Large language models (LLMs) have demonstrated the potential to mimic human social intelligence. However, most studies focus on simplistic and static self-report or performance-based tests, which limits the depth and validity of the analysis. In this paper, we developed a novel framework, InterIntent, to assess LLMs’ social intelligence by mapping their ability to understand and manage intentions in a game setting. We focus on four dimensions of social intelligence: situational awareness, self-regulation, self-awareness, and theory of mind. Each dimension is linked to a specific game task: intention selection, intention following, intention summarization, and intention guessing. Our findings indicate that while LLMs exhibit high proficiency in selecting intentions, achieving an accuracy of 88%, their ability to infer the intentions of others is significantly weaker, trailing human performance by 20%. Additionally, game performance correlates with intention understanding, highlighting the importance of the four components towards success in this game. These findings underline the crucial role of intention understanding in evaluating LLMs’ social intelligence and highlight the potential of using social deduction games as a complex testbed to enhance LLM evaluation. InterIntent contributes a structured approach to bridging the evaluation gap in social intelligence within multiplayer LLM-based games.

Social science research has shown that candidates with names indicative of certain races or genders often face discrimination in employment practices. Similarly, Large Language Models (LLMs) have demonstrated racial and gender biases in various applications. In this study, we utilize GPT-3.5-Turbo and Llama 3-70B-Instruct to simulate hiring decisions and salary recommendations for candidates with 320 first names that strongly signal their race and gender, across over 750,000 prompts. Our empirical results indicate a preference among these models for hiring candidates with White female-sounding names over other demographic groups across 40 occupations. Additionally, even among candidates with identical qualifications, salary recommendations vary by as much as 5% between different subgroups. A comparison with real-world labor data reveals inconsistent alignment with U.S. labor market characteristics, underscoring the necessity of risk investigation of LLM-powered systems.

Vision-language models (VLMs) pre-trained on extensive datasets can inadvertently learn biases by correlating gender information with specific objects or scenarios. Current methods, which focus on modifying inputs and monitoring changes in the model’s output probability scores, often struggle to comprehensively understand bias from the perspective of model components. We propose a framework that incorporates causal mediation analysis to measure and map the pathways of bias generation and propagation within VLMs. Our framework is applicable to a wide range of vision-language and multimodal tasks. In this work, we apply it to the object detection task and implement it on the GLIP model. This approach allows us to identify the direct effects of interventions on model bias and the indirect effects of interventions on bias mediated through different model components. Our results show that image features are the primary contributors to bias, with significantly higher impacts than text features, specifically accounting for 32.57% and 12.63% of the bias in the MSCOCO and PASCAL-SENTENCE datasets, respectively. Notably, the image encoder’s contribution surpasses that of the text encoder and the deep fusion encoder. Further experimentation confirms that contributions from both language and vision modalities are aligned and non-conflicting. Consequently, focusing on blurring gender representations within the image encoder which contributes most to the model bias, reduces bias efficiently by 22.03% and 9.04% in the MSCOCO and PASCAL-SENTENCE datasets, respectively, with minimal performance loss or increased computational demands.

In a plethora of recent work, large language models (LLMs) demonstrated impressive reasoning ability, but many proposed downstream reasoning tasks only focus on performance-wise evaluation. Two fundamental questions persist: 1) how consistent is the reasoning, and 2) can models detect unreliable reasoning? In this paper, we investigate self-contradictory (Self-Contra) reasoning, where the model reasoning does not support answers. To answer 1), we define and assess the Self-Contra rate across three datasets and delve into finer-grained categories of Self-Contra reasoning. We find that LLMs often contradict themselves in reasoning tasks involving contextual information understanding or commonsense. The model may generate correct answers by taking shortcuts in reasoning or overlooking contextual evidence, leading to compromised reasoning. For 2), we task the state-of-the-art model GPT-4 with identifying Self-Contra reasoning and finer-grained fallacies. We find that finer-grained aided detection can improve GPT-4’s ability to detect Self-Contra. However, it is only able to detect Self-Contra with a 52.2% F1 score, much lower compared to 66.7% for humans. Our results indicate that current LLMs lack the robustness necessary for reliable reasoning and we emphasize the urgent need for establishing best practices in comprehensive reasoning evaluations beyond pure performance-based metrics.

People from different social and demographic groups express diverse perspectives and conflicting opinions on a broad set of topics such as product reviews, healthcare, law, and politics. A fair summary should provide a comprehensive coverage of diverse perspectives without underrepresenting certain groups. However, current work in summarization metrics and Large Language Models (LLMs) evaluation has not explored fair abstractive summarization. In this paper, we systematically investigate fair abstractive summarization for user-generated data. We first formally define fairness in abstractive summarization as not underrepresenting perspectives of any groups of people, and we propose four reference-free automatic metrics by measuring the differences between target and source perspectives. We evaluate nine LLMs, including three GPT models, four LLaMA models, PaLM 2, and Claude, on six datasets collected from social media, online reviews, and recorded transcripts. Experiments show that both the model-generated and the human-written reference summaries suffer from low fairness. We conduct a comprehensive analysis of the common factors influencing fairness and propose three simple but effective methods to alleviate unfair summarization. Our dataset and code are available at https://github.com/psunlpgroup/FairSumm.

Reinforcement learning from human feedback (RLHF) is a vital strategy for enhancing model capability in language models. However, annotating preference data for RLHF is a resource-intensive and creativity-demanding process, while existing automatic generation methods face limitations in data diversity and quality. In response, we present Safer-Instruct, a novel pipeline for automatically constructing large-scale preference data. Our approach leverages reversed instruction tuning, instruction induction, and expert model evaluation to efficiently generate high-quality preference data without human annotators. To verify the effectiveness of Safer-Instruct, we apply the pipeline to construct a safety preference dataset as a case study. Finetuning an Alpaca model on this synthetic dataset not only demonstrates improved harmlessness but also outperforms models fine-tuned on human-annotated safety preference data, all the while maintaining a competitive edge in downstream tasks. Importantly, our Safer-Instruct framework is versatile and can be applied to generate preference data across various domains, extending its utility beyond safety preferences. It addresses the challenges in preference data acquisition and advances the development of more capable and responsible AI systems. For dataset and code implementation, see https://github.com/uscnlp-lime/safer-instruct/.

In English speaking assessment, pretrained large language models (LLMs) such as BERT can score constructed response items as accurately as human raters. Less research has investigated whether LLMs perpetuate or exacerbate biases, which would pose problems for the fairness and validity of the test. This study examines gender and native language (L1) biases in human and automated scores, using an off-the-shelf (OOS) BERT model. Analyses focus on a specific type of bias known as differential item functioning (DIF), which compares examinees of similar English language proficiency. Results show that there is a moderate amount of DIF, based on examinees’ L1 background in grade band 912. DIF is higher when scored by an OOS BERT model, indicating that BERT may exacerbate this bias; however, in practical terms, the degree to which BERT exacerbates DIF is very small. Additionally, there is more DIF for longer speaking items and for older examinees, but BERT does not exacerbate these patterns of DIF.

A common limitation of diagnostic tests for detecting social biases in NLP models is that they may only detect stereotypic associations that are pre-specified by the designer of the test. Since enumerating all possible problematic associations is infeasible, it is likely these tests fail to detect biases that are present in a model but not pre-specified by the designer. To address this limitation, we propose SODAPOP (SOcial bias Discovery from Answers about PeOPle), an approach for automatic social bias discovery in social commonsense question-answering. The SODAPOP pipeline generates modified instances from the Social IQa dataset (Sap et al., 2019b) by (1) substituting names associated with different demographic groups, and (2) generating many distractor answers from a masked language model. By using a social commonsense model to score the generated distractors, we are able to uncover the model’s stereotypic associations between demographic groups and an open set of words. We also test SODAPOP on debiased models and show the limitations of multiple state-of-the-art debiasing algorithms.

We ask the question: Are there widespread disparities in machine translations of names across race/ethnicity, and gender? We hypothesize that the translation quality of names and surrounding context will be lower for names associated with US racial and ethnic minorities due to these systems’ tendencies to standardize language to predominant language patterns. We develop a dataset of names that are strongly demographically aligned and propose a translation evaluation procedure based on round-trip translation. We analyze the effect of name demographics on translation quality using generalized linear mixed effects models and find that the ability of translation systems to correctly translate female-associated names is significantly lower than male-associated names. This effect is particularly pronounced for female-associated names that are also associated with racial (Black) and ethnic (Hispanic) minorities. This disparity in translation quality between social groups for something as personal as someone’s name has significant implications for people’s professional, personal, and cultural identities, self-worth and ease of communication. Our findings suggest that more MT research is needed to improve the translation of names and to provide high-quality service for users regardless of gender, race, and ethnicity.

Recent advancements in Large Language Models empower them to follow freeform instructions, including imitating generic or specific demographic personas in conversations. We define generic personas to represent demographic groups, such as “an Asian person”, whereas specific personas may take the form of specific popular Asian names like “Yumi”. While the adoption of personas enriches user experiences by making dialogue systems more engaging and approachable, it also casts a shadow of potential risk by exacerbating social biases within model responses, thereby causing societal harm through interactions with users. In this paper, we systematically study “persona biases”, which we define to be the sensitivity of dialogue models’ harmful behaviors contingent upon the personas they adopt. We categorize persona biases into biases in harmful expression and harmful agreement, and establish a comprehensive evaluation framework to measure persona biases in five aspects: Offensiveness, Toxic Continuation, Regard, Stereotype Agreement, and Toxic Agreement. Additionally, we propose to investigate persona biases by experimenting with UNIVERSALPERSONA, a systematically constructed persona dataset encompassing various types of both generic and specific model personas. Through benchmarking on four different models- including Blender, ChatGPT, Alpaca, and Vicuna- our study uncovers significant persona biases in dialogue systems. Our findings also underscore the pressing need to revisit the use of personas in dialogue agents to ensure safe application.

Recent advances in natural language processing and transformer-based models have made it easier to implement accurate, automated English speech assessments. Yet, without careful examination, applications of these models may exacerbate social prejudices based on gender and race. This study addresses the need to examine potential biases of transformer-based models in the context of automated English speech assessment. For this purpose, we developed a BERT-based automated speech assessment system and investigated gender and racial bias of examinees’ automated scores. Gender and racial bias was measured by examining differential item functioning (DIF) using an item response theory framework. Preliminary results, which focused on a single verbal-response item, showed no statistically significant DIF based on gender or race for automated scores.

Recent studies show that Natural Language Processing (NLP) technologies propagate societal biases about demographic groups associated with attributes such as gender, race, and nationality. To create interventions and mitigate these biases and associated harms, it is vital to be able to detect and measure such biases. While existing works propose bias evaluation and mitigation methods for various tasks, there remains a need to cohesively understand the biases and the specific harms they measure, and how different measures compare with each other. To address this gap, this work presents a practical framework of harms and a series of questions that practitioners can answer to guide the development of bias measures. As a validation of our framework and documentation questions, we also present several case studies of how existing bias measures in NLP—both intrinsic measures of bias in representations and extrinsic measures of bias of downstream applications—can be aligned with different harms and how our proposed documentation questions facilitates more holistic understanding of what bias measures are measuring.

Large pre-trained language models have shown remarkable performance over the past few years. These models, however, sometimes learn superficial features from the dataset and cannot generalize to the distributions that are dissimilar to the training scenario. There have been several approaches proposed to reduce model’s reliance on these bias features which can improve model robustness in the out-of-distribution setting. However, existing methods usually use a fixed low-capacity model to deal with various bias features, which ignore the learnability of those features. In this paper, we analyze a set of existing bias features and demonstrate there is no single model that works best for all the cases. We further show that by choosing an appropriate bias model, we can obtain a better robustness result than baselines with a more sophisticated model design.

Robustness and counterfactual bias are usually evaluated on a test dataset. However, are these evaluations robust? If the test dataset is perturbed slightly, will the evaluation results keep the same? In this paper, we propose a “double perturbation” framework to uncover model weaknesses beyond the test dataset. The framework first perturbs the test dataset to construct abundant natural sentences similar to the test data, and then diagnoses the prediction change regarding a single-word substitution. We apply this framework to study two perturbation-based approaches that are used to analyze models’ robustness and counterfactual bias in English. (1) For robustness, we focus on synonym substitutions and identify vulnerable examples where prediction can be altered. Our proposed attack attains high success rates (96.0%-99.8%) in finding vulnerable examples on both original and robustly trained CNNs and Transformers. (2) For counterfactual bias, we focus on substituting demographic tokens (e.g., gender, race) and measure the shift of the expected prediction among constructed sentences. Our method is able to reveal the hidden model biases not directly shown in the test dataset. Our code is available at https://github.com/chong-z/nlp-second-order-attack.

Humor plays an important role in human languages and it is essential to model humor when building intelligence systems. Among different forms of humor, puns perform wordplay for humorous effects by employing words with double entendre and high phonetic similarity. However, identifying and modeling puns are challenging as puns usually involved implicit semantic or phonological tricks. In this paper, we propose Pronunciation-attentive Contextualized Pun Recognition (PCPR) to perceive human humor, detect if a sentence contains puns and locate them in the sentence. PCPR derives contextualized representation for each word in a sentence by capturing the association between the surrounding context and its corresponding phonetic symbols. Extensive experiments are conducted on two benchmark datasets. Results demonstrate that the proposed approach significantly outperforms the state-of-the-art methods in pun detection and location tasks. In-depth analyses verify the effectiveness and robustness of PCPR.

Multilingual representations embed words from many languages into a single semantic space such that words with similar meanings are close to each other regardless of the language. These embeddings have been widely used in various settings, such as cross-lingual transfer, where a natural language processing (NLP) model trained on one language is deployed to another language. While the cross-lingual transfer techniques are powerful, they carry gender bias from the source to target languages. In this paper, we study gender bias in multilingual embeddings and how it affects transfer learning for NLP applications. We create a multilingual dataset for bias analysis and propose several ways for quantifying bias in multilingual representations from both the intrinsic and extrinsic perspectives. Experimental results show that the magnitude of bias in the multilingual representations changes differently when we align the embeddings to different target spaces and that the alignment direction can also have an influence on the bias in transfer learning. We further provide recommendations for using the multilingual word representations for downstream tasks.

Advanced machine learning techniques have boosted the performance of natural language processing. Nevertheless, recent studies, e.g., (CITATION) show that these techniques inadvertently capture the societal bias hidden in the corpus and further amplify it. However, their analysis is conducted only on models’ top predictions. In this paper, we investigate the gender bias amplification issue from the distribution perspective and demonstrate that the bias is amplified in the view of predicted probability distribution over labels. We further propose a bias mitigation approach based on posterior regularization. With little performance loss, our method can almost remove the bias amplification in the distribution. Our study sheds the light on understanding the bias amplification.

Recent developments in Neural Relation Extraction (NRE) have made significant strides towards Automated Knowledge Base Construction. While much attention has been dedicated towards improvements in accuracy, there have been no attempts in the literature to evaluate social biases exhibited in NRE systems. In this paper, we create WikiGenderBias, a distantly supervised dataset composed of over 45,000 sentences including a 10% human annotated test set for the purpose of analyzing gender bias in relation extraction systems. We find that when extracting spouse-of and hypernym (i.e., occupation) relations, an NRE system performs differently when the gender of the target entity is different. However, such disparity does not appear when extracting relations such as birthDate or birthPlace. We also analyze how existing bias mitigation techniques, such as name anonymization, word embedding debiasing, and data augmentation affect the NRE system in terms of maintaining the test performance and reducing biases. Unfortunately, due to NRE models rely heavily on surface level cues, we find that existing bias mitigation approaches have a negative effect on NRE. Our analysis lays groundwork for future quantifying and mitigating bias in NRE.

Machine learning techniques have been widely used in natural language processing (NLP). However, as revealed by many recent studies, machine learning models often inherit and amplify the societal biases in data. Various metrics have been proposed to quantify biases in model predictions. In particular, several of them evaluate disparity in model performance between protected groups and advantaged groups in the test corpus. However, we argue that evaluating bias at the corpus level is not enough for understanding how biases are embedded in a model. In fact, a model with similar aggregated performance between different groups on the entire data may behave differently on instances in a local region. To analyze and detect such local bias, we propose LOGAN, a new bias detection technique based on clustering. Experiments on toxicity classification and object classification tasks show that LOGAN identifies bias in a local region and allows us to better analyze the biases in model predictions.

Recent studies have shown that word embeddings exhibit gender bias inherited from the training corpora. However, most studies to date have focused on quantifying and mitigating such bias only in English. These analyses cannot be directly extended to languages that exhibit morphological agreement on gender, such as Spanish and French. In this paper, we propose new metrics for evaluating gender bias in word embeddings of these languages and further demonstrate evidence of gender bias in bilingual embeddings which align these languages with English. Finally, we extend an existing approach to mitigate gender bias in word embedding of these languages under both monolingual and bilingual settings. Experiments on modified Word Embedding Association Test, word similarity, word translation, and word pair translation tasks show that the proposed approaches can effectively reduce the gender bias while preserving the utility of the original embeddings.

In this paper, we quantify, analyze and mitigate gender bias exhibited in ELMo’s contextualized word vectors. First, we conduct several intrinsic analyses and find that (1) training data for ELMo contains significantly more male than female entities, (2) the trained ELMo embeddings systematically encode gender information and (3) ELMo unequally encodes gender information about male and female entities. Then, we show that a state-of-the-art coreference system that depends on ELMo inherits its bias and demonstrates significant bias on the WinoBias probing corpus. Finally, we explore two methods to mitigate such gender bias and show that the bias demonstrated on WinoBias can be eliminated.

As Natural Language Processing (NLP) and Machine Learning (ML) tools rise in popularity, it becomes increasingly vital to recognize the role they play in shaping societal biases and stereotypes. Although NLP models have shown success in modeling various applications, they propagate and may even amplify gender bias found in text corpora. While the study of bias in artificial intelligence is not new, methods to mitigate gender bias in NLP are relatively nascent. In this paper, we review contemporary studies on recognizing and mitigating gender bias in NLP. We discuss gender bias based on four forms of representation bias and analyze methods recognizing gender bias. Furthermore, we discuss the advantages and drawbacks of existing gender debiasing methods. Finally, we discuss future studies for recognizing and mitigating gender bias in NLP.

Word embedding models have become a fundamental component in a wide range of Natural Language Processing (NLP) applications. However, embeddings trained on human-generated corpora have been demonstrated to inherit strong gender stereotypes that reflect social constructs. To address this concern, in this paper, we propose a novel training procedure for learning gender-neutral word embeddings. Our approach aims to preserve gender information in certain dimensions of word vectors while compelling other dimensions to be free of gender influence. Based on the proposed method, we generate a Gender-Neutral variant of GloVe (GN-GloVe). Quantitative and qualitative experiments demonstrate that GN-GloVe successfully isolates gender information without sacrificing the functionality of the embedding model.

In this paper, we introduce a new benchmark for co-reference resolution focused on gender bias, WinoBias. Our corpus contains Winograd-schema style sentences with entities corresponding to people referred by their occupation (e.g. the nurse, the doctor, the carpenter). We demonstrate that a rule-based, a feature-rich, and a neural coreference system all link gendered pronouns to pro-stereotypical entities with higher accuracy than anti-stereotypical entities, by an average difference of 21.1 in F1 score. Finally, we demonstrate a data-augmentation approach that, in combination with existing word-embedding debiasing techniques, removes the bias demonstrated by these systems in WinoBias without significantly affecting their performance on existing datasets.

Language is increasingly being used to de-fine rich visual recognition problems with supporting image collections sourced from the web. Structured prediction models are used in these tasks to take advantage of correlations between co-occurring labels and visual input but risk inadvertently encoding social biases found in web corpora. In this work, we study data and models associated with multilabel object classification and visual semantic role labeling. We find that (a) datasets for these tasks contain significant gender bias and (b) models trained on these datasets further amplify existing bias. For example, the activity cooking is over 33% more likely to involve females than males in a training set, and a trained model further amplifies the disparity to 68% at test time. We propose to inject corpus-level constraints for calibrating existing structured prediction models and design an algorithm based on Lagrangian relaxation for collective inference. Our method results in almost no performance loss for the underlying recognition task but decreases the magnitude of bias amplification by 47.5% and 40.5% for multilabel classification and visual semantic role labeling, respectively。