This article introduces SM-FEEL-BG – the first Bulgarian-language package, containing 6 datasets with Social Media (SM) texts with emotion, feeling, and sentiment labels and 4 classifiers trained on them. All but one dataset from these are freely accessible for research purposes. The largest dataset contains 6000 Twitter, Telegram, and Facebook texts, manually annotated with 21 fine-grained emotion/feeling categories. The fine-grained labels are automatically merged into three coarse-grained sentiment categories, producing a dataset with two parallel sets of labels. Several classification experiments are run on different subsets of the fine-grained categories and their respective sentiment labels with a Bulgarian fine-tuned BERT. The highest Acc. reached was 0.61 for 16 emotions and 0.70 for 11 emotions (incl. 310 ChatGPT 4-generated texts). The sentiments Acc. of the 11 emotions dataset was also the highest (0.79). As Facebook posts cannot be shared, we ran experiments on the Twitter and Telegram subset of the 11 emotions dataset, obtaining 0.73 Acc. for emotions and 0.80 for sentiments. The article describes the annotation procedures, guidelines, experiments, and results. We believe that this package will be of significant benefit to researchers working on emotion detection and sentiment analysis in Bulgarian.

This paper presents an approach for training lightweight and robust language models for Bulgarian that mitigate gender, political, racial, and other biases in the data. Our method involves scraping content from major Bulgarian online media providers using a specialized procedure for source filtering, topic selection, and lexicon-based removal of inappropriate language during the pre-training phase. We continuously improve the models by incorporating new data from various domains, including social media, books, scientific literature, and linguistically modified corpora. Our motivation is to provide a solution that is sufficient for all natural language processing tasks in Bulgarian, and to address the lack of existing procedures for guaranteeing the robustness of such models.

Textual deepfakes can cause harm, especially on social media. At the moment, there are models trained to detect deepfake messages mainly for the English language, but no research or datasets currently exist for detecting them in most low-resource languages, such as Bulgarian. To address this gap, we explore three approaches. First, we machine translate an English-language social media dataset with bot messages into Bulgarian. However, the translation quality is unsatisfactory, leading us to create a new Bulgarian-language dataset with real social media messages and those generated by two language models (a new Bulgarian GPT-2 model – GPT-WEB-BG, and ChatGPT). We machine translate it into English and test existing English GPT-2 and ChatGPT detectors on it, achieving only 0.44-0.51 accuracy. Next, we train our own classifiers on the Bulgarian dataset, obtaining an accuracy of 0.97. Additionally, we apply the classifier with the highest results to a recently released Bulgarian social media dataset with manually fact-checked messages, which successfully identifies some of the messages as generated by Language Models (LM). Our results show that the use of machine translation is not suitable for textual deepfakes detection. We conclude that combining LM text detection with fact-checking is the most appropriate method for this task, and that identifying Bulgarian textual deepfakes is indeed possible.

The paper describes the Bulgarian Event Corpus (BEC). The annotation scheme is based on CIDOC-CRM ontology and on the English Framenet, adjusted for our task. It includes two main layers: named entities and events with their roles. The corpus is multi-domain and mainly oriented towards Social Sciences and Humanities (SSH). It will be used for: extracting knowledge and making it available through the Bulgaria-centric Knowledge Graph; further developing an annotation scheme that handles multiple domains in SSH; training automatic modules for the most important knowledge-based tasks, such as domain-specific and nested NER, NEL, event detection and profiling. Initial experiments were conducted on standard NER task due to complexity of the dataset and the rich NE annotation scheme. The results are promising with respect to some labels and give insights on handling better other ones. These experiments serve also as error detection modules that would help us in scheme re-design. They are a basis for further and more complex tasks, such as nested NER, NEL and event detection.

The paper describes a system for automatic summarization in English language of online news data that come from different non-English languages. The system is designed to be used in production environment for media monitoring. Automatic summarization can be very helpful in this domain when applied as a helper tool for journalists so that they can review just the important information from the news channels. However, like every software solution, the automatic summarization needs performance monitoring and assured safe environment for the clients. In media monitoring environment the most problematic features to be addressed are: the copyright issues, the factual consistency, the style of the text and the ethical norms in journalism. Thus, the main contribution of our present work is that the above mentioned characteristics are successfully monitored in neural automatic summarization models and improved with the help of validation, fact-preserving and fact-checking procedures.

The paper reports on the usage of deep learning methods for improving a Named Entity Recognition (NER) training corpus and for predicting and annotating new types in a test corpus. We show how the annotations in a type-based corpus of named entities (NE) were populated as occurrences within it, thus ensuring density of the training information. A deep learning model was adopted for discovering inconsistencies in the initial annotation and for learning new NE types. The evaluation results get improved after data curation, randomization and deduplication.

This paper reports on experiments with different stacks of word embeddings and evaluation of their usefulness for Bulgarian downstream tasks such as Named Entity Recognition and Classification (NERC) and Part-of-speech (POS) Tagging. Word embeddings stay in the core of the development of NLP, with several key language models being created over the last two years like FastText (CITATION), ElMo (CITATION), BERT (CITATION) and Flair (CITATION). Stacking or combining different word embeddings is another technique used in this paper and still not reported for Bulgarian NERC. Well-established architecture is used for the sequence tagging task such as BI-LSTM-CRF, and different pre-trained language models are combined in the embedding layer to decide which combination of them scores better.