People naturally vary in their annotations for subjective questions and some of this variation is thought to be due to the person’s sociodemographic characteristics. LLMs have also been used to label data, but recent work has shown that models perform poorly when prompted with sociodemographic attributes, suggesting limited inherent sociodemographic knowledge. Here, we ask whether LLMs can be trained to be accurate sociodemographic models of annotator variation. Using a curated dataset of five tasks with standardized sociodemographics, we show that models do improve in sociodemographic prompting when trained but that this performance gain is largely due to models learning annotator-specific behaviour rather than sociodemographic behaviours. Across all tasks, our results suggest that models learn little meaningful connection between sociodemographics and annotation, raising doubts about the current use of LLMs for simulating sociodemographic variation and behaviour.

Personalized content moderation can protect users from harm while facilitating free expression by tailoring moderation decisions to individual preferences rather than enforcing universal rules. However, content moderation that is fully personalized to individual preferences, no matter what these preferences are, may lead to even the most hazardous types of content being propagated on social media. In this paper, we explore this risk using hate speech as a case study. Certain types of hate speech are illegal in many countries. We show that, while fully personalized hate speech detection models increase overall user welfare (as measured by user-level classification performance), they also make predictions that violate such legal hate speech boundaries, especially when tailored to users who tolerate highly hateful content. To address this problem, we enforce legal boundaries in personalized hate speech detection by overriding predictions from personalized models with those from a boundary classifier. This approach significantly reduces legal violations while minimally affecting overall user welfare. Our findings highlight both the promise and the risks of personalized moderation, and offer a practical solution to balance user preferences with legal and ethical obligations.

For datasets to accurately represent diverse opinions in a population, they must preserve variation in data labels while filtering out spam or low-quality responses. How can we balance annotator reliability and representation? We empirically evaluate how a range of heuristics for annotator filtering affect the preservation of variation on subjective tasks. We find that these methods, designed for contexts in which variation from a single ground-truth label is considered noise, often remove annotators who disagree instead of spam annotators, introducing suboptimal tradeoffs between accuracy and label diversity. We find that conservative settings for annotator removal (<5%) are best, after which all tested methods increase the mean absolute error from the true average label. We analyze performance on synthetic spam to observe that these methods often assume spam annotators are less random than real spammers tend to be: most spammers are distributionally indistinguishable from real annotators, and the minority that are distinguishable tend to give fixed answers, not random ones. Thus, tasks requiring the preservation of variation reverse the intuition of existing spam filtering methods: spammers tend to be less random than non-spammers, so metrics that assume variation is spam fare worse. These results highlight the need for spam removal methods that account for label diversity.

Many NLP tasks exhibit human label variation, where different annotators give different labels to the same texts. This variation is known to depend, at least in part, on the sociodemographics of annotators. Recent research aims to model individual annotator behaviour rather than predicting aggregated labels, and we would expect that sociodemographic information is useful for these models. On the other hand, the ecological fallacy states that aggregate group behaviour, such as the behaviour of the average female annotator, does not necessarily explain individual behaviour. To account for sociodemographics in models of individual annotator behaviour, we introduce group-specific layers to multi-annotator models. In a series of experiments for toxic content detection, we find that explicitly accounting for sociodemographic attributes in this way does not significantly improve model performance. This result shows that individual annotation behaviour depends on much more than just sociodemographics.

Counterspeech on social media is rare. Consequently, it is difficult to collect naturally occurring examples, in particular for languages without annotated datasets. In this work, we study methods to increase the relevance of social media samples for counterspeech annotation when we lack annotated resources. We use the example of sourcing German data for counterspeech annotations from Twitter. We monitor tweets from German politicians and activists to collect replies. To select relevant replies we a) find replies that match German abusive keywords or b) label replies for counterspeech using a multilingual classifier fine-tuned on English data. For both approaches and a baseline setting, we annotate a random sample and use bootstrap sampling to estimate the amount of counterspeech. We find that neither the multilingual model nor the keyword approach achieve significantly higher counts of true counterspeech than the baseline. Thus, keyword lists or multi-lingual classifiers are likely not worth the added complexity beyond purposive data collection: Already without additional filtering, we gather a meaningful sample with 7,4% true counterspeech.

Many annotation tasks in natural language processing are highly subjective in that there can be different valid and justified perspectives on what is a proper label for a given example. This also applies to the judgment of argument quality, where the assignment of a single ground truth is often questionable. At the same time, there are generally accepted concepts behind argumentation that form a common ground. To best represent the interplay of individual and shared perspectives, we consider a continuum of approaches ranging from models that fully aggregate perspectives into a majority label to “share nothing”-architectures in which each annotator is considered in isolation from all other annotators. In between these extremes, inspired by models used in the field of recommender systems, we investigate the extent to which architectures that predict labels for single annotators but include layers that model the relations between different annotators are beneficial. By means of two tasks of argument quality classification (argument concreteness and validity/novelty of conclusions), we show that recommender architectures increase the averaged annotator-individual F1-scores up to 43% over a majority-label model. Our findings indicate that approaches to subjectivity can benefit from relating individual perspectives.

We propose to study the evolution of concepts by learning to complete diachronic analogies between lists of terms which relate to the same concept at different points in time. We present a number of models based on operations on word embedddings that correspond to different assumptions about the characteristics of diachronic analogies and change in concept vocabularies. These are tested in a quantitative evaluation for nine different concepts on a corpus of Dutch newspapers from the 1950s and 1980s. We show that a model which treats the concept terms as analogous and learns weights to compensate for diachronic changes (weighted linear combination) is able to more accurately predict the missing term than a learned transformation and two baselines for most of the evaluated concepts. We also find that all models tend to be coherent in relation to the represented concept, but less discriminative in regard to other concepts. Additionally, we evaluate the effect of aligning the time-specific embedding spaces using orthogonal Procrustes, finding varying effects on performance, depending on the model, concept and evaluation metric. For the weighted linear combination, however, results improve with alignment in a majority of cases. All related code is released publicly.