Simon Clematide

Andrianos Michail, Elias Schuhmacher, Juri Opitz et al.

Dense retrieval models exhibit positional bias: retrieval effectiveness degrades when relevant information appears later in a passage (Zeng et al., 2025). We ask whether this bias can be reduced at inference time, without retraining and without sacrificing overall retrieval effectiveness. To this end, we adapt inference-time attention calibration (Schuhmacher et al., 2026) to downstream retrieval and extend it with a strength coefficient lambda that interpolates between the original and fully calibrated attention distributions. Across three embedding models on SQuAD-PosQ and FineWeb-PosQ, we examine how basket size, calibrated layer set, and strength affect the trade-off between positional fairness and retrieval effectiveness, finding that partial calibration frequently outperforms full calibration. A single configuration (B=128, lambda=0.5, 50% layer depth) improves the harmonic mean of nDCG@10 across positional groups on FineWeb-PosQ for all three models without per-model tuning, and applies to both <s>-pooled and last-token-pooled architectures. This default configuration transfers without modification to PosIR, which spans 10 languages and 31 domains, reducing the Position Sensitivity Index in all 16 length-quartile x model x retrieval-setting combinations, while preserving or improving aggregate nDCG@10. We release our extended codebase at https://github.com/impresso/fair-sentence-transformers

Phillip Benjamin Ströbel, Simon Clematide, Martin Volk et al.

We apply the TrOCR framework to real-world, historical manuscripts and show that TrOCR per se is a strong model, ideal for transfer learning. TrOCR has been trained on English only, but it can adapt to other languages that use the Latin alphabet fairly easily and with little training material. We compare TrOCR against a SOTA HTR framework (Transkribus) and show that it can beat such systems. This finding is essential since Transkribus performs best when it has access to baseline information, which is not needed at all to fine-tune TrOCR.

Andrianos Michail, Simon Clematide, Juri Opitz

The task of determining whether two texts are paraphrases has long been a challenge in NLP. However, the prevailing notion of paraphrase is often quite simplistic, offering only a limited view of the vast spectrum of paraphrase phenomena. Indeed, we find that evaluating models in a paraphrase dataset can leave uncertainty about their true semantic understanding. To alleviate this, we create PARAPHRASUS, a benchmark designed for multi-dimensional assessment, benchmarking and selection of paraphrase detection models. We find that paraphrase detection models under our fine-grained evaluation lens exhibit trade-offs that cannot be captured through a single classification dataset. Furthermore, PARAPHRASUS allows prompt calibration for different use cases, tailoring LLM models to specific strictness levels. PARAPHRASUS includes 3 challenges spanning over 10 datasets, including 8 repurposed and 2 newly annotated; we release it along with a benchmarking library at https://github.com/impresso/paraphrasus

Andrianos Michail, Stefanos Konstantinou, Simon Clematide

This paper describes the submission of UZH_CLyp for the SemEval 2023 Task 9 "Multilingual Tweet Intimacy Analysis". We achieved second-best results in all 10 languages according to the official Pearson's correlation regression evaluation measure. Our cross-lingual transfer learning approach explores the benefits of using a Head-First Fine-Tuning method (HeFiT) that first updates only the regression head parameters and then also updates the pre-trained transformer encoder parameters at a reduced learning rate. Additionally, we study the impact of using a small set of automatically generated examples (in our case, from ChatGPT) for low-resource settings where no human-labeled data is available. Our study shows that HeFiT stabilizes training and consistently improves results for pre-trained models that lack domain adaptation to tweets. Our study also shows a noticeable performance increase in cross-lingual learning when synthetic data is used, confirming the usefulness of current text generation systems to improve zero-shot baseline results. Finally, we examine how possible inconsistencies in the annotated data contribute to cross-lingual interference issues.

Elias Schuhmacher, Andrianos Michail, Juri Opitz et al.

To be discoverable in an embedding-based search process, each part of a document should be reflected in its embedding representation. To quantify any potential reflection biases, we introduce a permutation-based evaluation framework. With this, we observe that state-of-the-art embedding models exhibit systematic positional and language biases when documents are longer and consist of multiple segments. Specifically, early segments and segments in higher-resource languages like English are over-represented, while later segments and segments in lower-resource languages are marginalized. In our further analysis, we find that the positional bias stems from front-loaded attention distributions in pooling-token embeddings, where early tokens receive more attention. To mitigate this issue, we introduce an inference-time attention calibration method that redistributes attention more evenly across document positions, increasing discoverabiltiy of later segments. Our evaluation framework and attention calibration is available at https://github.com/impresso/fair-sentence-transformers

Andrianos Michail, Simon Clematide, Rico Sennrich

The evaluation of cross-lingual semantic search models is often limited to existing datasets from tasks such as information retrieval and semantic textual similarity. We introduce Cross-Lingual Semantic Discrimination (CLSD), a lightweight evaluation task that requires only parallel sentences and a Large Language Model (LLM) to generate adversarial distractors. CLSD measures an embedding model's ability to rank the true parallel sentence above semantically misleading but lexically similar alternatives. As a case study, we construct CLSD datasets for German--French in the news domain. Our experiments show that models fine-tuned for retrieval tasks benefit from pivoting through English, whereas bitext mining models perform best in direct cross-lingual settings. A fine-grained similarity analysis further reveals that embedding models differ in their sensitivity to linguistic perturbations. We release our code and datasets under AGPL-3.0: https://github.com/impresso/cross_lingual_semantic_discrimination

Juri Opitz, Corina Raclé, Emanuela Boros et al.

HIPE-2026 is a CLEF evaluation lab dedicated to person-place relation extraction from noisy, multilingual historical texts. Building on the HIPE-2020 and HIPE-2022 campaigns, it extends the series toward semantic relation extraction by targeting the task of identifying person--place associations in multiple languages and time periods. Systems are asked to classify relations of two types - $at$ ("Has the person ever been at this place?") and $isAt$ ("Is the person located at this place around publication time?") - requiring reasoning over temporal and geographical cues. The lab introduces a three-fold evaluation profile that jointly assesses accuracy, computational efficiency, and domain generalization. By linking relation extraction to large-scale historical data processing, HIPE-2026 aims to support downstream applications in knowledge-graph construction, historical biography reconstruction, and spatial analysis in digital humanities.

Kenneth Enevoldsen, Isaac Chung, Imene Kerboua et al. · cambridge, meta-ai

Text embeddings are typically evaluated on a limited set of tasks, which are constrained by language, domain, and task diversity. To address these limitations and provide a more comprehensive evaluation, we introduce the Massive Multilingual Text Embedding Benchmark (MMTEB) - a large-scale, community-driven expansion of MTEB, covering over 500 quality-controlled evaluation tasks across 250+ languages. MMTEB includes a diverse set of challenging, novel tasks such as instruction following, long-document retrieval, and code retrieval, representing the largest multilingual collection of evaluation tasks for embedding models to date. Using this collection, we develop several highly multilingual benchmarks, which we use to evaluate a representative set of models. We find that while large language models (LLMs) with billions of parameters can achieve state-of-the-art performance on certain language subsets and task categories, the best-performing publicly available model is multilingual-e5-large-instruct with only 560 million parameters. To facilitate accessibility and reduce computational cost, we introduce a novel downsampling method based on inter-task correlation, ensuring a diverse selection while preserving relative model rankings. Furthermore, we optimize tasks such as retrieval by sampling hard negatives, creating smaller but effective splits. These optimizations allow us to introduce benchmarks that drastically reduce computational demands. For instance, our newly introduced zero-shot English benchmark maintains a ranking order similar to the full-scale version but at a fraction of the computational cost.

Andreas Säuberli, Simon Clematide

Reading comprehension tests are used in a variety of applications, reaching from education to assessing the comprehensibility of simplified texts. However, creating such tests manually and ensuring their quality is difficult and time-consuming. In this paper, we explore how large language models (LLMs) can be used to generate and evaluate multiple-choice reading comprehension items. To this end, we compiled a dataset of German reading comprehension items and developed a new protocol for human and automatic evaluation, including a metric we call text informativity, which is based on guessability and answerability. We then used this protocol and the dataset to evaluate the quality of items generated by Llama 2 and GPT-4. Our results suggest that both models are capable of generating items of acceptable quality in a zero-shot setting, but GPT-4 clearly outperforms Llama 2. We also show that LLMs can be used for automatic evaluation by eliciting item reponses from them. In this scenario, evaluation results with GPT-4 were the most similar to human annotators. Overall, zero-shot generation with LLMs is a promising approach for generating and evaluating reading comprehension test items, in particular for languages without large amounts of available data.

Juri Opitz, Lucas Möller, Andrianos Michail et al.

Text embeddings are a fundamental component in many NLP tasks, including classification, regression, clustering, and semantic search. However, despite their ubiquitous application, challenges persist in interpreting embeddings and explaining similarities between them. In this work, we provide a structured overview of methods specializing in inherently interpretable text embeddings and text similarity explanation, an underexplored research area. We characterize the main ideas, approaches, and trade-offs. We compare means of evaluation, discuss overarching lessons learned and finally identify opportunities and open challenges for future research.

Hongji Li, Andrianos Michail, Reto Gubelmann et al.

Controllable and transparent text generation has been a long-standing goal in NLP. Almost as long-standing is a general idea for addressing this challenge: Parsing text to a symbolic representation, and generating from it. However, earlier approaches were hindered by parsing and generation insufficiencies. Using modern parsers and a safety supervision mechanism, we show how close current methods come to this goal. Concretely, we propose the Sentence Smith framework for English, which has three steps: 1. Parsing a sentence into a semantic graph. 2. Applying human-designed semantic manipulation rules. 3. Generating text from the manipulated graph. A final entailment check (4.) verifies the validity of the applied transformation. To demonstrate our framework's utility, we use it to induce hard negative text pairs that challenge text embedding models. Since the controllable generation makes it possible to clearly isolate different types of semantic shifts, we can evaluate text embedding models in a fine-grained way, also addressing an issue in current benchmarking where linguistic phenomena remain opaque. Human validation confirms that our transparent generation process produces texts of good quality. Notably, our way of generation is very resource-efficient, since it relies only on smaller neural networks.

Andrianos Michail, Corina Julia Raclé, Juri Opitz et al.

The growing volume of digitized historical texts requires effective semantic search using text embeddings. However, pre-trained multilingual models face challenges with historical content due to OCR noise and outdated spellings. This study examines multilingual embeddings for cross-lingual semantic search in historical Luxembourgish (LB), a low-resource language. We collect historical Luxembourgish news articles from various periods and use GPT-4o for sentence segmentation and translation, generating 20,000 parallel training sentences per language pair. Additionally, we create a semantic search (Historical LB Bitext Mining) evaluation set and find that existing models perform poorly on cross-lingual search for historical Luxembourgish. Using our historical and additional modern parallel training data, we adapt several multilingual embedding models through contrastive learning or knowledge distillation and increase accuracy significantly for all models. We release our adapted models and historical Luxembourgish-German/French/English bitexts to support further research.

Phillip Benjamin Ströbel, Simon Clematide, Martin Volk et al.

The evaluation of Handwritten Text Recognition (HTR) models during their development is straightforward: because HTR is a supervised problem, the usual data split into training, validation, and test data sets allows the evaluation of models in terms of accuracy or error rates. However, the evaluation process becomes tricky as soon as we switch from development to application. A compilation of a new (and forcibly smaller) ground truth (GT) from a sample of the data that we want to apply the model on and the subsequent evaluation of models thereon only provides hints about the quality of the recognised text, as do confidence scores (if available) the models return. Moreover, if we have several models at hand, we face a model selection problem since we want to obtain the best possible result during the application phase. This calls for GT-free metrics to select the best model, which is why we (re-)introduce and compare different metrics, from simple, lexicon-based to more elaborate ones using standard language models and masked language models (MLM). We show that MLM-based evaluation can compete with lexicon-based methods, with the advantage that large and multilingual transformers are readily available, thus making compiling lexical resources for other metrics superfluous.

Yue Ding, Karolis Martinkus, Damian Pascual et al.

Different studies of the embedding space of transformer models suggest that the distribution of contextual representations is highly anisotropic - the embeddings are distributed in a narrow cone. Meanwhile, static word representations (e.g., Word2Vec or GloVe) have been shown to benefit from isotropic spaces. Therefore, previous work has developed methods to calibrate the embedding space of transformers in order to ensure isotropy. However, a recent study (Cai et al. 2021) shows that the embedding space of transformers is locally isotropic, which suggests that these models are already capable of exploiting the expressive capacity of their embedding space. In this work, we conduct an empirical evaluation of state-of-the-art methods for isotropy calibration on transformers and find that they do not provide consistent improvements across models and tasks. These results support the thesis that, given the local isotropy, transformers do not benefit from additional isotropy calibration.

Raphaël Barman, Maud Ehrmann, Simon Clematide et al.

The massive amounts of digitized historical documents acquired over the last decades naturally lend themselves to automatic processing and exploration. Research work seeking to automatically process facsimiles and extract information thereby are multiplying with, as a first essential step, document layout analysis. If the identification and categorization of segments of interest in document images have seen significant progress over the last years thanks to deep learning techniques, many challenges remain with, among others, the use of finer-grained segmentation typologies and the consideration of complex, heterogeneous documents such as historical newspapers. Besides, most approaches consider visual features only, ignoring textual signal. In this context, we introduce a multimodal approach for the semantic segmentation of historical newspapers that combines visual and textual features. Based on a series of experiments on diachronic Swiss and Luxembourgish newspapers, we investigate, among others, the predictive power of visual and textual features and their capacity to generalize across time and sources. Results show consistent improvement of multimodal models in comparison to a strong visual baseline, as well as better robustness to high material variance.

Peter Makarov, Simon Clematide

We employ imitation learning to train a neural transition-based string transducer for morphological tasks such as inflection generation and lemmatization. Previous approaches to training this type of model either rely on an external character aligner for the production of gold action sequences, which results in a suboptimal model due to the unwarranted dependence on a single gold action sequence despite spurious ambiguity, or require warm starting with an MLE model. Our approach only requires a simple expert policy, eliminating the need for a character aligner or warm start. It also addresses familiar MLE training biases and leads to strong and state-of-the-art performance on several benchmarks.

Peter Makarov, Tatiana Ruzsics, Simon Clematide

This paper presents the submissions by the University of Zurich to the SIGMORPHON 2017 shared task on morphological reinflection. The task is to predict the inflected form given a lemma and a set of morpho-syntactic features. We focus on neural network approaches that can tackle the task in a limited-resource setting. As the transduction of the lemma into the inflected form is dominated by copying over lemma characters, we propose two recurrent neural network architectures with hard monotonic attention that are strong at copying and, yet, substantially different in how they achieve this. The first approach is an encoder-decoder model with a copy mechanism. The second approach is a neural state-transition system over a set of explicit edit actions, including a designated COPY action. We experiment with character alignment and find that naive, greedy alignment consistently produces strong results for some languages. Our best system combination is the overall winner of the SIGMORPHON 2017 Shared Task 1 without external resources. At a setting with 100 training samples, both our approaches, as ensembles of models, outperform the next best competitor.