This paper presents a set of approaches to tackle the named entity recognition and relation extraction from scientific literature, specifically targeting the gut-brain axis related terms and relationships between them. The proposed methods participated in the GutBrainIE Task at CLEF 2025 BioASQ Lab. The solutions rely on fine-tuned BERT-based models (BioBERT , BiomedNLP ELECTRA , BioBERT PubMed) and GLiNER for the named entity recognition task, ATLOP and REBEL fine-tuning for the relation extraction task. Hybrid models and ensemble of models are also demonstrated for end-to-end tasks. Notably, one of our proposed solutions ranked 2nd on the most dificult task of the challenge - Ternary Mention-based Relation Extraction, achieving micro-F1 37.29%. Our best system for Named Entity Recognition over the test set achieved micro-F1 80.1%. On the Binary Tag-based Relation Extraction subtask, our best solution achieved micro-F1 65.38% on the test set and on the Ternary Tag-based Relation Extraction subtask, our best result was micro-F1 63.72%. All of the proposed approaches demonstrated good performance, consistently outperforming baseline results across all subtasks of the GutBrainIE Task at CLEF 2025 BioASQ Lab.
An increasing amount of research indicates that there exists a complex interaction between the gut
and the brain [
One step in this direction is to use the wealth of scientific publications in the biomedical field available in PubMed. PubMed contains an extensive collection of peer-reviewed articles from leading scientific journals and conferences in the field of biomedicine and ofers a rich resource for systematically identifying, extracting and synthesizing new scientific insights. However, the rapid pace of publication of new articles and the richness of the resource make the manual exploration of current results challenging, therefore it is important to provide tools for automated tracking of new developments to speed up the research in the area.
By applying modern natural language processing (NLP) techniques to scientific literature, researchers can eficiently collect and analyze evidences, reveal hidden patterns, explore comorbidities and risk factors, and accelerate the discovery of new connections within the gut-brain axis, which may ultimately help to better understanding, prevention and treatment of related diseases.
In this paper, we present our research in developing deep learning and hybrid models for NLP,
for automated information extraction from PubMed articles. The results we present are part of the
GutBrainIE Task at CLEF 2025 BioASQ Lab [
The proposed solutions below are based on deep learning models including GLiNER [
The paper is organized as follows: Section 2 outlines the related work, state-of-the art models and achievements on the tasks; Section 3 defines the tasks and data; in Section 4 are listed the approaches applied for NER; in Section 5 are presented the results of NER; Section 6 discusses the approaches to RE, Section 7 - the results of RE and in Section 8 are provided discussion and conclusions of the study.
State-of-the-art NLP approaches for information extraction from Biomedical literature use range of
techniques from deep learning (BERT, BiLSTM-CRF, XML-R and LLMs), classical ML methods (CRF,
SVM) and even traditional solutions like dictionary- and rule-based ones. The best F1 scores achieved
for the NER task on PubMed articles are up to 89.3% and for RE the performance vary in the range
47.7-88.5%, depending on the types of relations and entities [
One of the famous approaches to the problem is to use the innovative method GLiNER [
PubMed is one of the largest databases for biomedical literature, comprising more than 38 million
citations. The documents provided by the challenge organizers consist of a title, abstract, author
and journal metadata of a publication, retrieved from PubMed with explicit focus on the gut-brain
interplay and its implications in neurological and mental health. To foster the development of efective
Information Extraction (IE) systems within the context of the the GutBrainIE Task at CLEF 2025 BioASQ
Lab1 [
The goal of this task is to classify a text span (entity mention) into one of 13 pre-defined categories/labels (see the X-axis of Figure 1). We will use only the term "category" in the next sections as the term 1https://hereditary.dei.unipd.it/challenges/gutbrainie/2025/ "label" used by the organizers can be ambiguous in terms of data description (meaning (1) alternative name of the entity and also (2) entity category in the context of the given task). However, some of the categories have only very few annotations. This poses a potential challenge for larger pretrained models which need a substantial amount of data and resources to generalize well. For example, Figure 1 clearly shows the distribution between labels in the Platinum-standard collection. This plot remains almost unchanged for any other annotated dataset. The dominating labels are Disorder-Disease-Finding (DDF), chemical, bacteria, human, and microbiome. Categories like food, gene, drug, and statistical technique are rarely found in the texts.
1,232 1,200 1,000 cy 800 n e u req 600 F 400 200
0 { } {
F DD
ical chem bacteria
an hum
e icrobiom m
t en dietarysupplem
edical technique biom
ical location anatom statistdicrualgtechnique 3.2. Task 6.2 - RE The second main task is on RE and consists of three subtasks with increasing dificulty. The first of them is Binary Tag-based Relation Extraction (BT-RE) - requiring to identify whether two entity categories are in relation. Given the input text and/or the found entities, a binary relation is defined as an ordered tuple in the following example format: Where DDF and human are both named entity categories. It is important to note that the least represented categories are usually those involved in relations with only a small set of other categories - objects. For instance, gene only participates in relations with three other categories, in contrast to human or microbiome which appear as subject in relations to more than twice other labels.
In the next subtask - Ternary Tag-based Relation Extraction (TT-RE) the relation type (predicate)
between the entity categories is also considered. An example of a ternary tag-based relation is:
include target, impact, influence , change efect , located in, is linked to, and others. All relation types
are asymmetric, meaning that (subject, predicate, object) != (object, predicate, subject). On average,
influence , target, located in are some of the most common predicates. It is worth noting that the
Silverstandard collection has the largest number of annotated relations - 10 616, which is 5 times more than
the relations annotated in the Gold collection - 1 994. Even though the number of documents is diferent,
the average number of found relations per document is 21.27 for Silver, versus 9.59 for Gold. The highest
quality collection - Platinum, has total of 1 455 relations and an average of 13.11 relations per document.
For more detailed description on the statistics for the diferent relation types in the data, refer to the
extended overview paper [
Finally, the last subtask on Ternary Mention-based Relation Extraction (TM-RE) aims to extract every text mention of the entities with their corresponding relation. The annotated training data has the following format: { } "subject_text_span": "IgA-Biome", "subject_label": "microbiome", "predicate": "located in", "object_text_span": "AR and TD patients", "object_label": "human"
Across all subtasks, a true positive is only considered when there is a full match of the found instance. Therefore, the possibility of partial matches and potentially accounting for some of the systems errors is eliminated.
As a potential solution to the problem of scarcity in a number of important categories, as observed, PubMed API is used to fetch more articles and augment the training data. As a matching criterion, only articles in the period (2015/01/01 - 2025/01/01) are considered, and a specially designed search query is
created to extract all articles, where the gut-brain axis or the gastrointestinal microbiome are defined as major topics, following the MeSH ontology.
To fully ensure that the newly extracted articles are related to the task requirements, the search query is extended to also account for specific labels relevant to the needed categories. This way, for the rare category gene the query extracts all relevant articles within the gut-brain axis that have their main topics defined to include genes, DNA, or genetics. The designed search query looks like this: ((Brain-Gut Axis[MeSH Major Topic]) OR (Gastrointestinal Microbiome[MeSH Major Topic])) AND (genes[MeSH Major Topic] OR DNA[MeSH Major Topic] OR genetics[MeSH Subheading])
Following the strategy shown on Figure 2 for all categories by changing the query and annotating the entities by distant supervision, usually with diferent combinations of systems, resulted in the creation of our own Bronze-standard annotated corpus with a total of 6728 articles.
For the NER task several diferent techniques are presented. Two diferent pretrained language models (LMs) - BERT [17] and ELECTRA [18] are considered. BERT-based models are pretrained on objectives such as Masked Language Modeling (MLM) and Next Sentence Prediction (NSP), in contrast to what ELECTRA-based models are trying to learn. This family of models uses a small generator network to corrupt the input, similar to BERT, by replacing some tokens with plausible alternatives and then turns to train a discriminator, learning whether each token was in the original input or was artificially generated. The results of changing the learning objective lead to better contextual representations and outperform BERT on the GLUE natural language understanding benchmark. Furthermore, the learned embeddings are especially more representative for small models, which are actually more suitable for the context of the gut-brain task. To leverage the capabilities of more compute eficient and smaller bidirectional LMs, fine-tuned GLiNER are also included in our experiments. GLiNER can extract any entity type by maximizing the probability of a span (i,j) to be of the correct entity type t and minimizing it, for the same span (i,j) to be of any other type t. Finally, our highest-precision and highest-recall system is a combination of several models - ensemble. When using multiple models, the final decision is based on pre-defined rules or majority voting. For example - if model X performs best for category statistical technique, then on test time we will take the predictions for this category, if any, only from model X. To fine-tune models from the BERT/ELECTRA family, standard token classification with Hugging Face is employed. The first step is to adjust the training data or in other words to "tag" the input in the widely adopted BIO-format, first introduced by Ramshaw and Marcus [ 19]. However, the training data is not always the same. This means, we experiment with diferent parts of the whole dataset. In some configurations, the models are only given the Platinum and Gold Standard collections. In others, they are fine-tuned on all collections including the expanded Bronze corpus created by us. To convert the provided format to BIO, ScispaCy [20] is used, a python package that provides spaCy 2 models for scientific text preprocessing. Due to its size and complexity, en_core_sci_sm model was chosen. The inference happens by creating a pipeline with the saved model and making the predictions on entity level, where the input is no longer a token, but a whole chunk of text. To choose the most successful tokenizer setup was selected through experiments with diferent maximum sequence lengths starting from 128 and going up to 512. For simplicity, no additional tricks are considred to expand this window. Most of the documents do not exceed 400 tokens, including the subwords, as shown in Figure 3. Experiments with larger maximum lengths like the one in kiddothe2b/biomedical-longformer-large do not show better performance for the data.
120 ts 100 n e um 80 c o d o 60 f r e b um 40 N 20 0
121 63 19 <=200 <=400 <=600 <=800
3
2 >800
To arrive at this conclusion, the data is tokenized with biobert-base-cased-v1.1. Although during ifne-tuning diferent tokenizers are used with the diferent models, almost all of them use the WordPiece tokenization algorithm. The models in these experiments are domain-specific, pretrained on large biomedical corpora, usually consisting of PubMed abstracts and/or PubMed Central (PMC) full texts.
Here is the list of our most successful models on the development set. Their results on the oficial test dataset are shown in the next section 5. The exact choice on the hyperparemeters of the submitted models is illustrated in Table 7.
• dmis-lab/biobert-base-cased-v1.1 [
This is a lightweight zero-shot NER model based on the GLiNER architecture. It is fine-tuned on all data excluding the Bronze collection. We refer to it as GLiNER. The model is not achieving the best performance on the dev set, but has the highest micro precision and recall, therefore F1-micro score on the test set.
Leveraging the power of a subset of systems, the most successful ensemble approach on the test set turns out to be, this time matching the expectations, with the best models on the development set. In particular, the highest precision and recall system includes monologg/biobert_v1.1_pubmed, pubmed-electra-base, GLiNER, monologg/biobert_v1.1_pubmed and another version of pubmed-electra-base fine-tuned on all data and the expanded Bronze corpus, but by changing the annotator model. It is referred in Table 3 as microsoft/BiomedNLP-BiomedELECTRA-base-uncased (v2). In this case, the additional data is annotated by pubmed-electra-base and GLiNER. In total, 5 diferent models are used. The strategy to form this ensemble is two-fold. First, error analysis on the development set shows per-label performance for each model. Then, on the basis of this, hand-crafted rules are prepared to take into account only the predictions for the selected categories during testing. For example, GLiNER is used only for gene and chemical. The whole model composition is shown on Table 3.
Categories bacteria statistical technique gene, chemical human anatomical location, animal, bacteria, biomedical technique, DDF, dietary supplement, drug, food, microbiome Table 4 lists the results of our models on both the development and test sets for micro precision. The second column shows which subset of the data is used for fine-tuning to achieve the results, including the model split for the ensemble. The ensemble approach discussed in the previous section, Subsection 4.2, improves on precision by nearly 10% compared to GLiNER on the development set. However, on the test set GLiNER is the only model out of all to perform better.
Similarly to the precision table, Table 5 shows once again the better performance of our best models on the development set, this time by micro recall. Although some of the other fine-tuned versions of pubmed-electra-base and biobert-base-v1.1 are not too far behind, the ensemble we form turns out to account for the biggest number of categories according to the authors annotations. In contrast, on the test set, the micro recall of all systems drops. The biggest drop is in pubmed-electra-base, while the most consistent model remains GLiNER.
Finally, Table 6 shows the F1-micro score. This metric is used to rank the systems in the oficial leaderboard3. GLiNER has the lowest score on the development set, because of its weak precision compared to the other systems. On the test set, all of the systems show decreasing performance, except for GLiNER. It is the only model to generalize well and improve on its scores, making it our most successful submission.
Before the final submission, progressive fine-tuning was applied to all of the selected models. Surprisingly, after including the development set and training for a small number of epochs, usually 3-5, the performance decreases on average by around 5% on F1-micro, including the ensemble using these models. In conclusion, lighter models indeed prove to be more beneficial within the context of this task and GLiNER is especially good for NER with limited amount of data. 3https://hereditary.dei.unipd.it/challenges/gutbrainie/2025/#six Dev R The error analysis on the dev set, shows that that only 15% of the false positives are not entities of interest. Obviously the model is good enough in guessing the mentions in the text but not so good in guessing their type. Also the model does not learn very well the annotation guidelines in terms of nested entities - e.g. it extracts gene mentions such as "16 S rDNA gene" which are part of a longer mention "amplification and sequencing of the V4 region of the 16 S rDNA gene" which is a biomedical technique in the gold standard. There are also cases of extracting entities which are true positives in our view but are missing in the gold standard such as "discriminant taxa analyses". Although NER is a classical NLP task with long history and significant progress, this problem remains challenging in its detail.
One of the applied strategies for RE is to use an autoregressive transformer-based approach to extract the
entities in relation within a document. For this purpose, Babelscape/rebel-large is employed. REBEL [
At this stage, the goal is to generate the target linearized triplets as accurately as possible from the input text.
What makes REBEL a good choice are its pretraining methods. The authors of the algorithm solved the problem with the scarce RE datasets by creating a new one. This so-called silver corpus consists of Wikipedia abstracts where Wikipedia hyperlinks are matched with WikiData entities. From these, all the present relations are extracted. In total, REBEL is able to find 220 relation types based on the data it has been trained on.
Here, REBEL is fine-tuned to solve all RE-related subtasks. Because of the nature of the model, first all entity mentions and their corresponding relations are extracted. To make the decoding fit our criteria for the other subtasks, the target is modified to also include the entity types along with the mentions. Therefore, the training linear triplets are in the following format: <triplet> head-entity-mention %head-entity-type% <subj> type-relation <rel> tail-entity-mention %tail-entity-type%
As a last postprocessing step, also aiming to improve precision, the pre-defined by the organisers rules
for domain and range of the predicates are applied. Therefore, only relations that are valid according to
the annotation guidelines [
To obtain the most successful version of REBEL, all annotated collections (excluding the development set) are used during fine-tuning. The most eficient combination of hyperparameters is with a batch size of 16 and a learning rate of 5e-5. Furthermore, the maximum length of the generated output during decoding is limited to 512 tokens with the number of beams for beam search set to 5. The model is trained for just 100 iterations on a single RTX 6000 ADA Generation GPU, suggesting strong potential for the encoder-decoder architecture in solving RE-related tasks. The training time was about 2 hours. During it, the model with the best F1-micro score on the development set was selected. Similarly, other experiments with diferent combinations of input data and hyperparameters showed decreasing performance.
The most successful approach for relation extraction proved to be the use of Adaptive Thresholding
and Localized Context Pooling (ATLOP) [
ATLOP introduces the "Adaptive Thresholding" (AT) mechanism to address the limitations of global thresholding in the multi-label scenario. This technique replaces the static global threshold with a learnable, entity-dependent threshold, aiming to reduce decision errors during inference. The core of Adaptive Thresholding is the introduction of a special "threshold class," denoted as TH. This TH class is treated and learned similarly to other actual relation classes within the model’s architecture. Its purpose is to act as a learned decision boundary. The main benefit of AT is that, instead of using the usual approach of optimizing a value in the range (0, 1) and picking the one that maximizes the evaluation metrics, AT provides a way to learn the optimal threshold during training. Those techniques, plus the use of "Localized Context Pooling" and some other innovations, make ATLOP an eficient and easy-to-use.
To try to improve the model’s performance, diferent pretrained models are used, such as: BERT [ 17],
RoBERTa [23], XLM-R [24], PubMedBERT [
7.1. Subtask 6.2.1 - BT-RE For the BT-RE subtask, results on the dev set are shown in Table 9, demonstrating that ATLOP is a very efective method for relation extraction. While REBEL is a good end-to-end alternative, performance is slightly worse than ATLOP models, which suggests that there is a strong potential for encoder-decoder or decoder-only models for relation extraction. Another observation is that task pretraining gives improvement for BiomedELECTRA, but it isn’t very significant. Here, all ATLOP models use NER predictions from the baseline GLiNER model.
The results on the test set are shown in Table 10. While XLM-R and REBEL do not fall too much behind BiomedELECTRA, BiomedELECTRA, as a base model with task pretraining, is the best-performing model. Here again, all ATLOP models use NER predictions from the baseline GLiNER model. rebel-large
Test Micro-P
Test Micro-R
Test Micro-F1 7.2. Subtask 6.2.2 - TT-RE The dev and test sets results for TT-RE are shown in Table 11 and Table 12, respectively. Here, the ifgures are similar to the ones from the previous task, with the main diference being that on the test set this time "ATLOP + BiomedELECTRA + TP" trained with 200 epochs performs slightly better compared to the one trained with 100 epochs. Here, all ATLOP models use the NER predictions from the baseline GLiNER model.
For the final subtask, Ternary Mention-based Relation Extraction (TM-RE), the results are again fairly similar. The results on the dev set can be found in Table 13 and again demonstrate that ATLOP with BiomedELECTRA as a base model is the best performing model, with XLM-R being slightly behind and rebel-large showing strong performance while being an end-to-end model. Here, all ATLOP models use the NER predictions from the baseline GLiNER model.
The results on the test set are shown in Table 14. This time XLM-R slightly outperforms BiomedELECTRA, and rebel-large is not much behind. Here, is observed significant degradation in performance compared to the previous tasks. This is partially due to the accumulation of errors following each task, as the final task contains each of the other subtasks in itself. This could also partially be due to the fact that the search space becomes much larger with the final subtask. Let be the input length (number of tokens). Let be the number of possible spans (ordered contiguous sub-sequences of tokens) in an input of length . Then the number of such spans is = (+1) . Let the number of possible classes 2 for entities be classes. Then, for the NER subtask, the search space is , which is exponential with respect to the input length. For the relation extraction subtask, the number of possible relations is 2 assuming that each entity can be in relation to each other since all relations could be in the document or none of them the total search space is 22 . Which in our case is much smaller because the number of possible types is much smaller than the number of tokens in a document. But we still see some performance degradation for this subtask. Some of the models, like ATLOP, require the predictions of the NER subtask, which explains the degradation of the performance. REBEL is a generative model, so technically it needs to search through an infinite search space to generate the answer, but potentially instead of generating the answer one could take take the averaged token log probabilities by giving each answer to the model, and could achieve better performance, but this was not done in time for the competition. For the TM-RE subtask, the search space increases to 22 where is the number of possible relation types, assuming that all relation types and all entities are possible. This is not a significant increase in the search space, which is also shown in the results as the performance for this subtask is just a few percent lower than the BT-RE subtask. For the TM-RE, the search space increases to (2 + 1)2 = (2 + 1)( (+1) )2 , assuming that each entity can be in relation with each other 2 and with every relation type, and assuming there could be an overlap between entities. Although this is an overestimation of the search space, it still shows that for this subtask, the search space is significantly larger, thus at least partially explaining the significantly lower results for the last subtask compared to the other subtasks. Here, all ATLOP models use the NER predictions from the baseline GLiNER model.
For subtasks 6.2.1, 6.2.2 and 6.2.3, a hybrid system with another participant in the challenge CLEANR4 [25]) is proposed. The results are combined by taking the union and intersection of both solutions to combine the systems’ strengths. The results of the union achieve higher results than any of the systems alone on subtasks 6.2.1 and 6.2.2. On subtask 6.2.3 our system alone demonstrates better capabilities. This means that both systems are complementary. Results are shown on Table 15 and Table 16.
CLEANR utilizes RAG as its approach to incorporate detailed training data through semantic retrieval processes in the prompt for the language model (LM). This few-shot approach, combined with dynamic retrieval, enables the system to be extended or “retrained” by simply adding or reweighting the training 4https://github.com/Dakantz/CLEANR samples. CLEANR extends the approach by introducing a reweighting of the samples in the retrieval process to prefer samples with a higher degree of confidence (i.e., prefer the Gold annotations over the Bronze annotations in our setting). A sentence-transformer system is used to embed the given training samples and store them in a Postgres database using the pgvector extension. CLEANR utilizes llama-cpp and llama-cpp-agent for both eficient inference of pre-trained models and constrained generation from a provided grammar. The grammar is generated using dynamically created models, which are transformed into the GBNF syntax, which is then used to constrain the LM output to the exact schema provided by the challenge. CLEANR’s output was combined with the results from ATLOP + BiomedELECTRA + TP (100 epochs).
This challenge is of major importance for bootstrapping the development of tools for automated analysis of gut-brain related literature and therefore facilitating the research in the area. The provided data is of good quantity as a starting point for fine-tuning deep learning models, however, some categories of named entities are underrepresented and need alternative approaches. E.g. for genes extraction, gazetteers may increase recall but then many eforts to remove noise are also necessary. And, of course, more annotated data would potentially help for better training of these models.
Our takeaways from the NER task are two: (i) GLiNER deserves more attention, it is our top performing system on the NER task, it shows best capabilities to generalize and it is worth exploring more in depth; (ii) progressive training of the models on the dev set is not showing efective results on the test set - all these models degrade their performance on the test set in comparison with the dev set. This could also mean that the dev and train set are not really close in means of annotation agreement.
For the relation extraction task and ATLOP especially, we come to the conclusion that the model is very sensitive to the initial set of named entities it works on. If the model is provided with more entities than the ones entering in relations then the performance heavily drops. E.g. the ATLOP results are worse when the golden named entities are provided as input than when working on the GLiNER entities. Therefore initial pre-processing of the supplied named entities improves the results.
The hybrid system in which our RE model was part of outperformed our own system which means that the results of both systems are complementary and not overlapping, therefore combing a transformerbased language model with an LLM-based approach seems to be promising for further research.
This work is partially supported by European Union’s Horizon research and innovation programme projects RES-Q PLUS [Grant Agreement No. 101057603] and HEREDITARY [Grant Agreement No. 101137074]. Views and opinions expressed are however those of the author only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them.
The authors have not employed any Generative AI tools. comprehensive survey, recent advancements, and future research directions, Neurocomputing (2024) 129171. [12] L. Luo, P.-T. Lai, C.-H. Wei, C. N. Arighi, Z. Lu, Biored: a rich biomedical relation extraction dataset, Briefings in Bioinformatics 23 (2022) bbac282. [13] A. Névéol, R. I. Doğan, Z. Lu, Semi-automatic semantic annotation of pubmed queries: a study on quality, eficiency, satisfaction, Journal of biomedical informatics 44 (2011) 310–318. [14] M. Sänger, U. Leser, Large-scale entity representation learning for biomedical relationship extraction, Bioinformatics 37 (2021) 236–242. [15] N. A. A. Hassan, R. A. A. A. A. Seoud, D. A. Salem, Open information extraction methodology for a new curated biomedical literature dataset, International Journal of Advanced Computer Science and Applications 14 (2023). [16] W. Zhou, K. Huang, T. Ma, J. Huang, Document-level relation extraction with adaptive thresholding and localized context pooling, Proceedings of the AAAI Conference on Artificial Intelligence 35 (2021) 14612–14620. doi:10.1609/aaai.v35i16.17717. [17] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. URL: https://arxiv.org/abs/1810.04805. arXiv:1810.04805. [18] M.-T. Clark, Kevin andLuong, Q. V. Le, C. D. Manning, Electra: Pre-training text encoders as discriminators rather than generators (2020). [19] L. Ramshaw, M. Marcus, Text chunking using transformation-based learning, in: Third Workshop on Very Large Corpora, 1995. URL: https://aclanthology.org/W95-0107/. [20] M. Neumann, D. King, I. Beltagy, W. Ammar, Scispacy: Fast and robust models for biomedical natural language processing, in: Proceedings of the 18th BioNLP Workshop and Shared Task (2019) 319-327, 2019. URL: https://arXiv:1902.07669. [21] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin,
Attention is all you need, 2023. URL: https://arxiv.org/abs/1706.03762. arXiv:1706.03762. [22] R. Jia, C. Wong, H. Poon, Document-level n-ary relation extraction with multiscale representation learning, in: J. Burstein, C. Doran, T. Solorio (Eds.), Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Association for Computational Linguistics, Minneapolis, Minnesota, 2019, pp. 3693–3704. URL: https://aclanthology.org/N19-1370/. doi:10. 18653/v1/N19-1370. [23] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V. Stoyanov, Roberta: A robustly optimized bert pretraining approach, 2019. URL: https://arxiv.org/abs/1907. 11692. arXiv:1907.11692. [24] A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, E. Grave, M. Ott, L. Zettlemoyer, V. Stoyanov, Unsupervised cross-lingual representation learning at scale, 2020.
URL: https://arxiv.org/abs/1911.02116. arXiv:1911.02116. [25] B. Kantz, P. Waldert, S. Lengauer, T. Schreck, Constrained linked entity annotation using rag (cleanr), in: G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 – Conference and Labs of the Evaluation Forum, CEUR Workshop Proceedings, CEUR-WS, 2025.