Research Group Barbara Plank

Overparametrized transformer networks are the state-of-the-art architecture for Large Language Models (LLMs). However, such models contain billions of parameters making large compute a necessity, while raising environmental concerns. To address these issues, we propose FinerCut, a new form of fine-grained layer pruning, which in contrast to prior work at the transformer block level, considers all self-attention and feed-forward network (FFN) layers within blocks as individual pruning candidates. FinerCut prunes layers whose removal causes minimal alternation to the model’s output – contributing to a new, lean, interpretable, and task-agnostic pruning method. Tested across 9 benchmarks, our approach retains 90% performance of Llama3-8B with 25% layers removed, and 95% performance of Llama3-70B with 30% layers removed, all without fine-tuning or post-pruning reconstruction. Strikingly, we observe intriguing results with FinerCut: 42% (34 out of 80) of the self-attention layers in Llama3-70B can be removed while preserving 99% of its performance – without additional fine-tuning after removal. Moreover, FinerCut provides a tool to inspect the types and locations of pruned layers, allowing to observe interesting pruning behaviors. For instance, we observe a preference for pruning self-attention layers, often at deeper consecutive decoder layers. We hope our insights inspire future efficient LLM architecture designs.

Knights and knaves problems represent a classic genre of logical puzzles where characters either tell the truth or lie. The objective is to logically deduce each character’s identity based on their statements. The challenge arises from the truth-telling or lying behavior, which influences the logical implications of each statement. Solving these puzzles requires not only direct deductions from individual statements, but the ability to assess the truthfulness of statements by reasoning through various hypothetical scenarios. As such, knights and knaves puzzles serve as compelling examples of suppositional reasoning. In this paper, we introduce TruthQuest, a benchmark for suppositional reasoning based on the principles of knights and knaves puzzles. Our benchmark presents problems of varying complexity, considering both the number of characters and the types of logical statements involved. Evaluations on TruthQuest show that large language models like Llama 3 and Mixtral-8x7B exhibit significant difficulties solving these tasks. A detailed error analysis of the models’ output reveals that lower-performing models exhibit a diverse range of reasoning errors, frequently failing to grasp the concept of truth and lies. In comparison, more proficient models primarily struggle with accurately inferring the logical implications of potentially false statements.

Human label variation (HLV) is a valuable source of information that arises when multiple human annotators provide different labels for valid reasons. In Natural Language Inference (NLI) earlier approaches to capturing HLV involve either collecting annotations from many crowd workers to represent human judgment distribution (HJD) or use expert linguists to provide detailed explanations for their chosen labels. While the former method provides denser HJD information, obtaining it is resource-intensive. In contrast, the latter offers richer textual information but it is challenging to scale up to many human judges. Besides, large language models (LLMs) are increasingly used as evaluators (‘LLM judges’) but with mixed results, and few works aim to study HJDs. This study proposes to exploit LLMs to approximate HJDs using a small number of expert labels and explanations. Our experiments show that a few explanations significantly improve LLMs’ ability to approximate HJDs with and without explicit labels, thereby providing a solution to scale up annotations for HJD. However, fine-tuning smaller soft-label aware models with the LLM-generated model judgment distributions (MJDs) presents partially inconsistent results: while similar in distance, their resulting fine-tuned models and visualized distributions differ substantially. We show the importance of complementing instance-level distance measures with a global-level shape metric and visualization to more effectively evaluate MJDs against human judgment distributions.

One of the major aspects contributing to the striking performance of large language models (LLMs) is the vast amount of factual knowledge accumulated during pre-training. Yet, many LLMs suffer from self-inconsistency, which raises doubts about their trustworthiness and reliability. This paper focuses on entity type ambiguity, analyzing the proficiency and consistency of state-of-the-art LLMs in applying factual knowledge when prompted with ambiguous entities. To do so, we propose an evaluation protocol that disentangles knowing from applying knowledge, and test state-of-the-art LLMs on 49 ambiguous entities. Our experiments reveal that LLMs struggle with choosing the correct entity reading, achieving an average accuracy of only 85%, and as low as 75% with underspecified prompts. The results also reveal systematic discrepancies in LLM behavior, showing that while the models may possess knowledge, they struggle to apply it consistently, exhibit biases toward preferred readings, and display self-inconsistencies. This highlights the need to address entity ambiguity in the future for more trustworthy LLMs.

Recent advances in Large Language Models (LLMs) have sparked wide interest in validating and comprehending the human-like cognitive-behavioral traits LLMs may capture and convey. These cognitive-behavioral traits include typically Attitudes, Opinions, Values (AOVs). However, measuring AOVs embedded within LLMs remains opaque, and different evaluation methods may yield different results. This has led to a lack of clarity on how different studies are related to each other and how they can be interpreted. This paper aims to bridge this gap by providing a comprehensive overview of recent works on the evaluation of AOVs in LLMs. Moreover, we survey related approaches in different stages of the evaluation pipeline in these works. By doing so, we address the potential and challenges with respect to understanding the model, human-AI alignment, and downstream application in social sciences. Finally, we provide practical insights into evaluation methods, model enhancement, and interdisciplinary collaboration, thereby contributing to the evolving landscape of evaluating AOVs in LLMs.

Climate change (CC) has attracted increasing attention in NLP in recent years. However, detecting the stance on CC in multimodal data is understudied and remains challenging due to a lack of reliable datasets. To improve the understanding of public opinions and communication strategies, this paper presents MultiClimate, the first open-source manually-annotated stance detection dataset with 100 CC-related YouTube videos and 4,209 frame-transcript pairs. We deploy state-of-the-art vision and language models, as well as multimodal models for MultiClimate stance detection. Results show that text-only BERT significantly outperforms image-only ResNet50 and ViT. Combining both modalities achieves state-of-the-art, 0.747/0.749 in accuracy/F1. Our 100M-sized fusion models also beat CLIP and BLIP, as well as the much larger 9B-sized multimodal IDEFICS and text-only Llama3 and Gemma2, indicating that multimodal stance detection remains challenging for large language models.

In the recent past, a popular way of evaluating natural language understanding (NLU), was to consider a model’s ability to perform natural language inference (NLI) tasks. In this paper, we investigate if NLI tasks, that are rarely used for LLM evaluation, can still be informative for evaluating LLMs. Focusing on five different NLI benchmarks across six models of different scales, we investigate if they are able to discriminate models of different size and quality and how their accuracies develop during training. Furthermore, we investigate the extent to which the softmax distributions of models align with human distributions in cases where statements are ambiguous or vague. Overall, our results paint a positive picture for the NLI tasks: we find that they are able to discriminate well between models at various stages of training, yet are not (all) saturated. Furthermore, we find that while the similarity of model distributions with human label distributions increases with scale, it is still much higher than the similarity between two populations of humans, making it a potentially interesting statistic to consider.

Large language models (LLMs) have recently shown impressive performance on tasks involving reasoning, leading to a lively debate on whether these models possess reasoning capabilities similar to humans. However, despite these successes, the depth of LLMs’ reasoning abilities remains uncertain. This uncertainty partly stems from the predominant focus on task performance, measured through shallow accuracy metrics, rather than a thorough investigation of the models’ reasoning behavior. This paper seeks to address this gap by providing a comprehensive review of studies that go beyond task accuracy, offering deeper insights into the models’ reasoning processes. Furthermore, we survey prevalent methodologies to evaluate the reasoning behavior of LLMs, emphasizing current trends and efforts towards more nuanced reasoning analyses. Our review suggests that LLMs tend to rely on surface-level patterns and correlations in their training data, rather than on sophisticated reasoning abilities. Additionally, we identify the need for further research that delineates the key differences between human and LLM-based reasoning. Through this survey, we aim to shed light on the complex reasoning processes within LLMs.

Multiple choice questions (MCQs) are commonly used to evaluate the capabilities of large language models (LLMs). One common way to evaluate the model response is to rank the candidate answers based on the log probability of the first token prediction. An alternative way is to examine the text output. Prior work has shown that first token probabilities lack robustness to changes in MCQ phrasing, and that first token probabilities do not match text answers for instruction-tuned models. Therefore, in this paper, we investigate the robustness of text answers. We show that the text answers are more robust to question perturbations than the first token probabilities, when the first token answers mismatch the text answers. The difference in robustness increases as the mismatch rate becomes greater. As the mismatch reaches over 50%, the text answer is more robust to option order changes than the debiased first token probabilities using state-of-the-art debiasing methods such as PriDe. Our findings provide further evidence for the benefits of text answer evaluation over first token probability evaluation.

This document provides the annotation guidelines for MaiBaam, a Bavarian corpus manually annotated with part-of-speech (POS) tags, syntactic dependencies, and German lemmas. MaiBaam belongs to the Universal Dependencies (UD) project, and our annotations elaborate on the general and German UD version 2 guidelines. In this document, we detail how to preprocess and tokenize Bavarian data, provide an overview of the POS tags and dependencies we use, explain annotation decisions that would also apply to closely related languages like German, and lastly we introduce and motivate decisions that are specific to Bavarian grammar.

Tree of Thoughts (ToT) is a reasoning strategy for Large Language Models (LLMs) that employs a generator to suggest reasoning steps and a discriminator to decide which steps to implement. ToT demonstrates strong performance on reasoning tasks, often surpassing simple methods such as Input-Output (IO) prompting and Chain-of-Thought (CoT) reasoning. However, ToT does not consistently outperform such simpler methods across all models, leaving large knowledge gaps on the conditions under which ToT is most beneficial. In this paper, we analyze the roles of the generator and discriminator separately to better understand the conditions when ToT is beneficial. We find that the generator plays a more critical role than the discriminator in driving the success of ToT. Scaling the generator leads to notable improvements in ToT performance, even when using a smaller model as the discriminator, whereas scaling the discriminator with a fixed generator yields only marginal gains. Our results show that models across different scales exhibit comparable discrimination capabilities, yet differ significantly in their generative performance for ToT.

Large vision-language models frequently struggle to accurately predict responses provided by multiple human annotators, particularly when those responses exhibit human uncertainty. In this study, we focus on the Visual Question Answering (VQA) task, and we comprehensively evaluate how well the state-of-the-art vision-language models correlate with the distribution of human responses. To do so, we categorize our samples based on their levels (low, medium, high) of human uncertainty in disagreement (HUD) and employ not only accuracy but also three new human-correlated metrics in VQA, to investigate the impact of HUD. To better align models with humans, we also verify the effect of common calibration and human calibration. Our results show that even BEiT3, currently the best model for this task, struggles to capture the multi-label distribution inherent in diverse human responses. Additionally, we observe that the commonly used accuracy-oriented calibration technique adversely affects BEiT3’s ability to capture HUD, further widening the gap between model predictions and human distributions. In contrast, we show the benefits of calibrating models towards human distributions for VQA, better aligning model confidence with human uncertainty. Our findings highlight that for VQA, the consistent alignment between human responses and model predictions is understudied and should become the next crucial target of future studies.

A fundamental question in interpretability research is to what extent neural networks, particularly language models, implement reusable functions via subnetworks that can be composed to perform more complex tasks. Recent developments in mechanistic interpretability have made progress in identifying subnetworks, often referred to as circuits, which represent the minimal computational subgraph responsible for a model’s behavior on specific tasks. However, most studies focus on identifying circuits for individual tasks without investigating how functionally similar circuits relate to each other. To address this gap, we examine the modularity of neural networks by analyzing circuits for highly compositional subtasks within a transformer-based language model. Specifically, given a probabilistic context-free grammar, we identify and compare circuits responsible for ten modular string-edit operations. Our results indicate that functionally similar circuits exhibit both notable node overlap and cross-task faithfulness. Moreover, we demonstrate that the circuits identified can be reused and combined through subnetwork set operations to represent more complex functional capabilities of the model.

There is increasing interest in looking at dialects in NLP. However, most work to date still treats dialects as discrete categories. For instance, evaluative work in variation-oriented NLP for English often works with Indian English or African-American Venacular English as homogeneous categories (Faisal et al., 2024; Ziems et al., 2023), yet even within one variety there is substantial variation. We examine within-dialect variation and show that performance critically varies within categories. We measure speech-to-text performance on Italian dialects, and empirically observe a geographical performance disparity. This disparity correlates substantially (-0.5) with linguistic similarity to the highest performing dialect variety. We cross-examine our results against dialectometry methods, and interpret the performance disparity to be due to a bias towards dialects that are more similar to the standard variety in the speech-to-text model examined. We additionally leverage geostatistical methods to predict zero-shot performance at unseen sites, and find the incorporation of geographical information to substantially improve prediction performance, indicating there to be geographical structure in the performance distribution.

Training a language model to be both helpful and harmless requires careful calibration of refusal behaviours: Models should refuse to follow malicious instructions or give harmful advice (e.g. ‘how do I kill someone?’’), but they should not refuse safe requests, even if they superficially resemble unsafe ones (e.g. ‘how do I kill a Python process?’’). Avoiding such false refusal, as prior work has shown, is challenging even for highly-capable language models. In this paper, we propose a simple and surgical method for mitigating false refusal in language models via single vector ablation. For a given model, we extract a false refusal vector and show that ablating this vector reduces false refusal rate without negatively impacting model safety and general model capabilities. We also show that our approach can be used for fine-grained calibration of model safety. Our approach is training-free and model-agnostic, making it useful for mitigating the problem of false refusal in current and future language models.

Natural language processing (NLP) has largely focused on modelling standardized languages. More recently, attention has increasingly shifted to local, non-standardized languages and dialects. However, the relevant speaker populations’ needs and wishes with respect to NLP tools are largely unknown. In this paper, we focus on dialects and regional languages related to German – a group of varieties that is heterogeneous in terms of prestige and standardization. We survey speakers of these varieties (N=327) and present their opinions on hypothetical language technologies for their dialects. Although attitudes vary among subgroups of our respondents, we find that respondents are especially in favour of potential NLP tools that work with dialectal input (especially audio input) such as virtual assistants, and less so for applications that produce dialectal output such as machine translation or spellcheckers.

Deductive reasoning plays a pivotal role in the formulation of sound and cohesive arguments. It allows individuals to draw conclusions that logically follow, given the truth value of the information provided. Recent progress in the domain of large language models (LLMs) has showcased their capability in executing deductive reasoning tasks. Nonetheless, a significant portion of research primarily assesses the accuracy of LLMs in solving such tasks, often overlooking a deeper analysis of their reasoning behavior. In this study, we draw upon principles from cognitive psychology to examine inferential strategies employed by LLMs, through a detailed evaluation of their responses to propositional logic problems. Our findings indicate that LLMs display reasoning patterns akin to those observed in humans, including strategies like supposition following or chain construction. Moreover, our research demonstrates that the architecture and scale of the model significantly affect its preferred method of reasoning, with more advanced models tending to adopt strategies more frequently than less sophisticated ones. Importantly, we assert that a model’s accuracy, that is the correctness of its final conclusion, does not necessarily reflect the validity of its reasoning process. This distinction underscores the necessity for more nuanced evaluation procedures in the field.

Human label variation arises when annotators assign different labels to the same item for valid reasons, while annotation errors occur when labels are assigned for invalid reasons. These two issues are prevalent in NLP benchmarks, yet existing research has studied them in isolation. To the best of our knowledge, there exists no prior work that focuses on teasing apart error from signal, especially in cases where signal is beyond black-and-white.To fill this gap, we introduce a systematic methodology and a new dataset, VariErr (variation versus error), focusing on the NLI task in English. We propose a 2-round annotation procedure with annotators explaining each label and subsequently judging the validity of label-explanation pairs.VariErr contains 7,732 validity judgments on 1,933 explanations for 500 re-annotated MNLI items. We assess the effectiveness of various automatic error detection (AED) methods and GPTs in uncovering errors versus human label variation. We find that state-of-the-art AED methods significantly underperform GPTs and humans. While GPT-4 is the best system, it still falls short of human performance. Our methodology is applicable beyond NLI, offering fertile ground for future research on error versus plausible variation, which in turn can yield better and more trustworthy NLP systems.

In legal decisions, split votes (SV) occur when judges cannot reach a unanimous decision, posing a difficulty for lawyers who must navigate diverse legal arguments and opinions. In high-stakes domains, %as human-AI interaction systems become increasingly important, understanding the alignment of perceived difficulty between humans and AI systems is crucial to build trust. However, existing NLP calibration methods focus on a classifier’s awareness of predictive performance, measured against the human majority class, overlooking inherent human label variation (HLV). This paper explores split votes as naturally observable human disagreement and value pluralism. We collect judges’ vote distributions from the European Court of Human Rights (ECHR), and present SV-ECHR, a case outcome classification (COC) dataset with SV information. We build a taxonomy of disagreement with SV-specific subcategories. We further assess the alignment of perceived difficulty between models and humans, as well as confidence- and human-calibration of COC models. We observe limited alignment with the judge vote distribution. To our knowledge, this is the first systematic exploration of calibration to human judgements in legal NLP. Our study underscores the necessity for further research on measuring and enhancing model calibration considering HLV in legal decision tasks.

The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model’s diverse response styles such as starting with ‘Sure’ or refusing to answer. Consequently, first-token evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output as well and ii) caution against relying solely on first-token evaluation.

Climate Change (CC) is a pressing topic of global importance, attracting increasing attention across research fields, from social sciences to Natural Language Processing (NLP). CC is also discussed in various settings and communication platforms, from academic publications to social media forums. Understanding who and what is mentioned in such data is a first critical step to gaining new insights into CC. We present CLIMATELI (CLIMATe Entity LInking), the first manually annotated CC dataset that links 3,087 entity spans to Wikipedia. Using CLIMATELI (CLIMATe Entity LInking), we evaluate existing entity linking (EL) systems on the CC topic across various genres and propose automated filtering methods for CC entities. We find that the performance of EL models notably lags behind humans at both token and entity levels. Testing within the scope of retaining or excluding non-nominal and/or non-CC entities particularly impacts the models’ performances.

Whether future AI models are fair, trustworthy, and aligned with the public’s interests rests in part on our ability to collect accurate data about what we want the models to do. However, collecting high-quality data is difficult, and few AI/ML researchers are trained in data collection methods. Recent research in data-centric AI has show that higher quality training data leads to better performing models, making this the right moment to introduce AI/ML researchers to the field of survey methodology, the science of data collection. We summarize insights from the survey methodology literature and discuss how they can improve the quality of training and feedback data. We also suggest collaborative research ideas into how biases in data collection can be mitigated, making models more accurate and human-centric.

This paper presents our system developed for the SemEval-2024 Task 1: Semantic Textual Relatedness (STR), on Track C: Cross-lingual. The task aims to detect semantic relatedness of two sentences from the same languages. For cross-lingual approach we developed a set of linguistics-inspired models trained with several task-specific strategies. We 1) utilize language vectors for selection of donor languages; 2) investigate the multi-source approach for training; 3) use transliteration of non-latin script to study impact of ‘script gap’; 4) opt machine translation for data augmentation. We additionally compare the performance of XLM-RoBERTa and Furina with the same training strategy. Our submission achieved the first place in the C8 (Kinyarwanda) test.

There is an increasing trend towards evaluating NLP models with LLM-generated judgments instead of human judgments. In the absence of a comparison against human data, this raises concerns about the validity of these evaluations; in case they are conducted with proprietary models, this also raises concerns over reproducibility. We provide JUDGE-BENCH, a collection of 20 NLP datasets with human annotations, and comprehensively evaluate 11 current LLMs, covering both open-weight and proprietary models, for their ability to replicate the annotations. Our evaluations show that each LLM exhibits a large variance across datasets in its correlation to human judgments. We conclude that LLMs are not yet ready to systematically replace human judges in NLP.

Despite the success of the Universal Dependencies (UD) project exemplified by its impressive language breadth, there is still a lack in `within-language breadth’: most treebanks focus on standard languages. Even for German, the language with the most annotations in UD, so far no treebank exists for one of its language varieties spoken by over 10M people: Bavarian. To contribute to closing this gap, we present the first multi-dialect Bavarian treebank (MaiBaam) manually annotated with part-of-speech and syntactic dependency information in UD, covering multiple text genres (wiki, fiction, grammar examples, social, non-fiction). We highlight the morphosyntactic differences between the closely-related Bavarian and German and showcase the rich variability of speakers’ orthographies. Our corpus includes 15k tokens, covering dialects from all Bavarian-speaking areas spanning three countries. We provide baseline parsing and POS tagging results, which are lower than results obtained on German and vary substantially between different graph-based parsers. To support further research on Bavarian syntax, we make our dataset, language-specific guidelines and code publicly available.

Polar questions are common in dialogue and expect exactly one of two answers (yes/no). It is however not uncommon for speakers to bypass these expected choices and answer, for example, ‘Islands are generally by the sea’ to the question: ‘An island? By the sea?’. While such answers are natural in spoken dialogues, conversational systems still struggle to interpret them. Seminal work to interpret indirect answers were made in recent years—but only for English and with strict question formulations. In this work, we present a new corpus for French and Spanish—IndirectQA —where we mine subtitle data for indirect answers to study the labeling task with six different labels, while broadening polar questions to include also implicit polar questions (statements that trigger a yes/no-answer which are not necessarily formulated as a question). We opted for subtitles since they are a readily available source of conversation in various languages, but also come with peculiarities and challenges which we will discuss. Overall, we provide the first results on French and Spanish. They show that the task is challenging: the baseline accuracy scores drop from 61.43 on English to 44.06 for French and Spanish.

Named Entity Recognition (NER) is a fundamental task to extract key information from texts, but annotated resources are scarce for dialects. This paper introduces the first dialectal NER dataset for German, BarNER, with 161K tokens annotated on Bavarian Wikipedia articles (bar-wiki) and tweets (bar-tweet), using a schema adapted from German CoNLL 2006 and GermEval. The Bavarian dialect differs from standard German in lexical distribution, syntactic construction, and entity information. We conduct in-domain, cross-domain, sequential, and joint experiments on two Bavarian and three German corpora and present the first comprehensive NER results on Bavarian. Incorporating knowledge from the larger German NER (sub-)datasets notably improves on bar-wiki and moderately on bar-tweet. Inversely, training first on Bavarian contributes slightly to the seminal German CoNLL 2006 corpus. Moreover, with gold dialect labels on Bavarian tweets, we assess multi-task learning between five NER and two Bavarian-German dialect identification tasks and achieve NER SOTA on bar-wiki. We substantiate the necessity of our low-resource BarNER corpus and the importance of diversity in dialects, genres, and topics in enhancing model performance.

Digital assistants perform well in high-resource languages like English, where tasks like slot and intent detection (SID) are well-supported. Many recent SID datasets start including multiple language varieties. However, it is unclear how realistic these translated datasets are. Therefore, we extend one such dataset, namely xSID-0.4, to include two underrepresented languages: Bavarian, a German dialect, and Lithuanian, a Baltic language. Both language variants have limited speaker populations and are often not included in multilingual projects. In addition to translations we provide “natural” queries to digital assistants generated by native speakers. We further include utterances from another dataset for Bavarian to build the richest SID dataset available today for a low-resource dialect without standard orthography. We then set out to evaluate models trained on English in a zero-shot scenario on our target language variants. Our evaluation reveals that translated data can produce overly optimistic scores. However, the error patterns in translated and natural datasets are highly similar. Cross-dataset experiments demonstrate that data collection methods influence performance, with scores lower than those achieved with single-dataset translations. This work contributes to enhancing SID datasets for underrepresented languages, yielding NaLiBaSID, a new evaluation dataset for Bavarian and Lithuanian.

In this work, we analyze the uncertainty that is inherently present in the labels used for supervised machine learning in natural language inference (NLI). In cases where multiple annotations per instance are available, neither the majority vote nor the frequency of individual class votes is a trustworthy representation of the labeling uncertainty. We propose modeling the votes via a Bayesian mixture model to recover the data-generating process, i.e., the “true” latent classes, and thus gain insight into the class variations. This will enable a better understanding of the confusion happening during the annotation process. We also assess the stability of the proposed estimation procedure by systematically varying the numbers of i) instances and ii) labels. Thereby, we observe that few instances with many labels can predict the latent class borders reasonably well, while the estimation fails for many instances with only a few labels. This leads us to conclude that multiple labels are a crucial building block for properly analyzing label uncertainty.

Named Entity Recognition (NER) is a key information extraction task with a long-standing tradition. While recent studies address and aim to correct annotation errors via re-labeling efforts, little is known about the sources of human label variation, such as text ambiguity, annotation error, or guideline divergence. This is especially the case for high-quality datasets and beyond English CoNLL03. This paper studies disagreements in expert-annotated named entity datasets for three languages: English, Danish, and Bavarian. We show that text ambiguity and artificial guideline changes are dominant factors for diverse annotations among high-quality revisions. We survey student annotations on a subset of difficult entities and substantiate the feasibility and necessity of manifold annotations for understanding named entity ambiguities from a distributional perspective.

Mainstream cross-lingual task-oriented dialogue (ToD) systems leverage the transfer learning paradigm by training a joint model for intent recognition and slot-filling in English and applying it, zero-shot, to other languages.We address a gap in prior research, which often overlooked the transfer to lower-resource colloquial varieties due to limited test data.Inspired by prior work on English varieties, we craft and manually evaluate perturbation rules that transform German sentences into colloquial forms and use them to synthesize test sets in four ToD datasets.Our perturbation rules cover 18 distinct language phenomena, enabling us to explore the impact of each perturbation on slot and intent performance.Using these new datasets, we conduct an experimental evaluation across six different transformers.Here, we demonstrate that when applied to colloquial varieties, ToD systems maintain their intent recognition performance, losing 6% (4.62 percentage points) in accuracy on average. However, they exhibit a significant drop in slot detection, with a decrease of 31% (21 percentage points) in slot F1 score.Our findings are further supported by a transfer experiment from Standard American English to synthetic Urban African American Vernacular English.

With the rise of increasingly powerful and user-facing NLP systems, there is growing interest in assessing whether they have a good representation of uncertainty by evaluating the quality of their predictive distribution over outcomes. We identify two main perspectives that drive starkly different evaluation protocols. The first treats predictive probability as an indication of model confidence; the second as an indication of human label variation. We discuss their merits and limitations, and take the position that both are crucial for trustworthy and fair NLP systems, but that exploiting a single predictive distribution is limiting. We recommend tools and highlight exciting directions towards models with disentangled representations of uncertainty about predictions and uncertainty about human labels.

The labor market is changing rapidly, prompting increased interest in the automatic extraction of occupational skills from text. With the advent of English benchmark job description datasets, there is a need for systems that handle their diversity well. We tackle the complexity in occupational skill datasets tasks—combining and leveraging multiple datasets for skill extraction, to identify rarely observed skills within a dataset, and overcoming the scarcity of skills across datasets. In particular, we investigate the retrieval-augmentation of language models, employing an external datastore for retrieving similar skills in a dataset-unifying manner. Our proposed method, Nearest Neighbor Occupational Skill Extraction (NNOSE) effectively leverages multiple datasets by retrieving neighboring skills from other datasets in the datastore. This improves skill extraction without additional fine-tuning. Crucially, we observe a performance gain in predicting infrequent patterns, with substantial gains of up to 30% span-F1 in cross-dataset settings.

In Natural Language Processing, entity linking (EL) has centered around Wikipedia, but yet remains underexplored for the job market domain. Disambiguating skill mentions can help us get insight into the current labor market demands. In this work, we are the first to explore EL in this domain, specifically targeting the linkage of occupational skills to the ESCO taxonomy (le Vrang et al., 2014). Previous efforts linked coarse-grained (full) sentences to a corresponding ESCO skill. In this work, we link more fine-grained span-level mentions of skills. We tune two high-performing neural EL models, a bi-encoder (Wu et al., 2020) and an autoregressive model (Cao et al., 2021), on a synthetically generated mention–skill pair dataset and evaluate them on a human-annotated skill-linking benchmark. Our findings reveal that both models are capable of linking implicit mentions of skills to their correct taxonomy counterparts. Empirically, BLINK outperforms GENRE in strict evaluation, but GENRE performs better in loose evaluation (accuracy@k).

Annotation tools are the starting point for creating Natural Language Processing (NLP) datasets. There is a wide variety of tools available; setting up these tools is however a hindrance. We propose EEVEE, an annotation tool focused on simplicity, efficiency, and ease of use. It can run directly in the browser (no setup required) and uses tab-separated files (as opposed to character offsets or task-specific formats) for annotation. It allows for annotation of multiple tasks on a single dataset and supports four task-types: sequence labeling, span labeling, text classification and seq2seq.

Instruction tuning has become an integral part of training pipelines for Large Language Models (LLMs) and has been shown to yield strong performance gains. In an orthogonal line of research, Annotation Error Detection (AED) has emerged as a tool for detecting quality problems in gold standard labels. So far, however, the application of AED methods has been limited to classification tasks. It is an open question how well AED methods generalize to language generation settings, which are becoming more widespread via LLMs. In this paper, we present a first and novel benchmark for AED on instruction tuning data: DONKII. It comprises three instruction-tuning datasets enriched with error annotations by experts and semi-automatic methods. We also provide a novel taxonomy of error types for instruction-tuning data. We find that all three datasets contain clear errors, which sometimes propagate directly into instruction-tuned LLMs. We propose four AED baselines for the generative setting and evaluate them extensively on the newly introduced dataset. Our results show that the choice of the right AED method and model size is indeed crucial and derive practical recommendations for how to use AED methods to clean instruction-tuning data.

The convergence of embodied agents and large language models (LLMs) has brought significant advancements to embodied instruction following.Particularly, the strong reasoning capabilities of LLMs make it possible for robots to perform long-horizon tasks without expensive annotated demonstrations.However, public benchmarks for testing the long-horizon reasoning capabilities of language-conditioned robots in various scenarios are still missing. To fill this gap, this work focuses on the tabletopmanipulation task and releases a simulation benchmark,textit{LoHoRavens}, which covers various long-horizonreasoning aspects spanning color, size, space, arithmeticsand reference.Furthermore, there is a key modality bridging problem forlong-horizon manipulation tasks with LLMs: how toincorporate the observation feedback during robot executionfor the LLM’s closed-loop planning, which is however less studied by prior work. We investigate two methods of bridging the modality gap: caption generation and learnable interface for incorporating explicit and implicit observation feedback to the LLM, respectively.These methods serve as the two baselines for our proposed benchmark. Experiments show that both methods struggle to solve most tasks, indicating long-horizon manipulation tasks are still challenging for current popular models.We expect the proposed public benchmark and baselines can help the community develop better models for long-horizon tabletop manipulation tasks.

In Natural Language Generation (NLG) tasks, for any input, multiple communicative goals are plausible, and any goal can be put into words, or produced, in multiple ways. We characterise the extent to which human production varies lexically, syntactically, and semantically across four NLG tasks, connecting human production variability to aleatoric or data uncertainty. We then inspect the space of output strings shaped by a generation system’s predicted probability distribution and decoding algorithm to probe its uncertainty. For each test input, we measure the generator’s calibration to human production variability. Following this instance-level approach, we analyse NLG models and decoding strategies, demonstrating that probing a generator with multiple samples and, when possible, multiple references, provides the level of detail necessary to gain understanding of a model’s representation of uncertainty.

Language understanding is a multi-faceted cognitive capability, which the Natural Language Processing (NLP) community has striven to model computationally for decades. Traditionally, facets of linguistic intelligence have been compartmentalized into tasks with specialized model architectures and corresponding evaluation protocols. With the advent of large language models (LLMs) the community has witnessed a dramatic shift towards general purpose, task-agnostic approaches powered by generative models. As a consequence, the traditional compartmentalized notion of language tasks is breaking down, followed by an increasing challenge for evaluation and analysis. At the same time, LLMs are being deployed in more real-world scenarios, including previously unforeseen zero-shot setups, increasing the need for trustworthy and reliable systems. Therefore, we argue that it is time to rethink what constitutes tasks and model evaluation in NLP, and pursue a more holistic view on language, placing trustworthiness at the center. Towards this goal, we review existing compartmentalized approaches for understanding the origins of a model’s functional capacity, and provide recommendations for more multi-faceted evaluation protocols.

Label aggregation such as majority voting is commonly used to resolve annotator disagreement in dataset creation. However, this may disregard minority values and opinions. Recent studies indicate that learning from individual annotations outperforms learning from aggregated labels, though they require a considerable amount of annotation. Active learning, as an annotation cost-saving strategy, has not been fully explored in the context of learning from disagreement. We show that in the active learning setting, a multi-head model performs significantly better than a single-head model in terms of uncertainty estimation. By designing and evaluating acquisition functions with annotator-specific heads on two datasets, we show that group-level entropy works generally well on both datasets. Importantly, it achieves performance in terms of both prediction and uncertainty estimation comparable to full-scale training from disagreement, while saving 70% of the annotation budget.

In legal NLP, Case Outcome Classification (COC) must not only be accurate but also trustworthy and explainable. Existing work in explainable COC has been limited to annotations by a single expert. However, it is well-known that lawyers may disagree in their assessment of case facts. We hence collect a novel dataset RaVE: Rationale Variation in ECHR, which is obtained from two experts in the domain of international human rights law, for whom we observe weak agreement. We study their disagreements and build a two-level task-independent taxonomy, supplemented with COC-specific subcategories. To our knowledge, this is the first work in the legal NLP that focuses on human label variation. We quantitatively assess different taxonomy categories and find that disagreements mainly stem from underspecification of the legal context, which poses challenges given the typically limited granularity and noise in COC metadata. We further assess the explainablility of state-of-the-art COC models on RaVE and observe limited agreement between models and experts. Overall, our case study reveals hitherto underappreciated complexities in creating benchmark datasets in legal NLP that revolve around identifying aspects of a case’s facts supposedly relevant for its outcome.

Representational spaces learned via language modeling are fundamental to Natural Language Processing (NLP), however there has been limited understanding regarding how and when during training various types of linguistic information emerge and interact. Leveraging a novel information theoretic probing suite, which enables direct comparisons of not just task performance, but their representational subspaces, we analyze nine tasks covering syntax, semantics and reasoning, across 2M pre-training steps and five seeds. We identify critical learning phases across tasks and time, during which subspaces emerge, share information, and later disentangle to specialize. Across these phases, syntactic knowledge is acquired rapidly after 0.5% of full training. Continued performance improvements primarily stem from the acquisition of open-domain knowledge, while semantics and reasoning tasks benefit from later boosts to long-range contextualization and higher specialization. Measuring cross-task similarity further reveals that linguistically related tasks share information throughout training, and do so more during the critical phase of learning than before or after. Our findings have implications for model interpretability, multi-task learning, and learning from limited data.

Manually annotated datasets are crucial for training and evaluating Natural Language Processing models. However, recent work has discovered that even widely-used benchmark datasets contain a substantial number of erroneous annotations. This problem has been addressed with Annotation Error Detection (AED) models, which can flag such errors for human re-annotation. However, even though many of these AED methods assume a final curation step in which a human annotator decides whether the annotation is erroneous, they have been developed as static models without any human-in-the-loop component. In this work, we propose ActiveAED, an AED method that can detect errors more accurately by repeatedly querying a human for error corrections in its prediction loop. We evaluate ActiveAED on eight datasets spanning five different tasks and find that it leads to improvements over the state of the art on seven of them, with gains of up to six percentage points in average precision.

Recent advances of powerful Language Models have allowed Natural Language Generation (NLG) to emerge as an important technology that can not only perform traditional tasks like summarisation or translation, but also serve as a natural language interface to a variety of applications. As such, it is crucial that NLG systems are trustworthy and reliable, for example by indicating when they are likely to be wrong; and supporting multiple views, backgrounds and writing styles – reflecting diverse human sub-populations. In this paper, we argue that a principled treatment of uncertainty can assist in creating systems and evaluation protocols better aligned with these goals. We first present the fundamental theory, frameworks and vocabulary required to represent uncertainty. We then characterise the main sources of uncertainty in NLG from a linguistic perspective, and propose a two-dimensional taxonomy that is more informative and faithful than the popular aleatoric/epistemic dichotomy. Finally, we move from theory to applications and highlight exciting research directions that exploit uncertainty to power decoding, controllable generation, self-assessment, selective answering, active learning and more.

Despite much progress in recent years, the vast majority of work in natural language processing (NLP) is on standard languages with many speakers. In this work, we instead focus on low-resource languages and in particular non-standardized low-resource languages. Even within branches of major language families, often considered well-researched, little is known about the extent and type of available resources and what the major NLP challenges are for these language varieties. The first step to address this situation is a systematic survey of available corpora (most importantly, annotated corpora, which are particularly valuable for NLP research). Focusing on Germanic low-resource language varieties, we provide such a survey in this paper. Except for geolocation (origin of speaker or document), we find that manually annotated linguistic resources are sparse and, if they exist, mostly cover morphosyntax. Despite this lack of resources, we observe that interest in this area is increasing: there is active development and a growing research community. To facilitate research, we make our overview of over 80 corpora publicly available.

Recently, various intermediate layer distillation (ILD) objectives have been shown to improve compression of BERT models via Knowledge Distillation (KD). However, a comprehensive evaluation of the objectives in both task-specific and task-agnostic settings is lacking. To the best of our knowledge, this is the first work comprehensively evaluating distillation objectives in both settings. We show that attention transfer gives the best performance overall. We also study the impact of layer choice when initializing the student from the teacher layers, finding a significant impact on the performance in task-specific distillation. For vanilla KD and hidden states transfer, initialisation with lower layers of the teacher gives a considerable improvement over higher layers, especially on the task of QNLI (up to an absolute percentage change of 17.8 in accuracy). Attention transfer behaves consistently under different initialisation settings. We release our code as an efficient transformer-based model distillation framework for further studies.

One of the challenges with finetuning pretrained language models (PLMs) is that their tokenizer is optimized for the language(s) it was pretrained on, but brittle when it comes to previously unseen variations in the data. This can for instance be observed when finetuning PLMs on one language and evaluating them on data in a closely related language variety with no standardized orthography. Despite the high linguistic similarity, tokenization no longer corresponds to meaningful representations of the target data, leading to low performance in, e.g., part-of-speech tagging. In this work, we finetune PLMs on seven languages from three different families and analyze their zero-shot performance on closely related, non-standardized varieties. We consider different measures for the divergence in the tokenization of the source and target data, and the way they can be adjusted by manipulating the tokenization during the finetuning step. Overall, we find that the similarity between the percentage of words that get split into subwords in the source and target data (the split word ratio difference) is the strongest predictor for model performance on target data.

Calibration is a popular framework to evaluate whether a classifier knows when it does not know - i.e., its predictive probabilities are a good indication of how likely a prediction is to be correct. Correctness is commonly estimated against the human majority class. Recently, calibration to human majority has been measured on tasks where humans inherently disagree about which class applies. We show that measuring calibration to human majority given inherent disagreements is theoretically problematic, demonstrate this empirically on the ChaosNLI dataset, and derive several instance-level measures of calibration that capture key statistical properties of human judgements - including class frequency, ranking and entropy.

With the increase in availability of large pre-trained language models (LMs) in Natural Language Processing (NLP), it becomes critical to assess their fit for a specific target task a priori—as fine-tuning the entire space of available LMs is computationally prohibitive and unsustainable. However, encoder transferability estimation has received little to no attention in NLP. In this paper, we propose to generate quantitative evidence to predict which LM, out of a pool of models, will perform best on a target task without having to fine-tune all candidates. We provide a comprehensive study on LM ranking for 10 NLP tasks spanning the two fundamental problem types of classification and structured prediction. We adopt the state-of-the-art Logarithm of Maximum Evidence (LogME) measure from Computer Vision (CV) and find that it positively correlates with final LM performance in 94% of the setups.In the first study of its kind, we further compare transferability measures with the de facto standard of human practitioner ranking, finding that evidence from quantitative metrics is more robust than pure intuition and can help identify unexpected LM candidates.

Linguistic information is encoded at varying timescales (subwords, phrases, etc.) and communicative levels, such as syntax and semantics. Contextualized embeddings have analogously been found to capture these phenomena at distinctive layers and frequencies. Leveraging these findings, we develop a fully learnable frequency filter to identify spectral profiles for any given task. It enables vastly more granular analyses than prior handcrafted filters, and improves on efficiency. After demonstrating the informativeness of spectral probing over manual filters in a monolingual setting, we investigate its multilingual characteristics across seven diverse NLP tasks in six languages. Our analyses identify distinctive spectral profiles which quantify cross-task similarity in a linguistically intuitive manner, while remaining consistent across languages—highlighting their potential as robust, lightweight task descriptors.

Human variation in labeling is often considered noise. Annotation projects for machine learning (ML) aim at minimizing human label variation, with the assumption to maximize data quality and in turn optimize and maximize machine learning metrics. However, thisconventional practice assumes that there exists a ground truth, and neglects that there exists genuine human variation in labeling due to disagreement, subjectivity in annotation or multiple plausible answers. In this position paper, we argue that this big open problem of human label variation persists and critically needs more attention to move our field forward. This is because human label variation impacts all stages of the ML pipeline: data, modeling and evaluation. However, few works consider all of these dimensions jointly; and existing research is fragmented. We reconcile different previously proposed notions of human label variation, provide a repository of publicly-available datasets with un-aggregated labels, depict approaches proposed so far, identify gaps and suggest ways forward. As datasets are becoming increasingly available, we hope that this synthesized view on the ‘problem’ will lead to an open discussion on possible strategies to devise fundamentally new directions.

Relation Extraction (RE) has attracted increasing attention, but current RE evaluation is limited to in-domain evaluation setups. Little is known on how well a RE system fares in challenging, but realistic out-of-distribution evaluation setups. To address this gap, we propose CrossRE, a new, freely-available cross-domain benchmark for RE, which comprises six distinct text domains and includes multi-label annotations. An additional innovation is that we release meta-data collected during annotation, to include explanations and flags of difficult instances. We provide an empirical evaluation with a state-of-the-art model for relation classification. As the meta-data enables us to shed new light on the state-of-the-art model, we provide a comprehensive analysis on the impact of difficult cases and find correlations between model and human annotations. Overall, our empirical investigation highlights the difficulty of cross-domain RE. We release our dataset, to spur more research in this direction.

The field of Deep Learning (DL) has undergone explosive growth during the last decade, with a substantial impact on Natural Language Processing (NLP) as well. Yet, compared to more established disciplines, a lack of common experimental standards remains an open challenge to the field at large. Starting from fundamental scientific principles, we distill ongoing discussions on experimental standards in NLP into a single, widely-applicable methodology. Following these best practices is crucial to strengthen experimental evidence, improve reproducibility and enable scientific progress. These standards are further collected in a public repository to help them transparently adapt to future needs.