Large Language Models (LLMs) have been observed to encode and perpetuate harmful associations present in the training data. We propose a theoretically grounded framework called StereoMap to gain insights into their perceptions of how demographic groups have been viewed by society. The framework is grounded in the Stereotype Content Model (SCM); a well-established theory from psychology. According to SCM, stereotypes are not all alike. Instead, the dimensions of Warmth and Competence serve as the factors that delineate the nature of stereotypes. Based on the SCM theory, StereoMap maps LLMs’ perceptions of social groups (defined by socio-demographic features) using the dimensions of Warmth and Competence. Furthermore, the framework enables the investigation of keywords and verbalizations of reasoning of LLMs’ judgments to uncover underlying factors influencing their perceptions. Our results show that LLMs exhibit a diverse range of perceptions towards these groups, characterized by mixed evaluations along the dimensions of Warmth and Competence. Furthermore, analyzing the reasonings of LLMs, our findings indicate that LLMs demonstrate an awareness of social disparities, often stating statistical data and research findings to support their reasoning. This study contributes to the understanding of how LLMs perceive and represent social groups, shedding light on their potential biases and the perpetuation of harmful associations.

As language models continue to be integrated into applications of personal and societal relevance, ensuring these models’ trustworthiness is crucial, particularly with respect to producing consistent outputs regardless of sensitive attributes. Given that first names may serve as proxies for (intersectional) socio-demographic representations, it is imperative to examine the impact of first names on commonsense reasoning capabilities. In this paper, we study whether a model’s reasoning given a specific input differs based on the first names provided. Our underlying assumption is that the reasoning about Alice should not differ from the reasoning about James. We propose and implement a controlled experimental framework to measure the causal effect of first names on commonsense reasoning, enabling us to distinguish between model predictions due to chance and caused by actual factors of interest. Our results indicate that the frequency of first names has a direct effect on model prediction, with less frequent names yielding divergent predictions compared to more frequent names. To gain insights into the internal mechanisms of models that are contributing to these behaviors, we also conduct an in-depth explainable analysis. Overall, our findings suggest that to ensure model robustness, it is essential to augment datasets with more diverse first names during the configuration stage.

Recent research has shown that large-scale pretrained language models, specifically transformers, tend to exhibit issues relating to racism, sexism, religion bias, and toxicity in general. Unfortunately, these pretrained language models are used almost universally in downstream tasks, and natural language processing is often applied to make real-world predictions. Thus, debiasing these language models as early in development as possible is increasingly crucial for preventing unintentional harms caused by natural language systems. To this end, we propose a new technique called partitioned contrastive gradient unlearning (PCGU), a gray-box method for debiasing pretrained masked language models. PCGU aims to optimize only the weights that contribute most to a specific domain of bias, doing so by computing a first-order approximation based on the gradients of contrastive sentence pairs. Our experiments show that PCGU is both low-cost and seems particularly effective at pinpointing the sources of implicit social bias in large pretrained transformers. Although we train using PCGU in the gender-profession domain only, we find that doing so can also partially mitigate bias across other domains. All code for our implementation and experiments can be found at https://github.com/CharlesYu2000/PCGU-UnlearningBias.

In text summarization, extractivity is defined as a measurement of the degree of overlap between a source document and its summary. Previous research has shown that the extractivity level of training data can influence both output extractivity and the amount of factual information (i.e. faithfulness) in outputs for abstractive summarization. However, it remains unclear if and how extractivity impacts the performance of abstractive models. In this work, we investigate the relationship between dataset extractivity and model performance by comparing the performance of trained models under different degrees of extractivity. We find that while low levels of extractivity can improve performance, as extractivity increases, performance is negatively impacted. Furthermore, through an analysis of the model’s copy continuity of content, we discover that higher extractivity leads to a greater tendency for the model to copy text continuously from the source document rather than identifying and summarizing important content that should be covered in the target summary. To address these issues, we propose a simple and effective method to design copy labels for fixing the model’s copying behaviors and train the model with a copy mechanism. The experimental results illustrate the effectiveness of our strategy in alleviating the negative impact on model performance resulting from high dataset extractivity, and that our method outperforms several competitive baselines.

Previous work has examined how debiasing language models affect downstream tasks, specifically, how debiasing techniques influence task performance and whether debiased models also make impartial predictions in downstream tasks or not. However, what we don’t understand well yet is why debiasing methods have varying impacts on downstream tasks and how debiasing techniques affect internal components of language models, i.e., neurons, layers, and attentions. In this paper, we decompose the internal mechanisms of debiasing language models with respect to gender by applying causal mediation analysis to understand the influence of debiasing methods on toxicity detection as a downstream task. Our findings suggest a need to test the effectiveness of debiasing methods with different bias metrics, and to focus on changes in the behavior of certain components of the models, e.g.,first two layers of language models, and attention heads.

This paper critically examines the current practices of benchmark dataset sharing in NLP and suggests a better way to inform reusers of the benchmark dataset. As the dataset sharing platform plays a key role not only in distributing the dataset but also in informing the potential reusers about the dataset, we believe data-sharing platforms should provide a comprehensive context of the datasets. We survey four benchmark dataset sharing platforms: HuggingFace, PaperswithCode, Tensorflow, and Pytorch to diagnose the current practices of how the dataset is shared which metadata is shared and omitted. To be specific, drawing on the concept of data curation which considers the future reuse when the data is made public, we advance the direction that benchmark dataset sharing platforms should take into consideration. We identify that four benchmark platforms have different practices of using metadata and there is a lack of consensus on what social impact metadata is. We believe the problem of missing a discussion around social impact in the dataset sharing platforms has to do with the failed agreement on who should be in charge. We propose that the benchmark dataset should develop social impact metadata and data curator should take a role in managing the social impact metadata.