Antske Fokkens

Sergey Troshin, Vlad Niculae, Antske Fokkens

Language models trained on large amounts of data are known to produce inappropriate content in some cases and require careful tuning to be used in the real world. We revisit an effective and modular approach for controllability of the language models, when an external expert model guides the decoding. Particularly, we zoom in into the parametrization choice of an external expert, highlighting the difference between low-rank and higher-rank parametrizations. Higher-rank experts are designed to support high flexibility when representing the rewards, leading to higher computational costs during decoding. However, we demonstrate that they might not use their full flexibility. By analyzing the recently proposed reward-augmented decoding approach (RAD), which uses a higher-rank expert model, we introduce a simpler but more efficient low-rank parametrization of the expert model enabling fast and effective guided decoding. We empirically show that the low-rank RAD performs on par with the more flexible RAD on a detoxification and a sentiment control task, while requiring only a single reward model call per generated token.

Pantea Haghighatkhah, Antske Fokkens, Pia Sommerauer et al.

Bias elimination and recent probing studies attempt to remove specific information from embedding spaces. Here it is important to remove as much of the target information as possible, while preserving any other information present. INLP is a popular recent method which removes specific information through iterative nullspace projections. Multiple iterations, however, increase the risk that information other than the target is negatively affected. We introduce two methods that find a single targeted projection: Mean Projection (MP, more efficient) and Tukey Median Projection (TMP, with theoretical guarantees). Our comparison between MP and INLP shows that (1) one MP projection removes linear separability based on the target and (2) MP has less impact on the overall space. Further analysis shows that applying random projections after MP leads to the same overall effects on the embedding space as the multiple projections of INLP. Applying one targeted (MP) projection hence is methodologically cleaner than applying multiple (INLP) projections that introduce random effects.

Urja Khurana, Ivar Vermeulen, Eric Nalisnick et al.

\textbf{Offensive Content Warning}: This paper contains offensive language only for providing examples that clarify this research and do not reflect the authors' opinions. Please be aware that these examples are offensive and may cause you distress. The subjectivity of recognizing \textit{hate speech} makes it a complex task. This is also reflected by different and incomplete definitions in NLP. We present \textit{hate speech} criteria, developed with perspectives from law and social science, with the aim of helping researchers create more precise definitions and annotation guidelines on five aspects: (1) target groups, (2) dominance, (3) perpetrator characteristics, (4) type of negative group reference, and the (5) type of potential consequences/effects. Definitions can be structured so that they cover a more broad or more narrow phenomenon. As such, conscious choices can be made on specifying criteria or leaving them open. We argue that the goal and exact task developers have in mind should determine how the scope of \textit{hate speech} is defined. We provide an overview of the properties of English datasets from \url{hatespeechdata.com} that may help select the most suitable dataset for a specific scenario.

Urja Khurana, Eric Nalisnick, Antske Fokkens et al.

Subjective tasks in NLP have been mostly relegated to objective standards, where the gold label is decided by taking the majority vote. This obfuscates annotator disagreement and the inherent uncertainty of the label. We argue that subjectivity should factor into model decisions and play a direct role via calibration under a selective prediction setting. Specifically, instead of calibrating confidence purely from the model's perspective, we calibrate models for subjective tasks based on crowd worker agreement. Our method, Crowd-Calibrator, models the distance between the distribution of crowd worker labels and the model's own distribution over labels to inform whether the model should abstain from a decision. On two highly subjective tasks, hate speech detection and natural language inference, our experiments show Crowd-Calibrator either outperforms or achieves competitive performance with existing selective prediction baselines. Our findings highlight the value of bringing human decision-making into model predictions.

Alessandra Polimeno, Myrthe Reuver, Sanne Vrijenhoek et al.

News recommender systems play an increasingly influential role in shaping information access within democratic societies. However, tailoring recommendations to users' specific interests can result in the divergence of information streams. Fragmented access to information poses challenges to the integrity of the public sphere, thereby influencing democracy and public discourse. The Fragmentation metric quantifies the degree of fragmentation of information streams in news recommendations. Accurate measurement of this metric requires the application of Natural Language Processing (NLP) to identify distinct news events, stories, or timelines. This paper presents an extensive investigation of various approaches for quantifying Fragmentation in news recommendations. These approaches are evaluated both intrinsically, by measuring performance on news story clustering, and extrinsically, by assessing the Fragmentation scores of different simulated news recommender scenarios. Our findings demonstrate that agglomerative hierarchical clustering coupled with SentenceBERT text representation is substantially better at detecting Fragmentation than earlier implementations. Additionally, the analysis of simulated scenarios yields valuable insights and recommendations for stakeholders concerning the measurement and interpretation of Fragmentation.

Angel Daza, Antske Fokkens, Tomaž Erjavec

Abbreviations present a significant challenge for NLP systems because they cause tokenization and out-of-vocabulary errors. They can also make the text less readable, especially in reference printed books, where they are extensively used. Abbreviations are especially problematic in low-resource settings, where systems are less robust to begin with. In this paper, we propose a new method for addressing the problems caused by a high density of domain-specific abbreviations in a text. We apply this method to the case of a Slovenian biographical lexicon and evaluate it on a newly developed gold-standard dataset of 51 Slovenian biographies. Our abbreviation identification method performs significantly better than commonly used ad-hoc solutions, especially at identifying unseen abbreviations. We also propose and present the results of a method for expanding the identified abbreviations in context.

Urja Khurana, Michiel van der Meer, Enrico Liscio et al.

Subjective judgments are part of several NLP datasets and recent work is increasingly prioritizing models whose outputs reflect this diversity of perspectives. Such responses allow us to shed light on minority voices, which are frequently marginalized or obscured by dominant perspectives. It remains a question whether our evaluation practices align with these models' objectives. This position paper proposes seven evaluation desiderata for subjectivity-sensitive models, rooted in how subjectivity is represented in NLP data and models. The desiderata are constructed in a top-down approach, keeping in mind the user-centric impact of such models. We scan the experimental setup of 60 papers and show that various aspects of subjectivity are still understudied: the distinction between ambiguous and polyphonic input, whether subjectivity is effectively expressed to the user, and a lack of interplay between different desiderata, amongst other gaps.

Jonathan Kamp, Lisa Beinborn, Antske Fokkens

Argument Unit Recognition and Classification aims at identifying argument units from text and classifying them as pro or against. One of the design choices that need to be made when developing systems for this task is what the unit of classification should be: segments of tokens or full sentences. Previous research suggests that fine-tuning language models on the token-level yields more robust results for classifying sentences compared to training on sentences directly. We reproduce the study that originally made this claim and further investigate what exactly token-based systems learned better compared to sentence-based ones. We develop systematic tests for analysing the behavioural differences between the token-based and the sentence-based system. Our results show that token-based models are generally more robust than sentence-based models both on manually perturbed examples and on specific subpopulations of the data.

Leon Eshuijs, Shihan Wang, Antske Fokkens

Specification gaming under Reinforcement Learning (RL) is known to cause LLMs to develop sycophantic, manipulative, or deceptive behavior, yet the conditions under which this occurs remain unclear. We train 11 instruction-tuned LLMs (0.5B--14B) with on-policy RL across 3 environments and find that model size acts as a safety buffer in some environments but enables greater harmful exploitation in others. Controlled ablations trace this reversal to environment-specific features such as role framing and implicit gameability cues. We further show that most safety benchmarks do not predict RL-induced misalignment, except in the case of Sycophancy scores when the exploit relies on inferring the user's preference. Finally, we find that on-policy RL preserves a safety buffer inherent in the model's own generation distribution, one that is bypassed during off-policy settings.

Jonathan Kamp, Lisa Beinborn, Antske Fokkens

Feature attribution scores are used for explaining the prediction of a text classifier to users by highlighting a k number of tokens. In this work, we propose a way to determine the number of optimal k tokens that should be displayed from sequential properties of the attribution scores. Our approach is dynamic across sentences, method-agnostic, and deals with sentence length bias. We compare agreement between multiple methods and humans on an NLI task, using fixed k and dynamic k. We find that perturbation-based methods and Vanilla Gradient exhibit highest agreement on most method--method and method--human agreement metrics with a static k. Their advantage over other methods disappears with dynamic ks which mainly improve Integrated Gradient and GradientXInput. To our knowledge, this is the first evidence that sequential properties of attribution scores are informative for consolidating attribution signals for human interpretation.

Leon Eshuijs, Shihan Wang, Antske Fokkens

Fairness in classification tasks has traditionally focused on bias removal from neural representations, but recent trends favor algorithmic methods that embed fairness into the training process. These methods steer models towards fair performance, preventing potential elimination of valuable information that arises from representation manipulation. Reinforcement Learning (RL), with its capacity for learning through interaction and adjusting reward functions to encourage desired behaviors, emerges as a promising tool in this domain. In this paper, we explore the usage of RL to address bias in imbalanced classification by scaling the reward function to mitigate bias. We employ the contextual multi-armed bandit framework and adapt three popular RL algorithms to suit our objectives, demonstrating a novel approach to mitigating bias.

Urja Khurana, Eric Nalisnick, Antske Fokkens

When building a predictive model, it is often difficult to ensure that application-specific requirements are encoded by the model that will eventually be deployed. Consider researchers working on hate speech detection. They will have an idea of what is considered hate speech, but building a model that reflects their view accurately requires preserving those ideals throughout the workflow of data set construction and model training. Complications such as sampling bias, annotation bias, and model misspecification almost always arise, possibly resulting in a gap between the application specification and the model's actual behavior upon deployment. To address this issue for hate speech detection, we propose DefVerify: a 3-step procedure that (i) encodes a user-specified definition of hate speech, (ii) quantifies to what extent the model reflects the intended definition, and (iii) tries to identify the point of failure in the workflow. We use DefVerify to find gaps between definition and model behavior when applied to six popular hate speech benchmark datasets.

Sergey Troshin, Wafaa Mohammed, Yan Meng et al.

Diversity is an essential metric for evaluating the creativity of outputs generated by language models. Temperature-based sampling is a common strategy to increase diversity. However, for tasks that require high precision, e.g., mathematical reasoning, uncontrolled high temperature sampling, e.g., min-$p$ or top-$p$, degrades reasoning quality. We demonstrate that the loss of accuracy is caused by sampling incorrect continuations in sensitive decoding positions. To address this, in this paper, we propose \textbf{selective sampling}, a method that dynamically switches between greedy and high-temperature sampling based on a sampling risk metric. This risk metric estimates the likelihood of output errors when applying high-temperature sampling on the current token position. To predict sampling risk, we train a lightweight classifier on a small subset of verifiable problems. The trained classifier can be integrated with the base language model with minimal latency overhead. Experiments on mathematical reasoning tasks demonstrate that selective sampling enhances the quality-diversity trade-off, even in high-temperature settings.

Jonathan Kamp, Lisa Beinborn, Antske Fokkens

Post-hoc explanation methods are an important tool for increasing model transparency for users. Unfortunately, the currently used methods for attributing token importance often yield diverging patterns. In this work, we study potential sources of disagreement across methods from a linguistic perspective. We find that different methods systematically select different classes of words and that methods that agree most with other methods and with humans display similar linguistic preferences. Token-level differences between methods are smoothed out if we compare them on the syntactic span level. We also find higher agreement across methods by estimating the most important spans dynamically instead of relying on a fixed subset of size $k$. We systematically investigate the interaction between $k$ and spans and propose an improved configuration for selecting important tokens.

Myrthe Reuver, Suzan Verberne, Antske Fokkens

For a viewpoint-diverse news recommender, identifying whether two news articles express the same viewpoint is essential. One way to determine "same or different" viewpoint is stance detection. In this paper, we investigate the robustness of operationalization choices for few-shot stance detection, with special attention to modelling stance across different topics. Our experiments test pre-registered hypotheses on stance detection. Specifically, we compare two stance task definitions (Pro/Con versus Same Side Stance), two LLM architectures (bi-encoding versus cross-encoding), and adding Natural Language Inference knowledge, with pre-trained RoBERTa models trained with shots of 100 examples from 7 different stance detection datasets. Some of our hypotheses and claims from earlier work can be confirmed, while others give more inconsistent results. The effect of the Same Side Stance definition on performance differs per dataset and is influenced by other modelling choices. We found no relationship between the number of training topics in the training shots and performance. In general, cross-encoding out-performs bi-encoding, and adding NLI training to our models gives considerable improvement, but these results are not consistent across all datasets. Our results indicate that it is essential to include multiple datasets and systematic modelling experiments when aiming to find robust modelling choices for the concept `stance'.

Jonathan Kamp, Lisa Beinborn, Antske Fokkens

Human explanations of natural language, rationales, form a tool to assess whether models learn a label for the right reasons or rely on dataset-specific shortcuts. Sufficiency is a common metric for estimating the informativeness of rationales, but it provides limited insight into the effects of rationale information on model performance. We address this limitation by relating sufficiency to two modelling paradigms: the ability of models to identify which tokens are part of the rationale (through token classification) and the ability of improving model performance by incorporating rationales in the input (through attention regularisation). We find that highly informative rationales are not likely to help classify the instance correctly. Sufficiency conversely captures the classification impact of the non-rationalised context, which interferes with rationale information in the same input. We also find that incorporating rationale information in model inputs can boost cross-domain classification, but results are inconsistent per task and model type. Finally, sufficiency and token classification appear to be unrelated. These results exemplify the complexity of rationales, showing that metrics capable of systematically capturing this type of information merit further investigation.

Sergey Troshin, Irina Saparina, Antske Fokkens et al.

Large language models increasingly rely on explicit reasoning chains and can produce multiple plausible responses for a given context. We study the candidate sampler that produces the set of plausible responses contrasting the ancestral (parallel) sampling against two alternatives: enumeration, which asks the model to produce $n$ candidates in one pass, and iterative sampling, which proposes candidates sequentially while conditioning on the currently generated response set. Under matched budgets, we compare these samplers on quality, lexical and computation flow diversity, and efficiency. Our empirical results demonstrate that enumeration and iterative strategies result in higher diversity at comparable quality. Our findings highlight the potential of simple non-independent sampling strategies to improve response diversity without sacrificing generation quality.

Alicja Dobrzeniecka, Antske Fokkens, Pia Sommerauer

Amnesic probing is a technique used to examine the influence of specific linguistic information on the behaviour of a model. This involves identifying and removing the relevant information and then assessing whether the model's performance on the main task changes. If the removed information is relevant, the model's performance should decline. The difficulty with this approach lies in removing only the target information while leaving other information unchanged. It has been shown that Iterative Nullspace Projection (INLP), a widely used removal technique, introduces random modifications to representations when eliminating target information. We demonstrate that Mean Projection (MP) and LEACE, two proposed alternatives, remove information in a more targeted manner, thereby enhancing the potential for obtaining behavioural explanations through Amnesic Probing.

Leon Eshuijs, Shihan Wang, Antske Fokkens

Reliance on spurious correlations (shortcuts) has been shown to underlie many of the successes of language models. Previous work focused on identifying the input elements that impact prediction. We investigate how shortcuts are actually processed within the model's decision-making mechanism. We use actor names in movie reviews as controllable shortcuts with known impact on the outcome. We use mechanistic interpretability methods and identify specific attention heads that focus on shortcuts. These heads gear the model towards a label before processing the complete input, effectively making premature decisions that bypass contextual analysis. Based on these findings, we introduce Head-based Token Attribution (HTA), which traces intermediate decisions back to input tokens. We show that HTA is effective in detecting shortcuts in LLMs and enables targeted mitigation by selectively deactivating shortcut-related attention heads.

Urja Khurana, Eric Nalisnick, Antske Fokkens

Despite their success, modern language models are fragile. Even small changes in their training pipeline can lead to unexpected results. We study this phenomenon by examining the robustness of ALBERT (arXiv:1909.11942) in combination with Stochastic Weight Averaging (SWA) (arXiv:1803.05407) -- a cheap way of ensembling -- on a sentiment analysis task (SST-2). In particular, we analyze SWA's stability via CheckList criteria (arXiv:2005.04118), examining the agreement on errors made by models differing only in their random seed. We hypothesize that SWA is more stable because it ensembles model snapshots taken along the gradient descent trajectory. We quantify stability by comparing the models' mistakes with Fleiss' Kappa (Fleiss, 1971) and overlap ratio scores. We find that SWA reduces error rates in general; yet the models still suffer from their own distinct biases (according to CheckList).

Myrthe Reuver, Suzan Verberne, Roser Morante et al.

Cross-topic stance detection is the task to automatically detect stances (pro, against, or neutral) on unseen topics. We successfully reproduce state-of-the-art cross-topic stance detection work (Reimers et. al., 2019), and systematically analyze its reproducibility. Our attention then turns to the cross-topic aspect of this work, and the specificity of topics in terms of vocabulary and socio-cultural context. We ask: To what extent is stance detection topic-independent and generalizable across topics? We compare the model's performance on various unseen topics, and find topic (e.g. abortion, cloning), class (e.g. pro, con), and their interaction affecting the model's performance. We conclude that investigating performance on different topics, and addressing topic-specific vocabulary and context, is a future avenue for cross-topic stance detection.

Pia Sommerauer, Antske Fokkens

This paper presents an approach for investigating the nature of semantic information captured by word embeddings. We propose a method that extends an existing human-elicited semantic property dataset with gold negative examples using crowd judgments. Our experimental approach tests the ability of supervised classifiers to identify semantic features in word embedding vectors and com- pares this to a feature-identification method based on full vector cosine similarity. The idea behind this method is that properties identified by classifiers, but not through full vector comparison are captured by embeddings. Properties that cannot be identified by either method are not. Our results provide an initial indication that semantic properties relevant for the way entities interact (e.g. dangerous) are captured, while perceptual information (e.g. colors) is not represented. We conclude that, though preliminary, these results show that our method is suitable for identifying which properties are captured by embeddings.

Antske Fokkens, Serge ter Braake, Niels Ockeloen et al.

This paper describes BiographyNet, a digital humanities project (2012-2016) that brings together researchers from history, computational linguistics and computer science. The project uses data from the Biography Portal of the Netherlands (BPN), which contains approximately 125,000 biographies from a variety of Dutch biographical dictionaries from the eighteenth century until now, describing around 76,000 individuals. BiographyNet's aim is to strengthen the value of the portal and comparable biographical datasets for historical research, by improving the search options and the presentation of its outcome, with a historically justified NLP pipeline that works through a user evaluated demonstrator. The project's main target group are professional historians. The project therefore worked with two key concepts: "provenance" -understood as a term allowing for both historical source criticism and for references to data-management and programming interventions in digitized sources; and "perspective" interpreted as inherent uncertainty concerning the interpretation of historical results.

Minh Le, Antske Fokkens

Error propagation is a common problem in NLP. Reinforcement learning explores erroneous states during training and can therefore be more robust when mistakes are made early in a process. In this paper, we apply reinforcement learning to greedy dependency parsing which is known to suffer from error propagation. Reinforcement learning improves accuracy of both labeled and unlabeled dependencies of the Stanford Neural Dependency Parser, a high performance greedy parser, while maintaining its efficiency. We investigate the portion of errors which are the result of error propagation and confirm that reinforcement learning reduces the occurrence of error propagation.