This work provides an explanatory view of how LLMs can apply moral reasoning to both criticize and defend sexist language. We assessed eight large language models, all of which demonstrated the capability to provide explanations grounded in varying moral perspectives for both critiquing and endorsing views that reflect sexist assumptions. With both human and automatic evaluation, we show that all eight models produce comprehensible and contextually relevant text, which is helpful in understanding diverse views on how sexism is perceived. Also, through analysis of moral foundations cited by LLMs in their arguments, we uncover the diverse ideological perspectives in models’ outputs, with some models aligning more with progressive or conservative views on gender roles and sexism.Based on our observations, we caution against the potential misuse of LLMs to justify sexist language. We also highlight that LLMs can serve as tools for understanding the roots of sexist beliefs and designing well-informed interventions. Given this dual capacity, it is crucial to monitor LLMs and design safety mechanisms for their use in applications that involve sensitive societal topics, such as sexism.

Negative public perceptions of people living in poverty can hamper policies and programs that aim to help the poor. One prominent example of social bias and discrimination against people in need is the persistent association of poverty with criminality. The phenomenon has two facets: first, the belief that poor people are more likely to engage in crime (e.g., stealing, mugging, violence) and second, the view that certain behaviors directly resulting from poverty (e.g., living outside, panhandling) warrant criminal punishment. In this paper, we use large language models (LLMs) to identify examples of crime–poverty association (CPA) in English social media texts. We analyze the online discourse on CPA across eight geographically-diverse countries, and find evidence that the CPA rates are higher within the sample obtained from the U.S. and Canada, as compared to the other countries such as South Africa, despite the latter having higher poverty, criminality, and inequality indexes. We further uncover and analyze the most common themes in CPA posts and find more negative and biased attitudes toward people living in poverty in posts from the U.S. and Canada. These results could partially be explained by cultural factors related to the tendency to overestimate the equality of opportunities and social mobility in the U.S. and Canada. These findings have consequences for policy-making and open a new path of research for poverty mitigation with the focus not only on the redistribution of wealth but also on the mitigation of bias and discrimination against people in need.

For decades, psychologists have been studying stereotypes using specially-designed rating scales to capture people’s beliefs and opinions about different social groups. Now, using NLP tools on extensive collections of text, we have the opportunity to study stereotypes “in the wild” and on a large scale. However, are we truly capturing the same information? In this paper we compare measurements along six psychologically-motivated, stereotype-relevant dimensions (Sociability, Morality, Ability, Assertiveness, Beliefs, and Status) for 10 groups, defined by occupation. We compute these measurements on stereotypical English sentences written by crowd-workers, stereotypical sentences generated by ChatGPT, and more general data collected from social media, and contrast the findings with traditional, survey-based results, as well as a spontaneous word-list generation task. We find that while the correlation with the traditional scales varies across dimensions, the free-text data can be used to specify the particular traits associated with each group, and provide context for numerical survey data.

Following on recent advances in large language models (LLMs) and subsequent chat models, a new wave of large vision–language models (LVLMs) has emerged. Such models can incorporate images as input in addition to text, and perform tasks such as visual question answering, image captioning, story generation, etc. Here, we examine potential gender and racial biases in such systems, based on the perceived characteristics of the people in the input images. To accomplish this, we present a new dataset PAIRS (PArallel Images for eveRyday Scenarios). The PAIRS dataset contains sets of AI-generated images of people, such that the images are highly similar in terms of background and visual content, but differ along the dimensions of gender (man, woman) and race (Black, white). By querying the LVLMs with such images, we observe significant differences in the responses according to the perceived gender or race of the person depicted.

Gender stereotypes are pervasive beliefs about individuals based on their gender that play a significant role in shaping societal attitudes, behaviours, and even opportunities. Recognizing the negative implications of gender stereotypes, particularly in online communications, this study investigates eleven strategies to automatically counteract and challenge these views. We present AI-generated gender-based counter-stereotypes to (self-identified) male and female study participants and ask them to assess their offensiveness, plausibility, and potential effectiveness. The strategies of counter-facts and broadening universals (i.e., stating that anyone can have a trait regardless of group membership) emerged as the most robust approaches, while humour, perspective-taking, counter-examples, and empathy for the speaker were perceived as less effective. Also, the differences in ratings were more pronounced for stereotypes about the different targets than between the genders of the raters. Alarmingly, many AI-generated counter-stereotypes were perceived as offensive and/or implausible. Our analysis and the collected dataset offer foundational insight into counter-stereotype generation, guiding future efforts to develop strategies that effectively challenge gender stereotypes in online interactions.

When harmful social stereotypes are expressed on a public platform, they must be addressed in a way that educates and informs both the original poster and other readers, without causing offence or perpetuating new stereotypes. In this paper, we synthesize findings from psychology and computer science to propose a set of potential counter-stereotype strategies. We then automatically generate such counter-stereotypes using ChatGPT, and analyze their correctness and expected effectiveness at reducing stereotypical associations. We identify the strategies of denouncing stereotypes, warning of consequences, and using an empathetic tone as three promising strategies to be further tested.

While many types of hate speech and online toxicity have been the focus of extensive research in NLP, toxic language stigmatizing poor people has been mostly disregarded. Yet, aporophobia, a social bias against the poor, is a common phenomenon online, which can be psychologically damaging as well as hindering poverty reduction policy measures. We demonstrate that aporophobic attitudes are indeed present in social media and argue that the existing NLP datasets and models are inadequate to effectively address this problem. Efforts toward designing specialized resources and novel socio-technical mechanisms for confronting aporophobia are needed.

Classifiers tend to learn a false causal relationship between an over-represented concept and a label, which can result in over-reliance on the concept and compromised classification accuracy. It is imperative to have methods in place that can compare different models and identify over-reliances on specific concepts. We consider three well-known abusive language classifiers trained on large English datasets and focus on the concept of negative emotions, which is an important signal but should not be learned as a sufficient feature for the label of abuse. Motivated by the definition of global sufficiency, we first examine the unwanted dependencies learned by the classifiers by assessing their accuracy on a challenge set across all decision thresholds. Further, recognizing that a challenge set might not always be available, we introduce concept-based explanation metrics to assess the influence of the concept on the labels. These explanations allow us to compare classifiers regarding the degree of false global sufficiency they have learned between a concept and a label.

We present a novel feature attribution method for explaining text classifiers, and analyze it in the context of hate speech detection. Although feature attribution models usually provide a single importance score for each token, we instead provide two complementary and theoretically-grounded scores – necessity and sufficiency – resulting in more informative explanations. We propose a transparent method that calculates these values by generating explicit perturbations of the input text, allowing the importance scores themselves to be explainable. We employ our method to explain the predictions of different hate speech detection models on the same set of curated examples from a test suite, and show that different values of necessity and sufficiency for identity terms correspond to different kinds of false positive errors, exposing sources of classifier bias against marginalized groups.

Previous works on the fairness of toxic language classifiers compare the output of models with different identity terms as input features but do not consider the impact of other important concepts present in the context. Here, besides identity terms, we take into account high-level latent features learned by the classifier and investigate the interaction between these features and identity terms. For a multi-class toxic language classifier, we leverage a concept-based explanation framework to calculate the sensitivity of the model to the concept of sentiment, which has been used before as a salient feature for toxic language detection. Our results show that although for some classes, the classifier has learned the sentiment information as expected, this information is outweighed by the influence of identity terms as input features. This work is a step towards evaluating procedural fairness, where unfair processes lead to unfair outcomes. The produced knowledge can guide debiasing techniques to ensure that important concepts besides identity terms are well-represented in training datasets.

Robustness of machine learning models on ever-changing real-world data is critical, especially for applications affecting human well-being such as content moderation. New kinds of abusive language continually emerge in online discussions in response to current events (e.g., COVID-19), and the deployed abuse detection systems should be updated regularly to remain accurate. In this paper, we show that general abusive language classifiers tend to be fairly reliable in detecting out-of-domain explicitly abusive utterances but fail to detect new types of more subtle, implicit abuse. Next, we propose an interpretability technique, based on the Testing Concept Activation Vector (TCAV) method from computer vision, to quantify the sensitivity of a trained model to the human-defined concepts of explicit and implicit abusive language, and use that to explain the generalizability of the model on new data, in this case, COVID-related anti-Asian hate speech. Extending this technique, we introduce a novel metric, Degree of Explicitness, for a single instance and show that the new metric is beneficial in suggesting out-of-domain unlabeled examples to effectively enrich the training data with informative, implicitly abusive texts.

In an effort to guarantee that machine learning model outputs conform with human moral values, recent work has begun exploring the possibility of explicitly training models to learn the difference between right and wrong. This is typically done in a bottom-up fashion, by exposing the model to different scenarios, annotated with human moral judgements. One question, however, is whether the trained models actually learn any consistent, higher-level ethical principles from these datasets – and if so, what? Here, we probe the Allen AI Delphi model with a set of standardized morality questionnaires, and find that, despite some inconsistencies, Delphi tends to mirror the moral principles associated with the demographic groups involved in the annotation process. We question whether this is desirable and discuss how we might move forward with this knowledge.

Motivations for methods in explainable artificial intelligence (XAI) often include detecting, quantifying and mitigating bias, and contributing to making machine learning models fairer. However, exactly how an XAI method can help in combating biases is often left unspecified. In this paper, we briefly review trends in explainability and fairness in NLP research, identify the current practices in which explainability methods are applied to detect and mitigate bias, and investigate the barriers preventing XAI methods from being used more widely in tackling fairness issues.

Age-related stereotypes are pervasive in our society, and yet have been under-studied in the NLP community. Here, we present a method for extracting age-related stereotypes from Twitter data, generating a corpus of 300,000 over-generalizations about four contemporary generations (baby boomers, generation X, millennials, and generation Z), as well as “old” and “young” people more generally. By employing word-association metrics, semi-supervised topic modelling, and density-based clustering, we uncover many common stereotypes as reported in the media and in the psychological literature, as well as some more novel findings. We also observe trends consistent with the existing literature, namely that definitions of “young” and “old” age appear to be context-dependent, stereotypes for different generations vary across different topics (e.g., work versus family life), and some age-based stereotypes are distinct from generational stereotypes. The method easily extends to other social group labels, and therefore can be used in future work to study stereotypes of different social categories. By better understanding how stereotypes are formed and spread, and by tracking emerging stereotypes, we hope to eventually develop mitigating measures against such biased statements.

Stereotypical language expresses widely-held beliefs about different social categories. Many stereotypes are overtly negative, while others may appear positive on the surface, but still lead to negative consequences. In this work, we present a computational approach to interpreting stereotypes in text through the Stereotype Content Model (SCM), a comprehensive causal theory from social psychology. The SCM proposes that stereotypes can be understood along two primary dimensions: warmth and competence. We present a method for defining warmth and competence axes in semantic embedding space, and show that the four quadrants defined by this subspace accurately represent the warmth and competence concepts, according to annotated lexicons. We then apply our computational SCM model to textual stereotype data and show that it compares favourably with survey-based studies in the psychological literature. Furthermore, we explore various strategies to counter stereotypical beliefs with anti-stereotypes. It is known that countering stereotypes with anti-stereotypical examples is one of the most effective ways to reduce biased thinking, yet the problem of generating anti-stereotypes has not been previously studied. Thus, a better understanding of how to generate realistic and effective anti-stereotypes can contribute to addressing pressing societal concerns of stereotyping, prejudice, and discrimination.

The state of being alone can have a substantial impact on our lives, though experiences with time alone diverge significantly among individuals. Psychologists distinguish between the concept of solitude, a positive state of voluntary aloneness, and the concept of loneliness, a negative state of dissatisfaction with the quality of one’s social interactions. Here, for the first time, we conduct a large-scale computational analysis to explore how the terms associated with the state of being alone are used in online language. We present SOLO (State of Being Alone), a corpus of over 4 million tweets collected with query terms solitude, lonely, and loneliness. We use SOLO to analyze the language and emotions associated with the state of being alone. We show that the term solitude tends to co-occur with more positive, high-dominance words (e.g., enjoy, bliss) while the terms lonely and loneliness frequently co-occur with negative, low-dominance words (e.g., scared, depressed), which confirms the conceptual distinctions made in psychology. We also show that women are more likely to report on negative feelings of being lonely as compared to men, and there are more teenagers among the tweeters that use the word lonely than among the tweeters that use the word solitude.

NLP research has attained high performances in abusive language detection as a supervised classification task. While in research settings, training and test datasets are usually obtained from similar data samples, in practice systems are often applied on data that are different from the training set in topic and class distributions. Also, the ambiguity in class definitions inherited in this task aggravates the discrepancies between source and target datasets. We explore the topic bias and the task formulation bias in cross-dataset generalization. We show that the benign examples in the Wikipedia Detox dataset are biased towards platform-specific topics. We identify these examples using unsupervised topic modeling and manual inspection of topics’ keywords. Removing these topics increases cross-dataset generalization, without reducing in-domain classification performance. For a robust dataset design, we suggest applying inexpensive unsupervised methods to inspect the collected data and downsize the non-generalizable content before manually annotating for class labels.

Bigrams (two-word sequences) hold a special place in semantic composition research since they are the smallest unit formed by composing words. A semantic relatedness dataset that includes bigrams will thus be useful in the development of automatic methods of semantic composition. However, existing relatedness datasets only include pairs of unigrams (single words). Further, existing datasets were created using rating scales and thus suffer from limitations such as in consistent annotations and scale region bias. In this paper, we describe how we created a large, fine-grained, bigram relatedness dataset (BiRD), using a comparative annotation technique called Best–Worst Scaling. Each of BiRD’s 3,345 English term pairs involves at least one bigram. We show that the relatedness scores obtained are highly reliable (split-half reliability r= 0.937). We analyze the data to obtain insights into bigram semantic relatedness. Finally, we present benchmark experiments on using the relatedness dataset as a testbed to evaluate simple unsupervised measures of semantic composition. BiRD is made freely available to foster further research on how meaning can be represented and how meaning can be composed.

We present the SemEval-2018 Task 1: Affect in Tweets, which includes an array of subtasks on inferring the affectual state of a person from their tweet. For each task, we created labeled data from English, Arabic, and Spanish tweets. The individual tasks are: 1. emotion intensity regression, 2. emotion intensity ordinal classification, 3. valence (sentiment) regression, 4. valence ordinal classification, and 5. emotion classification. Seventy-five teams (about 200 team members) participated in the shared task. We summarize the methods, resources, and tools used by the participating teams, with a focus on the techniques and resources that are particularly useful. We also analyze systems for consistent bias towards a particular race or gender. The data is made freely available to further improve our understanding of how people convey emotions through language.

In this paper, we propose a regression system to infer the emotion intensity of a tweet. We develop a multi-aspect feature learning mechanism to capture the most discriminative semantic features of a tweet as well as the emotion information conveyed by each word in it. We combine six types of feature groups: (1) a tweet representation learned by an LSTM deep neural network on the training data, (2) a tweet representation learned by an LSTM network on a large corpus of tweets that contain emotion words (a distant supervision corpus), (3) word embeddings trained on the distant supervision corpus and averaged over all words in a tweet, (4) word and character n-grams, (5) features derived from various sentiment and emotion lexicons, and (6) other hand-crafted features. As part of the word embedding training, we also learn the distributed representations of multi-word expressions (MWEs) and negated forms of words. An SVR regressor is then trained over the full set of features. We evaluate the effectiveness of our ensemble feature sets on the SemEval-2018 Task 1 datasets and achieve a Pearson correlation of 72% on the task of tweet emotion intensity prediction.

Automatic machine learning systems can inadvertently accentuate and perpetuate inappropriate human biases. Past work on examining inappropriate biases has largely focused on just individual systems. Further, there is no benchmark dataset for examining inappropriate biases in systems. Here for the first time, we present the Equity Evaluation Corpus (EEC), which consists of 8,640 English sentences carefully chosen to tease out biases towards certain races and genders. We use the dataset to examine 219 automatic sentiment analysis systems that took part in a recent shared task, SemEval-2018 Task 1 ‘Affect in Tweets’. We find that several of the systems show statistically significant bias; that is, they consistently provide slightly higher sentiment intensity predictions for one race or one gender. We make the EEC freely available.

Being able to predict whether people agree or disagree with an assertion (i.e. an explicit, self-contained statement) has several applications ranging from predicting how many people will like or dislike a social media post to classifying posts based on whether they are in accordance with a particular point of view. We formalize this as two NLP tasks: predicting judgments of (i) individuals and (ii) groups based on the text of the assertion and previous judgments. We evaluate a wide range of approaches on a crowdsourced data set containing over 100,000 judgments on over 2,000 assertions. We find that predicting individual judgments is a hard task with our best results only slightly exceeding a majority baseline, but that judgments of groups can be more reliably predicted using a Siamese neural network, which outperforms all other approaches by a wide margin.

Rating scales are a widely used method for data annotation; however, they present several challenges, such as difficulty in maintaining inter- and intra-annotator consistency. Best–worst scaling (BWS) is an alternative method of annotation that is claimed to produce high-quality annotations while keeping the required number of annotations similar to that of rating scales. However, the veracity of this claim has never been systematically established. Here for the first time, we set up an experiment that directly compares the rating scale method with BWS. We show that with the same total number of annotations, BWS produces significantly more reliable results than the rating scale.

Existing Arabic sentiment lexicons have low coverage―with only a few thousand entries. In this paper, we present several large sentiment lexicons that were automatically generated using two different methods: (1) by using distant supervision techniques on Arabic tweets, and (2) by translating English sentiment lexicons into Arabic using a freely available statistical machine translation system. We compare the usefulness of new and old sentiment lexicons in the downstream application of sentence-level sentiment analysis. Our baseline sentiment analysis system uses numerous surface form features. Nonetheless, the system benefits from using additional features drawn from sentiment lexicons. The best result is obtained using the automatically generated Dialectal Hashtag Lexicon and the Arabic translations of the NRC Emotion Lexicon (accuracy of 66.6%). Finally, we describe a qualitative study of the automatic translations of English sentiment lexicons into Arabic, which shows that about 88% of the automatically translated entries are valid for English as well. Close to 10% of the invalid entries are caused by gross mistranslations, close to 40% by translations into a related word, and about 50% by differences in how the word is used in Arabic.

Sentiment composition is the determining of sentiment of a multi-word linguistic unit, such as a phrase or a sentence, based on its constituents. We focus on sentiment composition in phrases formed by at least one positive and at least one negative word ― phrases like ‘happy accident’ and ‘best winter break’. We refer to such phrases as opposing polarity phrases. We manually annotate a collection of opposing polarity phrases and their constituent single words with real-valued sentiment intensity scores using a method known as Best―Worst Scaling. We show that the obtained annotations are consistent. We explore the entries in the lexicon for linguistic regularities that govern sentiment composition in opposing polarity phrases. Finally, we list the current and possible future applications of the lexicon.

We can often detect from a person’s utterances whether he/she is in favor of or against a given target entity (a product, topic, another person, etc.). Here for the first time we present a dataset of tweets annotated for whether the tweeter is in favor of or against pre-chosen targets of interest―their stance. The targets of interest may or may not be referred to in the tweets, and they may or may not be the target of opinion in the tweets. The data pertains to six targets of interest commonly known and debated in the United States. Apart from stance, the tweets are also annotated for whether the target of interest is the target of opinion in the tweet. The annotations were performed by crowdsourcing. Several techniques were employed to encourage high-quality annotations (for example, providing clear and simple instructions) and to identify and discard poor annotations (for example, using a small set of check questions annotated by the authors). This Stance Dataset, which was subsequently also annotated for sentiment, can be used to better understand the relationship between stance, sentiment, entity relationships, and textual inference.