Malte Ostendorff

Mehdi Ali, Michael Fromm, Klaudia Thellmann et al.

The recent success of Large Language Models (LLMs) has been predominantly driven by curating the training dataset composition, scaling of model architectures and dataset sizes and advancements in pretraining objectives, leaving tokenizer influence as a blind spot. Shedding light on this underexplored area, we conduct a comprehensive study on the influence of tokenizer choice on LLM downstream performance by training 24 mono- and multilingual LLMs at a 2.6B parameter scale, ablating different tokenizer algorithms and parameterizations. Our studies highlight that the tokenizer choice can significantly impact the model's downstream performance and training costs. In particular, we find that the common tokenizer evaluation metrics fertility and parity are not always predictive of model downstream performance, rendering these metrics a questionable proxy for the model's downstream performance. Furthermore, we show that multilingual tokenizers trained on the five most frequent European languages require vocabulary size increases of factor three in comparison to English. While English-centric tokenizers have been applied to the training of multi-lingual LLMs in the past, we find that this approach results in a severe downstream performance degradation and additional training costs of up to 68%, due to an inefficient tokenization vocabulary.

Malte Ostendorff, Till Blume, Terry Ruas et al.

Document embeddings and similarity measures underpin content-based recommender systems, whereby a document is commonly represented as a single generic embedding. However, similarity computed on single vector representations provides only one perspective on document similarity that ignores which aspects make two documents alike. To address this limitation, aspect-based similarity measures have been developed using document segmentation or pairwise multi-class document classification. While segmentation harms the document coherence, the pairwise classification approach scales poorly to large scale corpora. In this paper, we treat aspect-based similarity as a classical vector similarity problem in aspect-specific embedding spaces. We represent a document not as a single generic embedding but as multiple specialized embeddings. Our approach avoids document segmentation and scales linearly w.r.t.the corpus size. In an empirical study, we use the Papers with Code corpus containing 157,606 research papers and consider the task, method, and dataset of the respective research papers as their aspects. We compare and analyze three generic document embeddings, six specialized document embeddings and a pairwise classification baseline in the context of research paper recommendations. As generic document embeddings, we consider FastText, SciBERT, and SPECTER. To compute the specialized document embeddings, we compare three alternative methods inspired by retrofitting, fine-tuning, and Siamese networks. In our experiments, Siamese SciBERT achieved the highest scores. Additional analyses indicate an implicit bias of the generic document embeddings towards the dataset aspect and against the method aspect of each research paper. Our approach of aspect-based document embeddings mitigates potential risks arising from implicit biases by making them explicit.

Luca Foppiano, Sotaro Takeshita, Pedro Ortiz Suarez et al.

SciLaD is a novel, large-scale dataset of scientific language constructed entirely using open-source frameworks and publicly available data sources. It comprises a curated English split containing over 10 million scientific publications and a multilingual, unfiltered TEI XML split including more than 35 million publications. We also publish the extensible pipeline for generating SciLaD. The dataset construction and processing workflow demonstrates how open-source tools can enable large-scale, scientific data curation while maintaining high data quality. Finally, we pre-train a RoBERTa model on our dataset and evaluate it across a comprehensive set of benchmarks, achieving performance comparable to other scientific language models of similar size, validating the quality and utility of SciLaD. We publish the dataset and evaluation pipeline to promote reproducibility, transparency, and further research in natural scientific language processing and understanding including scholarly document processing.

Tim Schopf, Emanuel Gerber, Malte Ostendorff et al.

Generic sentence embeddings provide a coarse-grained approximation of semantic textual similarity but ignore specific aspects that make texts similar. Conversely, aspect-based sentence embeddings provide similarities between texts based on certain predefined aspects. Thus, similarity predictions of texts are more targeted to specific requirements and more easily explainable. In this paper, we present AspectCSE, an approach for aspect-based contrastive learning of sentence embeddings. Results indicate that AspectCSE achieves an average improvement of 3.97% on information retrieval tasks across multiple aspects compared to the previous best results. We also propose using Wikidata knowledge graph properties to train models of multi-aspect sentence embeddings in which multiple specific aspects are simultaneously considered during similarity predictions. We demonstrate that multi-aspect embeddings outperform single-aspect embeddings on aspect-specific information retrieval tasks. Finally, we examine the aspect-based sentence embedding space and demonstrate that embeddings of semantically similar aspect labels are often close, even without explicit similarity training between different aspect labels.

Mehdi Ali, Michael Fromm, Klaudia Thellmann et al.

We present two multilingual LLMs, Teuken 7B-base and Teuken 7B-instruct, designed to embrace Europe's linguistic diversity by supporting all 24 official languages of the European Union. Trained on a dataset comprising around 60% non-English data and utilizing a custom multilingual tokenizer, our models address the limitations of existing LLMs that predominantly focus on English or a few high-resource languages. We detail the models' development principles, i.e., data composition, tokenizer optimization, and training methodologies. The models demonstrate strong performance across multilingual benchmarks, as evidenced by their performance on European versions of ARC, HellaSwag, and TruthfulQA.

Qian Ruan, Malte Ostendorff, Georg Rehm

Transformer-based language models usually treat texts as linear sequences. However, most texts also have an inherent hierarchical structure, i.e., parts of a text can be identified using their position in this hierarchy. In addition, section titles usually indicate the common topic of their respective sentences. We propose a novel approach to formulate, extract, encode and inject hierarchical structure information explicitly into an extractive summarization model based on a pre-trained, encoder-only Transformer language model (HiStruct+ model), which improves SOTA ROUGEs for extractive summarization on PubMed and arXiv substantially. Using various experimental settings on three datasets (i.e., CNN/DailyMail, PubMed and arXiv), our HiStruct+ model outperforms a strong baseline collectively, which differs from our model only in that the hierarchical structure information is not injected. It is also observed that the more conspicuous hierarchical structure the dataset has, the larger improvements our method gains. The ablation study demonstrates that the hierarchical position information is the main contributor to our model's SOTA performance.

Malte Ostendorff, Georg Rehm

Most Transformer language models are primarily pretrained on English text, limiting their use for other languages. As the model sizes grow, the performance gap between English and other languages with fewer compute and data resources increases even further. Consequently, more resource-efficient training methods are needed to bridge the gap for languages with fewer resources available. To address this problem, we introduce a cross-lingual and progressive transfer learning approach, called CLP-Transfer, that transfers models from a source language, for which pretrained models are publicly available, like English, to a new target language. As opposed to prior work, which focused on the cross-lingual transfer between two languages, we extend the transfer to the model size. Given a pretrained model in a source language, we aim for a same-sized model in a target language. Instead of training a model from scratch, we exploit a smaller model that is in the target language but requires much fewer resources. Both small and source models are then used to initialize the token embeddings of the larger model based on the overlapping vocabulary of the source and target language. All remaining weights are reused from the model in the source language. This approach outperforms the sole cross-lingual transfer and can save up to 80% of the training steps compared to the random initialization.

Malte Ostendorff, Elliott Ash, Terry Ruas et al.

Recommender systems assist legal professionals in finding relevant literature for supporting their case. Despite its importance for the profession, legal applications do not reflect the latest advances in recommender systems and representation learning research. Simultaneously, legal recommender systems are typically evaluated in small-scale user study without any public available benchmark datasets. Thus, these studies have limited reproducibility. To address the gap between research and practice, we explore a set of state-of-the-art document representation methods for the task of retrieving semantically related US case law. We evaluate text-based (e.g., fastText, Transformers), citation-based (e.g., DeepWalk, Poincaré), and hybrid methods. We compare in total 27 methods using two silver standards with annotations for 2,964 documents. The silver standards are newly created from Open Case Book and Wikisource and can be reused under an open license facilitating reproducibility. Our experiments show that document representations from averaged fastText word vectors (trained on legal corpora) yield the best results, closely followed by Poincaré citation embeddings. Combining fastText and Poincaré in a hybrid manner further improves the overall result. Besides the overall performance, we analyze the methods depending on document length, citation count, and the coverage of their recommendations. We make our source code, models, and datasets publicly available at https://github.com/malteos/legal-document-similarity/.

Malte Ostendorff, Till Blume, Saskia Ostendorff

Recent advances in the area of legal information systems have led to a variety of applications that promise support in processing and accessing legal documents. Unfortunately, these applications have various limitations, e.g., regarding scope or extensibility. Furthermore, we do not observe a trend towards open access in digital libraries in the legal domain as we observe in other domains, e.g., economics of computer science. To improve open access in the legal domain, we present our approach for an open source platform to transparently process and access Legal Open Data. This enables the sustainable development of legal applications by offering a single technology stack. Moreover, the approach facilitates the development and deployment of new technologies. As proof of concept, we implemented six technologies and generated metadata for more than 250,000 German laws and court decisions. Thus, we can provide users of our platform not only access to legal documents, but also the contained information.

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.

Nicolo' Brandizzi, Hammam Abdelwahab, Anirban Bhowmick et al.

This paper presents a comprehensive overview of the data preparation pipeline developed for the OpenGPT-X project, a large-scale initiative aimed at creating open and high-performance multilingual large language models (LLMs). The project goal is to deliver models that cover all major European languages, with a particular focus on real-world applications within the European Union. We explain all data processing steps, starting with the data selection and requirement definition to the preparation of the final filtered data. We distinguish between curated data and web data, as each of these categories is handled by distinct pipelines, with curated data undergoing minimal filtering and web data requiring extensive filtering and deduplication. This distinction guided the development of specialized algorithmic solutions for both pipelines. In addition to describing the processing methodologies, we provide an in-depth analysis of the datasets, increasing transparency and alignment with European data regulations. Finally, we share key insights and challenges faced during the project, offering recommendations for future endeavors in large-scale multilingual data preparation for LLMs.

Orhun Mersin Caglidil, Malte Ostendorff, Georg Rehm

Language models are trained mostly on Web data, which often contains social stereotypes and biases that the models can inherit. This has potentially negative consequences, as models can amplify these biases in downstream tasks or applications. However, prior research has primarily focused on the English language, especially in the context of gender bias. In particular, grammatically gender-neutral languages such as Turkish are underexplored despite representing different linguistic properties to language models with possibly different effects on biases. In this paper, we fill this research gap and investigate the significance of gender bias in Turkish language models. We build upon existing bias evaluation frameworks and extend them to the Turkish language by translating existing English tests and creating new ones designed to measure gender bias in the context of Türkiye. Specifically, we also evaluate Turkish language models for their embedded ethnic bias toward Kurdish people. Based on the experimental results, we attribute possible biases to different model characteristics such as the model size, their multilingualism, and the training corpora. We make the Turkish gender bias dataset publicly available.

Pedro Ortiz Suarez, Laurie Burchell, Catherine Arnett et al.

Language identification (LID) is a fundamental step in curating multilingual corpora. However, LID models still perform poorly for many languages, especially on the noisy and heterogeneous web data often used to train multilingual language models. In this paper, we introduce CommonLID, a community-driven, human-annotated LID benchmark for the web domain, covering 109 languages. Many of the included languages have been previously under-served, making CommonLID a key resource for developing more representative high-quality text corpora. We show CommonLID's value by using it, alongside five other common evaluation sets, to test eight popular LID models. We analyse our results to situate our contribution and to provide an overview of the state of the art. In particular, we highlight that existing evaluations overestimate LID accuracy for many languages in the web domain. We make CommonLID and the code used to create it available under an open, permissive license.

Ekaterina Borisova, Fabio Barth, Nils Feldhus et al.

Tables are among the most widely used tools for representing structured data in research, business, medicine, and education. Although LLMs demonstrate strong performance in downstream tasks, their efficiency in processing tabular data remains underexplored. In this paper, we investigate the effectiveness of both text-based and multimodal LLMs on table understanding tasks through a cross-domain and cross-modality evaluation. Specifically, we compare their performance on tables from scientific vs. non-scientific contexts and examine their robustness on tables represented as images vs. text. Additionally, we conduct an interpretability analysis to measure context usage and input relevance. We also introduce the TableEval benchmark, comprising 3017 tables from scholarly publications, Wikipedia, and financial reports, where each table is provided in five different formats: Image, Dictionary, HTML, XML, and LaTeX. Our findings indicate that while LLMs maintain robustness across table modalities, they face significant challenges when processing scientific tables.

Ben Hauptvogel, Malte Ostendorff, Georg Rehm et al.

Recent advancements in large language models (LLMs) have led to their increased application across various tasks, with reinforcement learning from human feedback (RLHF) being a crucial part of their training to align responses with user intentions. In the RLHF process, a reward model is trained using responses preferences determined by human labelers or AI systems, which then refines the LLM through reinforcement learning. This work introduces weak supervision as a strategy to extend RLHF datasets and enhance reward model performance. Weak supervision employs noisy or imprecise data labeling, reducing reliance on expensive manually labeled data. By analyzing RLHF datasets to identify heuristics that correlate with response preference, we wrote simple labeling functions and then calibrated a label model to weakly annotate unlabeled data. Our evaluation show that while weak supervision significantly benefits smaller datasets by improving reward model performance, its effectiveness decreases with larger, originally labeled datasets. Additionally, using an LLM to generate and then weakly label responses offers a promising method for extending preference data.

Martin Courtois, Malte Ostendorff, Leonhard Hennig et al.

Initially introduced as a machine translation model, the Transformer architecture has now become the foundation for modern deep learning architecture, with applications in a wide range of fields, from computer vision to natural language processing. Nowadays, to tackle increasingly more complex tasks, Transformer-based models are stretched to enormous sizes, requiring increasingly larger training datasets, and unsustainable amount of compute resources. The ubiquitous nature of the Transformer and its core component, the attention mechanism, are thus prime targets for efficiency research. In this work, we propose an alternative compatibility function for the self-attention mechanism introduced by the Transformer architecture. This compatibility function exploits an overlap in the learned representation of the traditional scaled dot-product attention, leading to a symmetric with pairwise coefficient dot-product attention. When applied to the pre-training of BERT-like models, this new symmetric attention mechanism reaches a score of 79.36 on the GLUE benchmark against 78.74 for the traditional implementation, leads to a reduction of 6% in the number of trainable parameters, and reduces the number of training steps required before convergence by half.

Malte Ostendorff, Nils Rethmeier, Isabelle Augenstein et al.

Learning scientific document representations can be substantially improved through contrastive learning objectives, where the challenge lies in creating positive and negative training samples that encode the desired similarity semantics. Prior work relies on discrete citation relations to generate contrast samples. However, discrete citations enforce a hard cut-off to similarity. This is counter-intuitive to similarity-based learning, and ignores that scientific papers can be very similar despite lacking a direct citation - a core problem of finding related research. Instead, we use controlled nearest neighbor sampling over citation graph embeddings for contrastive learning. This control allows us to learn continuous similarity, to sample hard-to-learn negatives and positives, and also to avoid collisions between negative and positive samples by controlling the sampling margin between them. The resulting method SciNCL outperforms the state-of-the-art on the SciDocs benchmark. Furthermore, we demonstrate that it can train (or tune) models sample-efficiently, and that it can be combined with recent training-efficient methods. Perhaps surprisingly, even training a general-domain language model this way outperforms baselines pretrained in-domain.

Malte Ostendorff, Corinna Breitinger, Bela Gipp

Literature recommendation systems (LRS) assist readers in the discovery of relevant content from the overwhelming amount of literature available. Despite the widespread adoption of LRS, there is a lack of research on the user-perceived recommendation characteristics for fundamentally different approaches to content-based literature recommendation. To complement existing quantitative studies on literature recommendation, we present qualitative study results that report on users' perceptions for two contrasting recommendation classes: (1) link-based recommendation represented by the Co-Citation Proximity (CPA) approach, and (2) text-based recommendation represented by Lucene's MoreLikeThis (MLT) algorithm. The empirical data analyzed in our study with twenty users and a diverse set of 40 Wikipedia articles indicate a noticeable difference between text- and link-based recommendation generation approaches along several key dimensions. The text-based MLT method receives higher satisfaction ratings in terms of user-perceived similarity of recommended articles. In contrast, the CPA approach receives higher satisfaction scores in terms of diversity and serendipity of recommendations. We conclude that users of literature recommendation systems can benefit most from hybrid approaches that combine both link- and text-based approaches, where the user's information needs and preferences should control the weighting for the approaches used. The optimal weighting of multiple approaches used in a hybrid recommendation system is highly dependent on a user's shifting needs.

Malte Ostendorff, Terry Ruas, Till Blume et al.

Traditional document similarity measures provide a coarse-grained distinction between similar and dissimilar documents. Typically, they do not consider in what aspects two documents are similar. This limits the granularity of applications like recommender systems that rely on document similarity. In this paper, we extend similarity with aspect information by performing a pairwise document classification task. We evaluate our aspect-based document similarity for research papers. Paper citations indicate the aspect-based similarity, i.e., the section title in which a citation occurs acts as a label for the pair of citing and cited paper. We apply a series of Transformer models such as RoBERTa, ELECTRA, XLNet, and BERT variations and compare them to an LSTM baseline. We perform our experiments on two newly constructed datasets of 172,073 research paper pairs from the ACL Anthology and CORD-19 corpus. Our results show SciBERT as the best performing system. A qualitative examination validates our quantitative results. Our findings motivate future research of aspect-based document similarity and the development of a recommender system based on the evaluated techniques. We make our datasets, code, and trained models publicly available.

Malte Ostendorff

To cope with the ever-growing information overload, an increasing number of digital libraries employ content-based recommender systems. These systems traditionally recommend related documents with the help of similarity measures. However, current document similarity measures simply distinguish between similar and dissimilar documents. This simplification is especially crucial for extensive documents, which cover various facets of a topic and are often found in digital libraries. Still, these similarity measures neglect to what facet the similarity relates. Therefore, the context of the similarity remains ill-defined. In this doctoral thesis, we explore contextual document similarity measures, i.e., methods that determine document similarity as a triple of two documents and the context of their similarity. The context is here a further specification of the similarity. For example, in the scientific domain, research papers can be similar with respect to their background, methodology, or findings. The measurement of similarity in regards to one or more given contexts will enhance recommender systems. Namely, users will be able to explore document collections by formulating queries in terms of documents and their contextual similarities. Thus, our research objective is the development and evaluation of a recommender system based on contextual similarity. The underlying techniques will apply established similarity measures and as well as neural approaches while utilizing semantic features obtained from links between documents and their text.

Georg Rehm, Peter Bourgonje, Stefanie Hegele et al.

In all domains and sectors, the demand for intelligent systems to support the processing and generation of digital content is rapidly increasing. The availability of vast amounts of content and the pressure to publish new content quickly and in rapid succession requires faster, more efficient and smarter processing and generation methods. With a consortium of ten partners from research and industry and a broad range of expertise in AI, Machine Learning and Language Technologies, the QURATOR project, funded by the German Federal Ministry of Education and Research, develops a sustainable and innovative technology platform that provides services to support knowledge workers in various industries to address the challenges they face when curating digital content. The project's vision and ambition is to establish an ecosystem for content curation technologies that significantly pushes the current state of the art and transforms its region, the metropolitan area Berlin-Brandenburg, into a global centre of excellence for curation technologies.

Georg Rehm, Karolina Zaczynska, Julián Moreno-Schneider et al.

Previous work of ours on Semantic Storytelling uses text analytics procedures including Named Entity Recognition and Event Detection. In this paper, we outline our longer-term vision on Semantic Storytelling and describe the current conceptual and technical approach. In the project that drives our research we develop AI-based technologies that are verified by partners from industry. One long-term goal is the development of an approach for Semantic Storytelling that has broad coverage and that is, furthermore, robust. We provide first results on experiments that involve discourse parsing, applied to a concrete use case, "Explore the Neighbourhood!", which is based on a semi-automatically collected data set with documents about noteworthy people in one of Berlin's districts. Though automatically obtaining annotations for coherence relations from plain text is a non-trivial challenge, our preliminary results are promising. We envision our approach to be combined with additional features (NER, coreference resolution, knowledge graphs

Sarah Schulz, Jurica Ševa, Samuel Rodriguez et al.

We present a new corpus comprising annotations of medical entities in case reports, originating from PubMed Central's open access library. In the case reports, we annotate cases, conditions, findings, factors and negation modifiers. Moreover, where applicable, we annotate relations between these entities. As such, this is the first corpus of this kind made available to the scientific community in English. It enables the initial investigation of automatic information extraction from case reports through tasks like Named Entity Recognition, Relation Extraction and (sentence/paragraph) relevance detection. Additionally, we present four strong baseline systems for the detection of medical entities made available through the annotated dataset.

Malte Ostendorff, Terry Ruas, Moritz Schubotz et al.

Many digital libraries recommend literature to their users considering the similarity between a query document and their repository. However, they often fail to distinguish what is the relationship that makes two documents alike. In this paper, we model the problem of finding the relationship between two documents as a pairwise document classification task. To find the semantic relation between documents, we apply a series of techniques, such as GloVe, Paragraph-Vectors, BERT, and XLNet under different configurations (e.g., sequence length, vector concatenation scheme), including a Siamese architecture for the Transformer-based systems. We perform our experiments on a newly proposed dataset of 32,168 Wikipedia article pairs and Wikidata properties that define the semantic document relations. Our results show vanilla BERT as the best performing system with an F1-score of 0.93, which we manually examine to better understand its applicability to other domains. Our findings suggest that classifying semantic relations between documents is a solvable task and motivates the development of recommender systems based on the evaluated techniques. The discussions in this paper serve as first steps in the exploration of documents through SPARQL-like queries such that one could find documents that are similar in one aspect but dissimilar in another.

Malte Ostendorff, Peter Bourgonje, Maria Berger et al.

In this paper, we focus on the classification of books using short descriptive texts (cover blurbs) and additional metadata. Building upon BERT, a deep neural language model, we demonstrate how to combine text representations with metadata and knowledge graph embeddings, which encode author information. Compared to the standard BERT approach we achieve considerably better results for the classification task. For a more coarse-grained classification using eight labels we achieve an F1- score of 87.20, while a detailed classification using 343 labels yields an F1-score of 64.70. We make the source code and trained models of our experiments publicly available