Christian Hardmeier

Dennis Ulmer, Jes Frellsen, Christian Hardmeier

We investigate the problem of determining the predictive confidence (or, conversely, uncertainty) of a neural classifier through the lens of low-resource languages. By training models on sub-sampled datasets in three different languages, we assess the quality of estimates from a wide array of approaches and their dependence on the amount of available data. We find that while approaches based on pre-trained models and ensembles achieve the best results overall, the quality of uncertainty estimates can surprisingly suffer with more data. We also perform a qualitative analysis of uncertainties on sequences, discovering that a model's total uncertainty seems to be influenced to a large degree by its data uncertainty, not model uncertainty. All model implementations are open-sourced in a software package.

Marco Antonio Stranisci, A Pranav, Rossana Damiano et al.

Modern language models rely on pretraining filters to remove undesirable content from training corpora and inference-time guardrails to suppress undesirable outputs during deployment. In this paper, we examine how these filtering and moderation decisions produce forms of epistemic erasure and reveal tensions both across automated systems and between these systems and human judgment. We audit four pretraining filters and three inference-time guardrails on Common Crawl sentences containing gender and regional-origin mentions, together with a manually annotated subset of 500 sentences. Our analysis shows that filtering and guardrail decisions are strongly associated with blocklist-based lexical cues, while frequently failing to flag content containing private information or explicit hate speech. At the same time, marginalized groups, particularly transgender people, women, and Central Americans, are significantly over-flagged across systems. Human annotators, by contrast, would retain 88.5\% of filter-flagged and 91.3\% of guardrail-flagged content, often recognizing representational harms arising from tensions of content removal that current systems fail to capture. Taken together, our findings document a form of epistemic erasure in which mentions of marginalized groups are disproportionately removed before pretraining and additionally suppressed again at inference time.

Dennis Ulmer, Elisa Bassignana, Max Müller-Eberstein et al.

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 support scientific progress. These standards are further collected in a public repository to help them transparently adapt to future needs.

Dennis Ulmer, Christian Hardmeier, Jes Frellsen

A lot of Machine Learning (ML) and Deep Learning (DL) research is of an empirical nature. Nevertheless, statistical significance testing (SST) is still not widely used. This endangers true progress, as seeming improvements over a baseline might be statistical flukes, leading follow-up research astray while wasting human and computational resources. Here, we provide an easy-to-use package containing different significance tests and utility functions specifically tailored towards research needs and usability.

Nils Grünefeld, Jes Frellsen, Christian Hardmeier

Existing methods for quantifying predictive uncertainty in neural networks are either computationally intractable for large language models or require access to training data that is typically unavailable. We derive a lightweight alternative through two approximations: a first-order Taylor expansion that expresses uncertainty in terms of the gradient of the prediction and the parameter covariance, and an isotropy assumption on the parameter covariance. Together, these yield epistemic uncertainty as the squared gradient norm and aleatoric uncertainty as the Bernoulli variance of the point prediction, from a single forward-backward pass through an unmodified pretrained model. We justify the isotropy assumption by showing that covariance estimates built from non-training data introduce structured distortions that isotropic covariance avoids, and that theoretical results on the spectral properties of large networks support the approximation at scale. Validation against reference Markov Chain Monte Carlo estimates on synthetic problems shows strong correspondence that improves with model size. We then use the estimates to investigate when each uncertainty type carries useful signal for predicting answer correctness in question answering with large language models, revealing a benchmark-dependent divergence: the combined estimate achieves the highest mean AUROC on TruthfulQA, where questions involve genuine conflict between plausible answers, but falls to near chance on TriviaQA's factual recall, suggesting that parameter-level uncertainty captures a fundamentally different signal than self-assessment methods.

Daniel Varab, Christian Hardmeier

Recent work has suggested that end-to-end system designs for cross-lingual summarization are competitive solutions that perform on par or even better than traditional pipelined designs. A closer look at the evidence reveals that this intuition is based on the results of only a handful of languages or using underpowered pipeline baselines. In this work, we compare these two paradigms for cross-lingual summarization on 39 source languages into English and show that a simple \textit{translate-then-summarize} pipeline design consistently outperforms even an end-to-end system with access to enormous amounts of parallel data. For languages where our pipeline model does not perform well, we show that system performance is highly correlated with publicly distributed BLEU scores, allowing practitioners to establish the feasibility of a language pair a priori. Contrary to recent publication trends, our result suggests that the combination of individual progress of monolingual summarization and translation tasks offers better performance than an end-to-end system, suggesting that end-to-end designs should be considered with care.

Nils Grünefeld, Bertram Højer, Philipp Mondorf et al.

Language model (LM) "reasoning", commonly described as Chain-of-Thought or test-time scaling, often improves benchmark performance, but the dynamics underlying this process remain poorly understood. We study these dynamics through the lens of uncertainty quantification by treating the "reasoning" traces, the intermediate token sequences generated by LMs, as evolving model states. We summarize each trace by an uncertainty trace profile: a small set of features describing the shape of the uncertainty signal over its trace, such as its slope and linearity. We find that across five LMs evaluated on GSM8K and ProntoQA, these profiles predict whether a trace yields a correct final answer with AUROC up to 0.807, improving markedly on recent related work. We reach AUROC 0.801 using only the first few hundred tokens of full traces, suggesting that errors can be detected early in the generation. A detailed comparison of correct and incorrect traces further reveals qualitatively distinct uncertainty profiles, with correct traces showing a steeper and less linear decline in uncertainty. Together, the results suggest that our method, grounded in decision-making under uncertainty, provides a principled lens for studying the generative process underlying LM "reasoning".

Paul Engelmann, Peter Brunsgaard Trolle, Christian Hardmeier

Dehumanization is a mental process that enables the exclusion and ill treatment of a group of people. In this paper, we present two data sets of dehumanizing text, a large, automatically collected corpus and a smaller, manually annotated data set. Both data sets include a combination of political discourse and dialogue from movie subtitles. Our methods give us a broad and varied amount of dehumanization data to work with, enabling further exploratory analysis and automatic classification of dehumanization patterns. Both data sets will be publicly released.

Dennis Ulmer, Alexandra Lorson, Ivan Titov et al.

Human users increasingly rely on natural language interactions with large language models (LLMs) in order to receive help on a large variety of tasks and problems. However, the trustworthiness and perceived legitimacy of LLMs is undermined by the fact that their output is frequently stated in very confident terms, even when its accuracy is questionable. Therefore, there is a need to signal the confidence of the language model to a user in order to reap the benefits of human-machine collaboration and mitigate potential harms. Verbalized uncertainty is the expression of confidence with linguistic means, an approach that integrates perfectly into language-based interfaces. Nevertheless, most recent research in natural language processing (NLP) overlooks the nuances surrounding human uncertainty communication and the data biases that influence machine uncertainty communication. We argue for anthropomimetic uncertainty, meaning that intuitive and trustworthy uncertainty communication requires a degree of linguistic authenticity and personalization to the user, which could be achieved by emulating human communication. We present a thorough overview over the research in human uncertainty communication, survey ongoing research, and perform additional analyses to demonstrate so-far overlooked biases in verbalized uncertainty. We conclude by pointing out unique factors in human-machine communication of uncertainty and deconstruct anthropomimetic uncertainty into future research directions for NLP.

Marco Antonio Stranisci, Christian Hardmeier

Data filtering strategies are a crucial component to develop safe Large Language Models (LLM), since they support the removal of harmful contents from pretraining datasets. There is a lack of research on the actual impact of these strategies on vulnerable groups to discrimination, though, and their effectiveness has not been yet systematically addressed. In this paper we present a benchmark study of data filtering strategies for harm reduction aimed at providing a systematic evaluation on these approaches. We provide an overview $55$ technical reports of English LMs and LLMs to identify the existing filtering strategies in literature and implement an experimental setting to test their impact against vulnerable groups. Our results show that the positive impact that strategies have in reducing harmful contents from documents has the side effect of increasing the underrepresentation of vulnerable groups to discrimination in datasets. WARNING: the paper could contain racist, sexist, violent, and generally offensive contents

Ahmed Ruby, Christian Hardmeier, Sara Stymne

Implicit discourse relation classification is a challenging task, as it requires inferring meaning from context. While contextual cues can be distributed across modalities and vary across languages, they are not always captured by text alone. To address this, we introduce an automatic method for distantly related and unrelated language pairs to construct a multilingual and multimodal dataset for implicit discourse relations in English, French, and Spanish. For classification, we propose a multimodal approach that integrates textual and acoustic information through Qwen2-Audio, allowing joint modeling of text and audio for implicit discourse relation classification across languages. We find that while text-based models outperform audio-based models, integrating both modalities can enhance performance, and cross-lingual transfer can provide substantial improvements for low-resource languages.

Andreas Motzfeldt, Joakim Edin, Casper L. Christensen et al.

In medical coding, experts map unstructured clinical notes to alphanumeric codes for diagnoses and procedures. We introduce Code Like Humans: a new agentic framework for medical coding with large language models. It implements official coding guidelines for human experts, and it is the first solution that can support the full ICD-10 coding system (+70K labels). It achieves the best performance to date on rare diagnosis codes (fine-tuned discriminative classifiers retain an advantage for high-frequency codes, to which they are limited). Towards future work, we also contribute an analysis of system performance and identify its `blind spots' (codes that are systematically undercoded).

Jorge del Pozo Lérida, Kamil Kojs, János Máté et al.

Large Language Models (LLMs) have become state-of-the-art in Machine Translation (MT), often trained on massive bilingual parallel corpora scraped from the web, that contain low-quality entries and redundant information, leading to significant computational challenges. Various data filtering methods exist to reduce dataset sizes, but their effectiveness largely varies based on specific language pairs and domains. This paper evaluates the impact of commonly used data filtering techniques, such as LASER, MUSE, and LaBSE, on English-Polish translation within the biomedical domain. By filtering the UFAL Medical Corpus, we created varying dataset sizes to fine-tune the mBART50 model, which was then evaluated using the SacreBLEU metric on the Khresmoi dataset, having the quality of translations assessed by bilingual speakers. Our results show that both LASER and MUSE can significantly reduce dataset sizes while maintaining or even enhancing performance. We recommend the use of LASER, as it consistently outperforms the other methods and provides the most fluent and natural-sounding translations.

Gongbo Tang, Christian Hardmeier

Most pronouns are referring expressions, computers need to resolve what do the pronouns refer to, and there are divergences on pronoun usage across languages. Thus, dealing with these divergences and translating pronouns is a challenge in machine translation. Mentions are referring candidates of pronouns and have closer relations with pronouns compared to general tokens. We assume that extracting additional mention features can help pronoun translation. Therefore, we introduce an additional mention attention module in the decoder to pay extra attention to source mentions but not non-mention tokens. Our mention attention module not only extracts features from source mentions, but also considers target-side context which benefits pronoun translation. In addition, we also introduce two mention classifiers to train models to recognize mentions, whose outputs guide the mention attention. We conduct experiments on the WMT17 English-German translation task, and evaluate our models on general translation and pronoun translation, using BLEU, APT, and contrastive evaluation metrics. Our proposed model outperforms the baseline Transformer model in terms of APT and BLEU scores, this confirms our hypothesis that we can improve pronoun translation by paying additional attention to source mentions, and shows that our introduced additional modules do not have negative effect on the general translation quality.

Gongbo Tang, Christian Hardmeier

Coreference resolution is the task of finding expressions that refer to the same entity in a text. Coreference models are generally trained on monolingual annotated data but annotating coreference is expensive and challenging. Hardmeier et al.(2013) have shown that parallel data contains latent anaphoric knowledge, but it has not been explored in end-to-end neural models yet. In this paper, we propose a simple yet effective model to exploit coreference knowledge from parallel data. In addition to the conventional modules learning coreference from annotations, we introduce an unsupervised module to capture cross-lingual coreference knowledge. Our proposed cross-lingual model achieves consistent improvements, up to 1.74 percentage points, on the OntoNotes 5.0 English dataset using 9 different synthetic parallel datasets. These experimental results confirm that parallel data can provide additional coreference knowledge which is beneficial to coreference resolution tasks.

Sharid Loáiciga, Luca Bevacqua, Christian Hardmeier

We present an unsupervised method to detect English unergative and unaccusative verbs. These categories allow us to identify verbs participating in the causative-inchoative alternation without knowing the semantic roles of the verb. The method is based on the generation of intransitive sentence variants of candidate verbs and probing a language model. We obtained results on par with similar approaches, with the added benefit of not relying on annotated resources.

Dennis Ulmer, Christian Hardmeier, Jes Frellsen

Popular approaches for quantifying predictive uncertainty in deep neural networks often involve distributions over weights or multiple models, for instance via Markov Chain sampling, ensembling, or Monte Carlo dropout. These techniques usually incur overhead by having to train multiple model instances or do not produce very diverse predictions. This comprehensive and extensive survey aims to familiarize the reader with an alternative class of models based on the concept of Evidential Deep Learning: For unfamiliar data, they aim to admit "what they don't know", and fall back onto a prior belief. Furthermore, they allow uncertainty estimation in a single model and forward pass by parameterizing distributions over distributions. This survey recapitulates existing works, focusing on the implementation in a classification setting, before surveying the application of the same paradigm to regression. We also reflect on the strengths and weaknesses compared to other existing methods and provide the most fundamental derivations using a unified notation to aid future research.

Christian Hardmeier, Marta R. Costa-jussà, Kellie Webster et al.

At the Workshop on Gender Bias in NLP (GeBNLP), we'd like to encourage authors to give explicit consideration to the wider aspects of bias and its social implications. For the 2020 edition of the workshop, we therefore requested that all authors include an explicit bias statement in their work to clarify how their work relates to the social context in which NLP systems are used. The programme committee of the workshops included a number of reviewers with a background in the humanities and social sciences, in addition to NLP experts doing the bulk of the reviewing. Each paper was assigned one of those reviewers, and they were asked to pay specific attention to the provided bias statements in their reviews. This initiative was well received by the authors who submitted papers to the workshop, several of whom said they received useful suggestions and literature hints from the bias reviewers. We are therefore planning to keep this feature of the review process in future editions of the workshop.

Ali Basirat, Christian Hardmeier, Joakim Nivre

We generalize principal component analysis for embedding words into a vector space. The generalization is made in two major levels. The first is to generalize the concept of the corpus as a counting process which is defined by three key elements vocabulary set, feature (annotation) set, and context. This generalization enables the principal word embedding method to generate word vectors with regard to different types of contexts and different types of annotations provided for a corpus. The second is to generalize the transformation step used in most of the word embedding methods. To this end, we define two levels of transformations. The first is a quadratic transformation, which accounts for different types of weighting over the vocabulary units and contextual features. Second is an adaptive non-linear transformation, which reshapes the data distribution to be meaningful to principal component analysis. The effect of these generalizations on the word vectors is intrinsically studied with regard to the spread and the discriminability of the word vectors. We also provide an extrinsic evaluation of the contribution of the principal word vectors on a word similarity benchmark and the task of dependency parsing. Our experiments are finalized by a comparison between the principal word vectors and other sets of word vectors generated with popular word embedding methods. The results obtained from our intrinsic evaluation metrics show that the spread and the discriminability of the principal word vectors are higher than that of other word embedding methods. The results obtained from the extrinsic evaluation metrics show that the principal word vectors are better than some of the word embedding methods and on par with popular methods of word embedding.

Liane Guillou, Christian Hardmeier, Preslav Nakov et al.

We describe the design, the evaluation setup, and the results of the 2016 WMT shared task on cross-lingual pronoun prediction. This is a classification task in which participants are asked to provide predictions on what pronoun class label should replace a placeholder value in the target-language text, provided in lemmatised and PoS-tagged form. We provided four subtasks, for the English-French and English-German language pairs, in both directions. Eleven teams participated in the shared task; nine for the English-French subtask, five for French-English, nine for English-German, and six for German-English. Most of the submissions outperformed two strong language-model based baseline systems, with systems using deep recurrent neural networks outperforming those using other architectures for most language pairs.

Eva Vanmassenhove, Christian Hardmeier, Andy Way

Speakers of different languages must attend to and encode strikingly different aspects of the world in order to use their language correctly (Sapir, 1921; Slobin, 1996). One such difference is related to the way gender is expressed in a language. Saying "I am happy" in English, does not encode any additional knowledge of the speaker that uttered the sentence. However, many other languages do have grammatical gender systems and so such knowledge would be encoded. In order to correctly translate such a sentence into, say, French, the inherent gender information needs to be retained/recovered. The same sentence would become either "Je suis heureux", for a male speaker or "Je suis heureuse" for a female one. Apart from morphological agreement, demographic factors (gender, age, etc.) also influence our use of language in terms of word choices or even on the level of syntactic constructions (Tannen, 1991; Pennebaker et al., 2003). We integrate gender information into NMT systems. Our contribution is two-fold: (1) the compilation of large datasets with speaker information for 20 language pairs, and (2) a simple set of experiments that incorporate gender information into NMT for multiple language pairs. Our experiments show that adding a gender feature to an NMT system significantly improves the translation quality for some language pairs.

Christian Hardmeier, Liane Guillou

Pronouns are a long-standing challenge in machine translation. We present a study of the performance of a range of rule-based, statistical and neural MT systems on pronoun translation based on an extensive manual evaluation using the PROTEST test suite, which enables a fine-grained analysis of different pronoun types and sheds light on the difficulties of the task. We find that the rule-based approaches in our corpus perform poorly as a result of oversimplification, whereas SMT and early NMT systems exhibit significant shortcomings due to a lack of awareness of the functional and referential properties of pronouns. A recent Transformer-based NMT system with cross-sentence context shows very promising results on non-anaphoric pronouns and intra-sentential anaphora, but there is still considerable room for improvement in examples with cross-sentence dependencies.

Liane Guillou, Christian Hardmeier

We compare the performance of the APT and AutoPRF metrics for pronoun translation against a manually annotated dataset comprising human judgements as to the correctness of translations of the PROTEST test suite. Although there is some correlation with the human judgements, a range of issues limit the performance of the automated metrics. Instead, we recommend the use of semi-automatic metrics and test suites in place of fully automatic metrics.

Yan Shao, Christian Hardmeier, Joakim Nivre

Word segmentation is a low-level NLP task that is non-trivial for a considerable number of languages. In this paper, we present a sequence tagging framework and apply it to word segmentation for a wide range of languages with different writing systems and typological characteristics. Additionally, we investigate the correlations between various typological factors and word segmentation accuracy. The experimental results indicate that segmentation accuracy is positively related to word boundary markers and negatively to the number of unique non-segmental terms. Based on the analysis, we design a small set of language-specific settings and extensively evaluate the segmentation system on the Universal Dependencies datasets. Our model obtains state-of-the-art accuracies on all the UD languages. It performs substantially better on languages that are non-trivial to segment, such as Chinese, Japanese, Arabic and Hebrew, when compared to previous work.

Yan Shao, Christian Hardmeier, Jörg Tiedemann et al.

We present a character-based model for joint segmentation and POS tagging for Chinese. The bidirectional RNN-CRF architecture for general sequence tagging is adapted and applied with novel vector representations of Chinese characters that capture rich contextual information and lower-than-character level features. The proposed model is extensively evaluated and compared with a state-of-the-art tagger respectively on CTB5, CTB9 and UD Chinese. The experimental results indicate that our model is accurate and robust across datasets in different sizes, genres and annotation schemes. We obtain state-of-the-art performance on CTB5, achieving 94.38 F1-score for joint segmentation and POS tagging.

Daniel Beck, Trevor Cohn, Christian Hardmeier et al.

Structural kernels are a flexible learning paradigm that has been widely used in Natural Language Processing. However, the problem of model selection in kernel-based methods is usually overlooked. Previous approaches mostly rely on setting default values for kernel hyperparameters or using grid search, which is slow and coarse-grained. In contrast, Bayesian methods allow efficient model selection by maximizing the evidence on the training data through gradient-based methods. In this paper we show how to perform this in the context of structural kernels by using Gaussian Processes. Experimental results on tree kernels show that this procedure results in better prediction performance compared to hyperparameter optimization via grid search. The framework proposed in this paper can be adapted to other structures besides trees, e.g., strings and graphs, thereby extending the utility of kernel-based methods.