chat (7B). Overall, the results shed light on understanding the bias behavior in pretrained language models.

In-Context Learning (ICL) and Instruction Tuning (IT) are two primary paradigms of adopting Large Language Models (LLMs) to downstream applications. However, they are significantly different. In ICL, a set of demonstrations is provided at the inference time, but the LLM’s parameters are not updated. In IT, a set of demonstrations is used to adjust the parameters of the LLM during training, but no demonstrations are provided at the inference time. Although a growing body of literature has explored ICL and IT, studies on these topics have largely been conducted in isolation, leading to a disconnect between these two paradigms. In this work, we explore the relationship between ICL and IT by examining how the hidden states of LLMs change in these two paradigms. Through carefully designed experiments conducted with LLaMA-2 and LLaMA-2-Chat (7B and 13B), we find that ICL and IT converge in LLM hidden states despite their apparent differences in implementation. Specifically, ICL changes an LLM’s hidden states as if its accompanying demonstrations were used to instructionally tune the model. Furthermore, the convergence between ICL and IT is largely contingent upon several factors related to the demonstration. Overall, this work offers a unique perspective to explore the connection between ICL and IT and sheds light on understanding the behaviors of LLMs.

Human-like biases and undesired social stereotypes exist in large pretrained language models. Given the wide adoption of these models in real-world applications, mitigating such biases has become an emerging and important task. In this paper, we propose an automatic method to mitigate the biases in pretrained language models. Different from previous debiasing work that uses external corpora to fine-tune the pretrained models, we instead directly probe the biases encoded in pretrained models through prompts. Specifically, we propose a variant of the beam search method to automatically search for biased prompts such that the cloze-style completions are the most different with respect to different demographic groups. Given the identified biased prompts, we then propose a distribution alignment loss to mitigate the biases. Experiment results on standard datasets and metrics show that our proposed Auto-Debias approach can significantly reduce biases, including gender and racial bias, in pretrained language models such as BERT, RoBERTa and ALBERT. Moreover, the improvement in fairness does not decrease the language models’ understanding abilities, as shown using the GLUE benchmark.

Machine learning models often suffer from a performance drop when they are applied to out-of-distribution (OOD) samples, i.e., those drawn far away from the training data distribution. Existing OOD detection work mostly focuses on identifying semantic-shift OOD samples, e.g., instances from unseen new classes. However, background-shift OOD detection, which identifies samples with domain or style-change, represents a more practical yet challenging task. In this paper, we propose Background-Aware Representation Learning (BARLE) for background-shift OOD detection in NLP. Specifically, we generate semantics-preserving background-shifted pseudo OOD samples from pretrained masked language models. We then contrast the in-distribution (ID) samples with their pseudo OOD counterparts. Unlike prior semantic-shift OOD detection work that often leverages an external text corpus, BARLE only uses ID data, which is more flexible and cost-efficient. In experiments across several text classification tasks, we demonstrate that BARLE is capable of improving background-shift OOD detection performance while maintaining ID classification accuracy. We further investigate the properties of the generated pseudo OOD samples, uncovering the working mechanism of BARLE.

There has been a recent wave of work assessing the fairness of machine learning models in general, and more specifically, on natural language processing (NLP) models built using machine learning techniques. While much work has highlighted biases embedded in state-of-the-art language models, and more recent efforts have focused on how to debias, research assessing the fairness and performance of biased/debiased models on downstream prediction tasks has been limited. Moreover, most prior work has emphasized bias along a single dimension such as gender or race. In this work, we benchmark multiple NLP models with regards to their fairness and predictive performance across a variety of NLP tasks. In particular, we assess intersectional bias - fairness across multiple demographic dimensions. The results show that while current debiasing strategies fare well in terms of the fairness-accuracy trade-off (generally preserving predictive power in debiased models), they are unable to effectively alleviate bias in downstream tasks. Furthermore, this bias is often amplified across dimensions (i.e., intersections). We conclude by highlighting possible causes and making recommendations for future NLP debiasing research.

Psychometric measures of ability, attitudes, perceptions, and beliefs are crucial for understanding user behavior in various contexts including health, security, e-commerce, and finance. Traditionally, psychometric dimensions have been measured and collected using survey-based methods. Inferring such constructs from user-generated text could allow timely, unobtrusive collection and analysis. In this paper we describe our efforts to construct a corpus for psychometric natural language processing (NLP) related to important dimensions such as trust, anxiety, numeracy, and literacy, in the health domain. We discuss our multi-step process to align user text with their survey-based response items and provide an overview of the resulting testbed which encompasses survey-based psychometric measures and accompanying user-generated text from 8,502 respondents. Our testbed also encompasses self-reported demographic information, including race, sex, age, income, and education - thereby affording opportunities for measuring bias and benchmarking fairness of text classification methods. We report preliminary results on use of the text to predict/categorize users’ survey response labels - and on the fairness of these models. We also discuss the important implications of our work and resulting testbed for future NLP research on psychometrics and fairness.

Twitter has become one of the quintessential social media platforms for user-generated content. Researchers and industry practitioners are increasingly interested in Twitter sentiments. Consequently, an array of commercial and freely available Twitter sentiment analysis tools have emerged, though it remains unclear how well these tools really work. This study presents the findings of a detailed benchmark analysis of Twitter sentiment analysis tools, incorporating 20 tools applied to 5 different test beds. In addition to presenting detailed performance evaluation results, a thorough error analysis is used to highlight the most prevalent challenges facing Twitter sentiment analysis tools. The results have important implications for various stakeholder groups, including social media analytics researchers, NLP developers, and industry managers and practitioners using social media sentiments as input for decision-making.