Prompt-based language models like GPT4 and LLaMa have been used for a wide variety of use cases such as simulating agents, searching for information, or for content analysis. For all of these applications and others, political biases in these models can affect their performance. Several researchers have attempted to study political bias in language models using evaluation suites based on surveys, such as the Political Compass Test (PCT), often finding a particular leaning favored by these models. However, there is some variation in the exact prompting techniques, leading to diverging findings, and most research relies on constrained-answer settings to extract model responses. Moreover, the Political Compass Test is not a scientifically valid survey instrument. In this work, we contribute a political bias measured informed by political science theory, building on survey design principles to test a wide variety of input prompts, while taking into account prompt sensitivity. We then prompt 11 different open and commercial models, differentiating between instruction-tuned and non-instruction-tuned models, and automatically classify their political stances from 88,110 responses. Leveraging this dataset, we compute political bias profiles across different prompt variations and find that while PCT exaggerates bias in certain models like GPT3.5, measures of political bias are often unstable, but generally more left-leaning for instruction-tuned models. Code and data are available at https://github.com/MaFa211/theory_grounded_pol_bias.

This paper investigates hybrid intelligence and collaboration between researchers of sexism and Large Language Models (LLMs), with afour-component pipeline. First, nine sexism researchers answer questions about their knowledge of sexism and of LLMs. They then participate in two interactive experiments involving an LLM (GPT3.5). The first experiment has experts assessing the model’s knowledgeabout sexism and suitability for use in research. The second experiment tasks them with creating three different definitions of sexism: anexpert-written definition, an LLM-written one, and a co-created definition. Lastly, zero-shot classification experiments use the three definitions from each expert in a prompt template for sexism detection, evaluating GPT4o on 2.500 texts sampled from five sexism benchmarks. We then analyze the resulting 67.500 classification decisions. The LLM interactions lead to longer and more complex definitions of sexism. Expert-written definitions on average perform poorly compared to LLM-generated definitions. However, some experts do improve classification performance with their co-created definitions of sexism, also experts who are inexperienced in using LLMs.

Despite a growing literature finding that large language models (LLMs) exhibit demographic biases, reports with whom they align best are hard to generalize or even contradictory. In this work, we examine the alignment of LLMs with human annotations in five offensive language datasets, comprising approximately 220K annotations. While demographic traits, particularly race, influence alignment, these effects vary across datasets and are often entangled with other factors. Confounders introduced in the annotation process—such as document difficulty, annotator sensitivity, and within-group agreement—account for more variation in alignment patterns than demographic traits. Alignment increases with annotator sensitivity and group agreement, and decreases with document difficulty. Our results underscore the importance of multi-dataset analyses and confounder-aware methodologies in developing robust measures of demographic bias.

Many applications of Large Language Models (LLMs) require them to either simulate people or offer personalized functionality, making the demographic representativeness of LLMs crucial for equitable utility. At the same time, we know little about the extent to which these models actually reflect the demographic attributes and behaviors of certain groups or populations, with conflicting findings in empirical research. To shed light on this debate, we review 211 papers on the demographic representativeness of LLMs. We find that while 29% of the studies report positive conclusions on the representativeness of LLMs, 30% of these do not evaluate LLMs across multiple demographic categories or within demographic subcategories. Another 35% and 47% of the papers concluding positively fail to specify these subcategories altogether for gender and race, respectively. Of the articles that do report subcategories, fewer than half include marginalized groups in their study. Finally, more than a third of the papers do not define the target population to whom their findings apply; of those that do define it either implicitly or explicitly, a large majority study only the U.S. Taken together, our findings suggest an inflated perception of LLM representativeness in the broader community. We recommend more precise evaluation methods and comprehensive documentation of demographic attributes to ensure the responsible use of LLMs for social applications.

The detection of sensitive content in large datasets is crucial for ensuring that shared and analysed data is free from harmful material. However, current moderation tools, such as external APIs, suffer from limitations in customisation, accuracy across diverse sensitive categories, and privacy concerns. Additionally, existing datasets and open-source models focus predominantly on toxic language, leaving gaps in detecting other sensitive categories such as substance abuse or self-harm. In this paper, we put forward a unified dataset tailored for social media content moderation across six sensitive categories: conflictual language, profanity,sexually explicit material, drug-related content, self-harm, and spam. By collecting and annotating data with consistent retrieval strategies and guidelines, we address the shortcomings of previous focalised research. Our analysis demonstrates that fine-tuning large language models (LLMs) on this novel dataset yields significant improvements in detection performance compared to open off-the-shelf models such as LLaMA, and even proprietary OpenAI models, which underperform by 10-15% overall. This limitation is even more pronounced on popular moderation APIs, which cannot be easily tailored to specific sensitive content categories, among others.

With over 60M articles, Wikipedia has become the largest platform for open and freely accessible knowledge. While it has more than 15B monthly visits, its content is believed to be inaccessible to many readers due to the lack of readability of its text. However, previous investigations of the readability of Wikipedia have been restricted to English only, and there are currently no systems supporting the automatic readability assessment of the 300+ languages in Wikipedia. To bridge this gap, we develop a multilingual model to score the readability of Wikipedia articles. To train and evaluate this model, we create a novel multilingual dataset spanning 14 languages, by matching articles from Wikipedia to simplified Wikipedia and online children encyclopedias. We show that our model performs well in a zero-shot scenario, yielding a ranking accuracy of more than 80% across 14 languages and improving upon previous benchmarks. These results demonstrate the applicability of the model at scale for languages in which there is no ground-truth data available for model fine-tuning. Furthermore, we provide the first overview on the state of readability in Wikipedia beyond English.

NLP models are used in a variety of critical social computing tasks, such as detecting sexist, racist, or otherwise hateful content. Therefore, it is imperative that these models are robust to spurious features. Past work has attempted to tackle such spurious features using training data augmentation, including Counterfactually Augmented Data (CADs). CADs introduce minimal changes to existing training data points and flip their labels; training on them may reduce model dependency on spurious features. However, manually generating CADs can be time-consuming and expensive. Hence in this work, we assess if this task can be automated using generative NLP models. We automatically generate CADs using Polyjuice, ChatGPT, and Flan-T5, and evaluate their usefulness in improving model robustness compared to manually-generated CADs. By testing both model performance on multiple out-of-domain test sets and individual data point efficacy, our results show that while manual CADs are still the most effective, CADs generated by ChatGPT come a close second. One key reason for the lower performance of automated methods is that the changes they introduce are often insufficient to flip the original label.

Counterfactually Augmented Data (CAD) aims to improve out-of-domain generalizability, an indicator of model robustness. The improvement is credited to promoting core features of the construct over spurious artifacts that happen to correlate with it. Yet, over-relying on core features may lead to unintended model bias. Especially, construct-driven CAD—perturbations of core features—may induce models to ignore the context in which core features are used. Here, we test models for sexism and hate speech detection on challenging data: non-hate and non-sexist usage of identity and gendered terms. On these hard cases, models trained on CAD, especially construct-driven CAD, show higher false positive rates than models trained on the original, unperturbed data. Using a diverse set of CAD—construct-driven and construct-agnostic—reduces such unintended bias.

As NLP models are increasingly deployed in socially situated settings such as online abusive content detection, it is crucial to ensure that these models are robust. One way of improving model robustness is to generate counterfactually augmented data (CAD) for training models that can better learn to distinguish between core features and data artifacts. While models trained on this type of data have shown promising out-of-domain generalizability, it is still unclear what the sources of such improvements are. We investigate the benefits of CAD for social NLP models by focusing on three social computing constructs — sentiment, sexism, and hate speech. Assessing the performance of models trained with and without CAD across different types of datasets, we find that while models trained on CAD show lower in-domain performance, they generalize better out-of-domain. We unpack this apparent discrepancy using machine explanations and find that CAD reduces model reliance on spurious features. Leveraging a novel typology of CAD to analyze their relationship with model performance, we find that CAD which acts on the construct directly or a diverse set of CAD leads to higher performance.

Social media sites like Twitter possess the potential to complement surveys that measure political opinions and, more specifically, political actors’ approval. However, new challenges related to the reliability and validity of social-media-based estimates arise. Various sentiment analysis and stance detection methods have been developed and used in previous research to measure users’ political opinions based on their content on social media. In this work, we attempt to gauge the efficacy of untargeted sentiment, targeted sentiment, and stance detection methods in labeling various political actors’ approval by benchmarking them across several datasets. We also contrast the performance of these pretrained methods that can be used in an off-the-shelf (OTS) manner against a set of models trained on minimal custom data. We find that OTS methods have low generalizability on unseen and familiar targets, while low-resource custom models are more robust. Our work sheds light on the strengths and limitations of existing methods proposed for understanding politicians’ approval from tweets.

While growing code-mixed content on Online Social Networks(OSN) provides a fertile ground for studying various aspects of code-mixing, the lack of automated text analysis tools render such studies challenging. To meet this challenge, a family of tools for analyzing code-mixed data such as language identifiers, parts-of-speech (POS) taggers, chunkers have been developed. Named Entity Recognition (NER) is an important text analysis task which is not only informative by itself, but is also needed for downstream NLP tasks such as semantic role labeling. In this work, we present an exploration of automatic NER of code-mixed data. We compare our method with existing off-the-shelf NER tools for social media content,and find that our systems outperforms the best baseline by 33.18 % (F1 score).

Code-mixing is a linguistic phenomenon where multiple languages are used in the same occurrence that is increasingly common in multilingual societies. Code-mixed content on social media is also on the rise, prompting the need for tools to automatically understand such content. Automatic Parts-of-Speech (POS) tagging is an essential step in any Natural Language Processing (NLP) pipeline, but there is a lack of annotated data to train such models. In this work, we present a unique language tagged and POS-tagged dataset of code-mixed English-Hindi tweets related to five incidents in India that led to a lot of Twitter activity. Our dataset is unique in two dimensions: (i) it is larger than previous annotated datasets and (ii) it closely resembles typical real-world tweets. Additionally, we present a POS tagging model that is trained on this dataset to provide an example of how this dataset can be used. The model also shows the efficacy of our dataset in enabling the creation of code-mixed social media POS taggers.