Temporal relation extraction in medical document analysis is crucial for understanding patient histories and treatment outcomes. This paper introduces a novel approach leveraging a bimodal model integrating textual content and a knowledge graph, to enhance temporal relation extraction. The paper presents ongoing research in constructing an optimal knowledge graph by augmenting PrimeKG with dynamically expanded information using a language model-generated knowledge graph, and further personalize the information with patient-specific graphs tailored for relation prediction. The pipeline for constructing this enriched knowledge graph is detailed, aiming to improve the capabilities of temporal relation extraction models. The preliminary results show that adding a simple knowledge graph to the temporal relation extraction model can significantly increase the performance, achieving new state-of-the-art results. While the research in using enhanced knowledge graphs is still ongoing, this paper lays the groundwork for leveraging common knowledge to advance temporal relation extraction in medical contexts. This approach holds promise for enhancing the understanding of patient histories and treatment outcomes, potentially leading to improved healthcare decision-making and patient care.

Understanding the relation between the meanings of words is an important part of comprehending natural language. Prior work has either focused on analysing lexical semantic relations in word embeddings or probing pretrained language models (PLMs), with some exceptions. Given the rarity of highly multilingual benchmarks, it is unclear to what extent PLMs capture relational knowledge and are able to transfer it across languages. To start addressing this question, we propose MultiLexBATS, a multilingual parallel dataset of lexical semantic relations adapted from BATS in 15 languages including low-resource languages, such as Bambara, Lithuanian, and Albanian. As experiment on cross-lingual transfer of relational knowledge, we test the PLMs’ ability to (1) capture analogies across languages, and (2) predict translation targets. We find considerable differences across relation types and languages with a clear preference for hypernymy and antonymy as well as romance languages.

This paper introduces the upgrade of a training corpus for linguistic annotation of modern standard Slovene. The enhancement spans both the size of the corpus and the depth of annotation layers. The revised SUK 1.0 corpus, building on its predecessor ssj500k 2.3, has doubled in size, containing over a million tokens. This expansion integrates three preexisting open-access datasets, all of which have undergone automatic tagging and meticulous manual review across multiple annotation layers, each represented in varying proportions. These layers span tokenization, segmentation, lemmatization, MULTEXT-East morphology, Universal Dependencies, JOS-SYN syntax, semantic role labeling, named entity recognition, and the newly incorporated coreferences. The paper illustrates the annotation processes for each layer while also presenting the results of the new CLASSLA-Stanza annotation tool, trained on the SUK corpus data. As one of the fundamental language resources of modern Slovene, the SUK corpus calls for constant development, as outlined in the concluding section.

This paper describes the University of Ljubljana (UL FRI) Group’s submissions to the shared task at the Balto-Slavic Natural Language Processing (BSNLP) 2021 Workshop. We experiment with multiple BERT-based models, pre-trained in multi-lingual, Croatian-Slovene-English and Slovene-only data. We perform training iteratively and on the concatenated data of previously available NER datasets. For the normalization task we use Stanza lemmatizer, while for entity matching we implemented a baseline using the Dedupe library. The performance of evaluations suggests that multi-source settings outperform less-resourced approaches. The best NER models achieve 0.91 F-score on Slovene training data splits while the best official submission achieved F-scores of 0.84 and 0.78 for relaxed partial matching and strict settings, respectively. In multi-lingual NER setting we achieve F-scores of 0.82 and 0.74.