Parallel natural language processing systems were previously successfully tested on the tasks of part-of-speech tagging and authorship attribution through mini-language modeling, for which they achieved significantly better results than independent methods in the cases of seven European languages. The aim of this paper is to present the advantages of using composite language models in the processing and evaluation of texts written in arbitrary highly inflective and morphology-rich natural language, particularly Serbian. A perplexity-based dataset, the main asset for the methodology assessment, was created using a series of generative pre-trained transformers trained on different representations of the Serbian language corpus and a set of sentences classified into three groups (expert translations, corrupted translations, and machine translations). The paper describes a comparative analysis of calculated perplexities in order to measure the classification capability of different models on two binary classification tasks. In the course of the experiment, we tested three standalone language models (baseline) and two composite language models (which are based on perplexities outputted by all three standalone models). The presented results single out a complex stacked classifier using a multitude of features extracted from perplexity vectors as the optimal architecture of composite language models for both tasks.

Considering the linguistic means for expressing the concept of a border in the Serbian language, Professor Predrag Piper listed several typical nouns that represent lexical means in that categorial-semantic complex. This paper investigates other nouns that serve this function in the Serbian language. The starting point is the assumption that the role of a boundary is probably performed by those nouns in the Serbian language that are more often than others used with prepositions, that is to say, preceded by a preposition in most cases. A list of all such nouns was made using the electronic corpus SrpKor, followed by excerpting only the nouns suggesting some type of a borderline which are most often used directly after the following prepositions: do, od, iz, na, oko, pred, u. Subsequently, a further selection was made based on semantic and pragmatic criteria, and thus the final list of delimiting nouns described in this paper was obtained.

This research presents the development and evaluation of Convolutional Neural Network (CNN)-based models for Named Entity Linking (NEL) in Serbian to the Wikidata knowledge base. It introduces four novel model configurations derived from two underlying Named Entity Recognition (NER) systems: the 7-class SrpCNNER2 and a specialized SrpCNNER2-loc model for locations. Different training strategies for entity linking were implemented, including training on all entity types ( srNEL-all ), the top three most frequent types ( srNEL-top3 ), locations only with NER ( srNEL-loc ), and a fully specialized location-only approach ( srNEL-LOCfull ). Models were trained on a corpus (73,493 sentences) comprising literary excerpts, legal documents, news articles, and synthetically generated sentences. The performance was evaluated on the sr-geography corpus derived from elementary school geography textbooks, using strict evaluation metrics and SrpCNNeL model as the baseline. The srNEL-all configuration, leveraging the comprehensive SrpCNNER2 NER base and trained on all entity types, achieves the highest performance on linking locations within the sr-geography corpus, attaining an F1 0.845, outperforming the SrpCNNeL baseline (F1 0.731) and models employing more specialized training regimes, such as srNEL-LOCfull (F1 0.411). The findings highlight the effectiveness of the proposed methods for Serbian NEL and underscore the benefits of incorporating broader entity context during training. While transformer-based dense-retrieval and RAG systems currently dominate multilingual entity linking research, this work prioritizes an efficient CNN-based architecture suitable for less-resource and morphologically rich languages, demonstrating competitive performance for Serbian. Although this system is faster and requires lower hardware requirements in terms of machine performance, the highest-performing model was compared with a transformer-based baseline, demonstrating its superiority over the CNN architecture. Future directions include further development of models for Serbian and exploration of hybrid architectures that combine efficiency with state-of-the-art performance.

Research on Serbian named entity recognition (NER) has evolved from early rule-based approaches to neural and transformer-based models, culminating in the TeslaNER+ dataset, which contains over 150,000 sentences annotated with eight entity types and linked to Wikidata. This paper explores the use of TeslaNER+ for relation extraction (RE), focusing on identifying and classifying semantic relations between entities in unstructured text. A predefined set of relations tailored to various domains, including literary texts is employed. After analyzing existing RE schemas, 19 relevant relations are selected. The methodology combines rule-based patterns with LLM-based prompting, leveraging lexically grounded patterns and prompts derived from key terms. Experiments were conducted on a subset of 23,000 sentences, and the paper reports initial results for selected relations. The FST approach achieves a precision of 0.77 and recall of 0.39, whereas the LLM model attains a higher recall (0.97) but lower precision (0.65). These results are approximate, as they are computed on a sampled subset rather than the full dataset. In terms of efficiency, the FST approach incurs no computational cost and enables fast extraction, while the LLM approach is significantly slower and involves additional expenses.