Social bias is shaped by the accumulation of social perceptions towards targets across various demographic identities. To fully understand such social bias in large language models (LLMs), it is essential to consider the composite of social perceptions from diverse perspectives among identities. Previous studies have either evaluated biases in LLMs by indirectly assessing the presence of sentiments towards demographic identities in the generated text or measuring the degree of alignment with given stereotypes. These methods have limitations in directly quantifying social biases at the level of distinct perspectives among identities. In this paper, we aim to investigate how social perceptions from various viewpoints contribute to the development of social bias in LLMs. To this end, we propose a novel strategy to intuitively quantify these social perceptions and suggest metrics that can evaluate the social biases within LLMs by aggregating diverse social perceptions. The experimental results show the quantitative demonstration of the social attitude in LLMs by examining social perception. The analysis we conducted shows that our proposed metrics capture the multi-dimensional aspects of social bias, enabling a fine-grained and comprehensive investigation of bias in LLMs.
Measuring Social Bias in Vision-Language Models with Face-Only Counterfactuals from Real Photos
School of Intelligence Science and Engineering, Harbin Institute of Technology (Shenzhen)（哈尔滨工业大学（深圳）智能科学与工程学院）；School of Information Science and Engineering, Ningbo University（宁波大学信息科学与工程学院）；School of Information Science and Technology, University of Science and Technology of China（中国科学技术大学信息科学与技术学院）
Boston University（波士顿大学）；Harvard University（哈佛大学）；Koç University（科奇大学）
Large language models trained on a mixture of NLP tasks that are converted into a text-to-text format using prompts, can generalize into novel forms of language and handle novel tasks. A large body of work within prompt engineering attempts to understand the effects of input forms and prompts in achieving superior performance. We consider an alternative measure and inquire whether the way in which an input is encoded affects social biases promoted in outputs. In this paper, we study T0, a large-scale multi-task text-to-text language model trained using prompt-based learning. We consider two di…
Researchers have devised numerous ways to quantify social biases vested in pretrained language models. As some language models are capable of generating coherent completions given a set of textual prompts, several prompting datasets have been proposed to measure biases between social groups -- posing language generation as a way of identifying biases. In this opinion paper, we analyze how specific choices of prompt sets, metrics, automatic tools and sampling strategies affect bias results. We find out that the practice of measuring biases through text completion is prone to yielding contradict…
Measuring Social Integration Through Participation: Categorizing Organizations and Leisure Activities in the Displaced Karelians Interview Archive using LLMs
University of Turku（图尔库大学）；ELLIS Institute Finland（芬兰ELLIS研究所）
Measuring and Narrowing the Compositionality Gap in Language Models
Paul G. Allen School of Computer Science & Engineering, University of Washington（华盛顿大学保罗·G·艾伦计算机科学与工程学院）；MosaicML（MosaicML公司）；Meta AI Research（Meta人工智能研究院）；Allen Institute for AI（艾伦人工智能研究所）
NormAd: A Framework for Measuring the Cultural Adaptability of Large Language Models
Language Technologies Institute, Carnegie Mellon University（卡内基梅隆大学语言技术研究所）；Paul G. Allen School of Computer Science & Engineering, University of Washington（华盛顿大学保罗·G·艾伦计算机科学与工程学院）
Fairness is a principal social value that can be observed in civilisations around the world. A manifestation of this is in social agreements, often described in texts, such as contracts. Yet, despite the prevalence of such, a fairness metric for texts describing a social act remains wanting. To address this, we take a step back to consider the problem based on first principals. Instead of using rules or templates, we utilise social psychology literature to determine the principal factors that humans use when making a fairness assessment. We then attempt to digitise these using word embeddings …
When writing, a person may need to anticipate questions from their audience, but different social groups may ask very different types of questions. If someone is writing about a problem they want to resolve, what kind of follow-up question will a domain expert ask, and could the writer better address the expert's information needs by rewriting their original post? In this paper, we explore the task of socially-aware question generation. We collect a data set of questions and posts from social media, including background information about the question-askers' social groups. We find that differe…
Johns Hopkins University（约翰斯·霍普金斯大学）；New York University（纽约大学）
The Word Embedding Association Test shows that GloVe and word2vec word embeddings exhibit human-like implicit biases based on gender, race, and other social constructs (Caliskan et al., 2017). Meanwhile, research on learning reusable text representations has begun to explore sentence-level texts, with some sentence encoders seeing enthusiastic adoption. Accordingly, we extend the Word Embedding Association Test to measure bias in sentence encoders. We then test several sentence encoders, including state-of-the-art methods such as ELMo and BERT, for the social biases studied in prior work and t…
The Florida Institute for Human and Machine Cognition（佛罗里达人类与机器认知研究所）；The University of North Carolina at Charlotte（北卡罗来纳大学夏洛特分校）；University of Albany NY（纽约州立大学奥尔巴尼分校）；Rensselaer Polytechnic Institute NY（纽约伦斯勒理工学院）
We present a paradigm for extensible lexicon development based on Lexical Conceptual Structure to support social engineering detection and response generation. We leverage the central notions of ask (elicitation of behaviors such as providing access to money) and framing (risk/reward implied by the ask). We demonstrate improvements in ask/framing detection through refinements to our lexical organization and show that response generation qualitatively improves as ask/framing detection performance improves. The paradigm presents a systematic and efficient approach to resource adaptation for impr…
It has recently been shown that word embeddings encode social biases, with a harmful impact on downstream tasks. However, to this point there has been no similar work done in the field of graph embeddings. We present the first study on social bias in knowledge graph embeddings, and propose a new metric suitable for measuring such bias. We conduct experiments on Wikidata and Freebase, and show that, as with word embeddings, harmful social biases related to professions are encoded in the embeddings with respect to gender, religion, ethnicity and nationality. For example, graph embeddings encode …
Stanford University（斯坦福大学）；Max Planck Institute SWS（马克斯·普朗克软件系统研究所）
Requests are at the core of many social media systems such as question & answer sites and online philanthropy communities. While the success of such requests is critical to the success of the community, the factors that lead community members to satisfy a request are largely unknown. Success of a request depends on factors like who is asking, how they are asking, when are they asking, and most critically what is being requested, ranging from small favors to substantial monetary donations. We present a case study of altruistic requests in an online community where all requests ask for the very …
Pretrained language models, especially masked language models (MLMs) have seen success across many NLP tasks. However, there is ample evidence that they use the cultural biases that are undoubtedly present in the corpora they are trained on, implicitly creating harm with biased representations. To measure some forms of social bias in language models against protected demographic groups in the US, we introduce the Crowdsourced Stereotype Pairs benchmark (CrowS-Pairs). CrowS-Pairs has 1508 examples that cover stereotypes dealing with nine types of bias, like race, religion, and age. In CrowS-Pai…
Instead of studying the properties of social relationship from an objective view, in this paper, we focus on individuals' subjective and asymmetric opinions on their interrelationships. Inspired by the theories from sociolinguistics, we investigate two individuals' opinions on their interrelationship with their interactive language features. Eliminating the difference of personal language style, we clarify that the asymmetry of interactive language feature values can indicate individuals' asymmetric opinions on their interrelationship. We also discuss how the degree of opinions' asymmetry is r…
University of Pennsylvania（宾夕法尼亚大学）；Nanyang Technological University（南洋理工大学）
A body of literature has demonstrated that users' mental health conditions, such as depression and anxiety, can be predicted from their social media language. There is still a gap in the scientific understanding of how psychological stress is expressed on social media. Stress is one of the primary underlying causes and correlates of chronic physical illnesses and mental health conditions. In this paper, we explore the language of psychological stress with a dataset of 601 social media users, who answered the Perceived Stress Scale questionnaire and also consented to share their Facebook and Tw…
School of Information（信息学院）；UC Berkeley（加州大学伯克利分校）
We present the task of modeling information propagation in literature, in which we seek to identify pieces of information passing from character A to character B to character C, only given a description of their activity in text. We describe a new pipeline for measuring information propagation in this domain and publish a new dataset for speaker attribution, enabling the evaluation of an important component of this pipeline on a wider range of literary texts than previously studied. Using this pipeline, we analyze the dynamics of information propagation in over 5,000 works of fiction, finding …
Institute for Human and Machine Cognition (IHMC)（人类与机器认知研究所（IHMC））；State University of New York, Albany（纽约州立大学奥尔巴尼分校）；University of North Carolina, Charlotte（北卡罗来纳大学夏洛特分校）；Rensselaer Polytechnic Institute（伦斯勒理工学院）
Social engineers attempt to manipulate users into undertaking actions such as downloading malware by clicking links or providing access to money or sensitive information. Natural language processing, computational sociolinguistics, and media-specific structural clues provide a means for detecting both the ask (e.g., buy gift card) and the risk/reward implied by the ask, which we call framing (e.g., lose your job, get a raise). We apply linguistic resources such as Lexical Conceptual Structure to tackle ask detection and also leverage structural clues such as links and their proximity to identi…
There are large amounts of insight and social discovery potential in mining crowd-sourced comments left on popular news forums like Reddit.com, Tumblr.com, Facebook.com and Hacker News. Unfortunately, due the overwhelming amount of participation with its varying quality of commentary, extracting value out of such data isn't always obvious nor timely. By designing efficient, single-pass and adaptive natural language filters to quickly prune spam, noise, copy-cats, marketing diversions, and out-of-context posts, we can remove over a third of entries and return the comments with a higher probabil…
Riveter: Measuring Power and Social Dynamics Between Entities
Allen Institute for AI（艾伦人工智能研究所）；Johns Hopkins University（约翰斯·霍普金斯大学）；Carnegie Mellon University（卡内基梅隆大学）；University of Washington（华盛顿大学）；Emory University（埃默里大学）
Knowledge graph question answering (KGQA) facilitates information access by leveraging structured data without requiring formal query language expertise from the user. Instead, users can express their information needs by simply asking their questions in natural language (NL). Datasets used to train KGQA models that would provide such a service are expensive to construct, both in terms of expert and crowdsourced labor. Typically, crowdsourced labor is used to improve template-based pseudo-natural questions generated from formal queries. However, the resulting datasets often fall short of repre…
University of Vermont（佛蒙特大学）；Charles River Analytics（查尔斯河分析公司）
Working from a dataset of 118 billion messages running from the start of 2009 to the end of 2019, we identify and explore the relative daily use of over 150 languages on Twitter. We find that eight languages comprise 80% of all tweets, with English, Japanese, Spanish, and Portuguese being the most dominant. To quantify social spreading in each language over time, we compute the 'contagion ratio': The balance of retweets to organic messages. We find that for the most common languages on Twitter there is a growing tendency, though not universal, to retweet rather than share new content. By the e…
Amazon Alexa AI-NU（亚马逊Alexa人工智能-西北大学合作团队）；UC Santa Barbara（加州大学圣巴巴拉分校）；UCLA（加州大学洛杉矶分校）
Recent advances in deep learning techniques have enabled machines to generate cohesive open-ended text when prompted with a sequence of words as context. While these models now empower many downstream applications from conversation bots to automatic storytelling, they have been shown to generate texts that exhibit social biases. To systematically study and benchmark social biases in open-ended language generation, we introduce the Bias in Open-Ended Language Generation Dataset (BOLD), a large-scale dataset that consists of 23,679 English text generation prompts for bias benchmarking across fiv…
This paper investigates the research question if senders of large amounts of irrelevant or unsolicited information - commonly called "spammers" - distort the network structure of social networks. Two large social networks are analyzed, the first extracted from the Twitter discourse about a big telecommunication company, and the second obtained from three years of email communication of 200 managers working for a large multinational company. This work compares network robustness and the stability of centrality and interaction metrics, as well as the use of language, after removing spammers and …
Sentiment and topic analysis are common methods used for social media monitoring. Essentially, these methods answers questions such as, "what is being talked about, regarding X", and "what do people feel, regarding X". In this paper, we investigate another venue for social media monitoring, namely issue ownership and agenda setting, which are concepts from political science that have been used to explain voter choice and electoral outcomes. We argue that issue alignment and agenda setting can be seen as a kind of semantic source similarity of the kind "how similar is source A to issue owner P,…
Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case …
Machines in the Crowd? Measuring the Footprint of Machine-Generated Text on Reddit
DIMES Dept., University of Calabria Italy；IT University of Copenhagen\ Centre for AI Denmark；DIMES Dept., University of Calabria；IT University of Copenhagen\ Centre for AI
Computational measures of linguistic diversity help us understand the linguistic landscape using digital language data. The contribution of this paper is to calibrate measures of linguistic diversity using restrictions on international travel resulting from the COVID-19 pandemic. Previous work has mapped the distribution of languages using geo-referenced social media and web data. The goal, however, has been to describe these corpora themselves rather than to make inferences about underlying populations. This paper shows that a difference-in-differences method based on the Herfindahl-Hirschman…
Asking effective questions is a powerful social skill. In this paper we seek to build computational models that learn to discriminate effective questions from ineffective ones. Armed with such a capability, future advanced systems can evaluate the quality of questions and provide suggestions for effective question wording. We create a large-scale, real-world dataset that contains over 400,000 questions collected from Reddit "Ask Me Anything" threads. Each thread resembles an online press conference where questions compete with each other for attention from the host. This dataset enables the de…
VisBias: Measuring Explicit and Implicit Social Biases in Vision Language Models
University of Southern California（南加州大学）；University of Bristol（布里斯托大学）；Centre National de la Recherche Scientifique（法国国家科学研究中心）；University of California, Los Angeles（加州大学洛杉矶分校）
Measuring Dimensions of Self-Presentation in Twitter Bios and their Links to Misinformation Sharing
University at Buffalo（布法罗大学）；GESIS Leibniz Institute for the Social Sciences（GESIS莱布尼茨社会科学研究所）；University of Pittsburgh（匹兹堡大学）；George Washington University（乔治华盛顿大学）；Northeastern University（东北大学）
