Marko Robnik‐Šikonja

Boštjan Vouk, Matej Guid, Marko Robnik-Šikonja

Feature construction can contribute to comprehensibility and performance of machine learning models. Unfortunately, it usually requires exhaustive search in the attribute space or time-consuming human involvement to generate meaningful features. We propose a novel heuristic approach for reducing the search space based on aggregation of instance-based explanations of predictive models. The proposed Explainable Feature Construction (EFC) methodology identifies groups of co-occurring attributes exposed by popular explanation methods, such as IME and SHAP. We empirically show that reducing the search to these groups significantly reduces the time of feature construction using logical, relational, Cartesian, numerical, and threshold num-of-N and X-of-N constructive operators. An analysis on 10 transparent synthetic datasets shows that EFC effectively identifies informative groups of attributes and constructs relevant features. Using 30 real-world classification datasets, we show significant improvements in classification accuracy for several classifiers and demonstrate the feasibility of the proposed feature construction even for large datasets. Finally, EFC generated interpretable features on a real-world problem from the financial industry, which were confirmed by a domain expert.

Kristian Miok, Encarnacion Hidalgo-Tenorio, Petya Osenova et al.

Parliamentary and legislative debate transcripts provide informative insight into elected politicians' opinions, positions, and policy preferences. They are interesting for political and social sciences as well as linguistics and natural language processing (NLP) research. While existing research studied individual parliaments, we apply advanced NLP methods to a joint and comparative analysis of six national parliaments (Bulgarian, Czech, French, Slovene, Spanish, and United Kingdom) between 2017 and 2020. We analyze emotions and sentiment in the transcripts from the ParlaMint dataset collection and assess if the age, gender, and political orientation of speakers can be detected from their speeches. The results show some commonalities and many surprising differences among the analyzed countries.

Dimitar Trajanov, Vangel Trajkovski, Makedonka Dimitrieva et al.

Natural language processing (NLP) is an area of artificial intelligence that applies information technologies to process the human language, understand it to a certain degree, and use it in various applications. This area has rapidly developed in the last few years and now employs modern variants of deep neural networks to extract relevant patterns from large text corpora. The main objective of this work is to survey the recent use of NLP in the field of pharmacology. As our work shows, NLP is a highly relevant information extraction and processing approach for pharmacology. It has been used extensively, from intelligent searches through thousands of medical documents to finding traces of adversarial drug interactions in social media. We split our coverage into five categories to survey modern NLP methodology, commonly addressed tasks, relevant textual data, knowledge bases, and useful programming libraries. We split each of the five categories into appropriate subcategories, describe their main properties and ideas, and summarize them in a tabular form. The resulting survey presents a comprehensive overview of the area, useful to practitioners and interested observers.

Boshko Koloski, Blaž Škrlj, Marko Robnik-Šikonja et al.

The cross-lingual transfer is a promising technique to solve tasks in less-resourced languages. In this empirical study, we compare two fine-tuning approaches combined with zero-shot and full-shot learning approaches for large language models in a cross-lingual setting. As fine-tuning strategies, we compare parameter-efficient adapter methods with fine-tuning of all parameters. As cross-lingual transfer strategies, we compare the intermediate-training (\textit{IT}) that uses each language sequentially and cross-lingual validation (\textit{CLV}) that uses a target language already in the validation phase of fine-tuning. We assess the success of transfer and the extent of catastrophic forgetting in a source language due to cross-lingual transfer, i.e., how much previously acquired knowledge is lost when we learn new information in a different language. The results on two different classification problems, hate speech detection and product reviews, each containing datasets in several languages, show that the \textit{IT} cross-lingual strategy outperforms \textit{CLV} for the target language. Our findings indicate that, in the majority of cases, the \textit{CLV} strategy demonstrates superior retention of knowledge in the base language (English) compared to the \textit{IT} strategy, when evaluating catastrophic forgetting in multiple cross-lingual transfers.

Kristian Miok, Matej Klemen, Blaz Škrlj et al.

Explaining the predictions of neural models in clinical NLP remains a significant challenge, especially for complex tasks involving long, unstructured medical texts. While post-hoc methods like LIME and SHAP are widely used, they often fall short when applied to clinical narratives. In this paper, we identify core limitations of token-level and perturbation-based explanation techniques through targeted demonstra- tions on a hospital length-of-stay prediction task. Our findings reveal issues such as overemphasis on non-informative tokens, instability in at- tributions, and high-confidence predictions for incoherent input variants. These results underscore the need for explanation strategies that are clin- ically meaningful, semantically grounded, and robust to linguistic noise.

Matej Ulčar, Marko Robnik-Šikonja

Large pretrained language models have recently conquered the area of natural language processing. As an alternative to predominant masked language modelling introduced in BERT, the T5 model has introduced a more general training objective, namely sequence to sequence transformation, which includes masked language model but more naturally fits text generation tasks such as machine translation, summarization, question answering, text simplification, dialogue systems, etc. The monolingual variants of T5 models have been limited to well-resourced languages, while the massively multilingual T5 model supports 101 languages. In contrast, we trained two different sized T5-type sequence to sequence models for morphologically rich Slovene language with much less resources and analyzed their behavior on 11 tasks. Concerning classification tasks, the SloT5 models mostly lag behind the monolingual Slovene SloBERTa model but are useful for the generative tasks.

Vuk Đuranović, Marko Robnik Šikonja

Large language models (LLMs) demonstrate strong performance in text summarization, yet their effectiveness drops significantly across languages with restricted training resources. This work addresses the challenge of query-focused summarization (QFS) in less-resourced languages, where labeled datasets and evaluation tools are limited. We present a novel QFS framework, QFS-Composer, that integrates query decomposition, question generation (QG), question answering (QA), and abstractive summarization to improve the factual alignment of a summary with user intent. We test our approach on the Slovenian language. To enable high-quality supervision and evaluation, we develop the Slovenian QA and QG models based on a Slovene LLM and adapt evaluation approaches for reference-free summary evaluation. Empirical evaluation shows that the QA-guided summarization pipeline yields improved consistency and relevance over baseline LLMs. Our work establishes an extensible methodology for advancing QFS in less-resourced languages.

Marko Hostnik, Marko Robnik-Šikonja

The use of large language models (LLMs) is becoming increasingly widespread among software developers. However, privacy and computational requirements are problematic with commercial solutions and the use of LLMs. In this work, we focus on using relatively small and efficient LLMs with 160M parameters that are suitable for local execution and augmentation with retrieval from local projects. We train two open transformer-based models, the generative GPT-2 and the retrieval-adapted RETRO, on open-source Python files, and empirically compare them, confirming the benefits of embedding-based retrieval. Furthermore, we improve our models' performance with In-context retrieval-augmented generation (RAG), which retrieves code snippets using the Jaccard similarity of tokens. We evaluate In-context RAG on larger models and determine that, despite its simplicity, the approach is more suitable than using the RETRO architecture. Experimental results indicate that In-context RAG improves the code completion baseline by over 26%, while RETRO improves over the similarly sized GPT-2 baseline by 12%. We highlight the key role of proper tokenization in achieving the full potential of LLMs in code completion.

Aleš Žagar, Marko Robnik-Šikonja

Text summarization is an essential task in natural language processing, and researchers have developed various approaches over the years, ranging from rule-based systems to neural networks. However, there is no single model or approach that performs well on every type of text. We propose a system that recommends the most suitable summarization model for a given text. The proposed system employs a fully connected neural network that analyzes the input content and predicts which summarizer should score the best in terms of ROUGE score for a given input. The meta-model selects among four different summarization models, developed for the Slovene language, using different properties of the input, in particular its Doc2Vec document representation. The four Slovene summarization models deal with different challenges associated with text summarization in a less-resourced language. We evaluate the proposed SloMetaSum model performance automatically and parts of it manually. The results show that the system successfully automates the step of manually selecting the best model.

Katja Logar, Marko Robnik-Šikonja

Question answering is one of the most challenging tasks in language understanding. Most approaches are developed for English, while less-resourced languages are much less researched. We adapt a successful English question-answering approach, called UnifiedQA, to the less-resourced Slovene language. Our adaptation uses the encoder-decoder transformer SloT5 and mT5 models to handle four question-answering formats: yes/no, multiple-choice, abstractive, and extractive. We use existing Slovene adaptations of four datasets, and machine translate the MCTest dataset. We show that a general model can answer questions in different formats at least as well as specialized models. The results are further improved using cross-lingual transfer from English. While we produce state-of-the-art results for Slovene, the performance still lags behind English.

Lazar Đoković, Marko Robnik-Šikonja

The sarcasm detection task in natural language processing tries to classify whether an utterance is sarcastic or not. It is related to sentiment analysis since it often inverts surface sentiment. Because sarcastic sentences are highly dependent on context, and they are often accompanied by various non-verbal cues, the task is challenging. Most of related work focuses on high-resourced languages like English. To build a sarcasm detection dataset for a less-resourced language, such as Slovenian, we leverage two modern techniques: a machine translation specific medium-size transformer model, and a very large generative language model. We explore the viability of translated datasets and how the size of a pretrained transformer affects its ability to detect sarcasm. We train ensembles of detection models and evaluate models' performance. The results show that larger models generally outperform smaller ones and that ensembling can slightly improve sarcasm detection performance. Our best ensemble approach achieves an $\text{F}_1$-score of 0.765 which is close to annotators' agreement in the source language.

Matej Klemen, Tjaša Arčon, Luka Terčon et al.

Empirical grammar research has become increasingly data-driven, but the systematic analysis of annotated corpora still requires substantial methodological and technical effort. We explore how agentic large language models (LLMs) can streamline this process by reasoning over annotated corpora and producing interpretable, data-grounded answers to linguistic questions. We introduce an agentic framework for corpus-grounded grammatical analysis that integrates concepts such as natural-language task interpretation, code generation, and data-driven reasoning. As a proof of concept, we apply it to Universal Dependencies (UD) corpora, testing it on multilingual grammatical tasks inspired by the World Atlas of Language Structures (WALS). The evaluation spans 13 word-order features and over 170 languages, assessing system performance across three complementary dimensions - dominant-order accuracy, order-coverage completeness, and distributional fidelity - which reflect how well the system generalizes, identifies, and quantifies word-order variations. The results demonstrate the feasibility of combining LLM reasoning with structured linguistic data, offering a first step toward interpretable, scalable automation of corpus-based grammatical inquiry.

Tjaša Arčon, Marko Robnik-Šikonja, Polona Tratnik

Artificial intelligence approaches are being adapted to many research areas, including digital humanities. We built a methodology for large-scale analyses in folkloristics. Using machine learning and natural language processing, we automatically detected motifs in a large collection of Cinderella variants and analysed their similarities and differences with clustering and dimensionality reduction. The results show that large language models detect complex interactions in tales, enabling computational analysis of extensive text collections and facilitating cross-lingual comparisons.

Dario Vajda, Domen Vreš, Marko Robnik-Šikonja

Large language models have emerged as effective machine translation systems. In this paper, we explore how a general instruction-tuned large language model can be improved for machine translation using relatively few easily produced data resources. Using Slovene as a use case, we improve the GaMS-9B-Instruct model using Direct Preference Optimization (DPO) training on a programmatically curated and enhanced subset of a public dataset. As DPO requires pairs of quality-ranked instances, we generated its training dataset by translating English Wikipedia articles using two LLMs, GaMS-9B-Instruct and EuroLLM-9B-Instruct. We ranked the resulting translations based on heuristics coupled with automatic evaluation metrics such as COMET. The evaluation shows that our fine-tuned model outperforms both models involved in the dataset generation. In comparison to the baseline models, the fine-tuned model achieved a COMET score gain of around 0.04 and 0.02, respectively, on translating Wikipedia articles. It also more consistently avoids language and formatting errors.

Kristian Miok, Blaz Škrlj, Daniela Zaharie et al.

Clinical language models often struggle to provide trustworthy predictions and explanations when applied to lengthy, unstructured electronic health records (EHRs). This work introduces TT-XAI, a lightweight and effective framework that improves both classification performance and interpretability through domain-aware keyword distillation and reasoning with large language models (LLMs). First, we demonstrate that distilling raw discharge notes into concise keyword representations significantly enhances BERT classifier performance and improves local explanation fidelity via a focused variant of LIME. Second, we generate chain-of-thought clinical explanations using keyword-guided prompts to steer LLMs, producing more concise and clinically relevant reasoning. We evaluate explanation quality using deletion-based fidelity metrics, self-assessment via LLaMA-3 scoring, and a blinded human study with domain experts. All evaluation modalities consistently favor the keyword-augmented method, confirming that distillation enhances both machine and human interpretability. TT-XAI offers a scalable pathway toward trustworthy, auditable AI in clinical decision support.

Tadej Škvorc, Nikola Ivačič, Sebastjan Hribar et al.

To improve the reading experience, many news sites organize news into topical collections, called stories. In this work, we present an approach for implementing real-time story identification for a news monitoring system that automatically collects news articles as they appear online and processes them in various ways. Story identification aims to assign each news article to a specific story that the article is covering. The process is similar to text clustering and topic modeling, but requires that articles be grouped based on particular events, places, and people, rather than general text similarity (as in clustering) or general (predefined) topics (as in topic modeling). We present an approach to story identification that is capable of functioning in real time, assigning articles to stories as they are published online. In the proposed approach, we combine text representation techniques, clustering algorithms, and online topic modeling methods. We combine various text representation methods to extract specific events and named entities necessary for story identification, showing that a mixture of online topic-modeling approaches such as BERTopic, DBStream, and TextClust can be adapted for story discovery. We evaluate our approach on a news dataset from Slovene media covering a period of 1 month. We show that our real-time approach produces sensible results as judged by human evaluators.

Tadej Škvorc, Marko Robnik-Šikonja

Many less-resourced languages struggle with a lack of large, task-specific datasets that are required for solving relevant tasks with modern transformer-based large language models (LLMs). On the other hand, many linguistic resources, such as dictionaries, are rarely used in this context despite their large information contents. We show how LLMs can be used to extend existing language resources in less-resourced languages for two important tasks: word-sense disambiguation (WSD) and word-sense induction (WSI). We approach the two tasks through the related but much more accessible word-in-context (WiC) task where, given a pair of sentences and a target word, a classification model is tasked with predicting whether the sense of a given word differs between sentences. We demonstrate that a well-trained model for this task can distinguish between different word senses and can be adapted to solve the WSD and WSI tasks. The advantage of using the WiC task, instead of directly predicting senses, is that the WiC task does not need pre-constructed sense inventories with a sufficient number of examples for each sense, which are rarely available in less-resourced languages. We show that sentence pairs for the WiC task can be successfully generated from dictionary examples using LLMs. The resulting prediction models outperform existing models on WiC, WSD, and WSI tasks. We demonstrate our methodology on the Slovene language, where a monolingual dictionary is available, but word-sense resources are tiny.

Matej Klemen, Martin Božič, Špela Arhar Holdt et al.

Spell-checkers are valuable tools that enhance communication by identifying misspelled words in written texts. Recent improvements in deep learning, and in particular in large language models, have opened new opportunities to improve traditional spell-checkers with new functionalities that not only assess spelling correctness but also the suitability of a word for a given context. In our work, we present and compare two new spell-checkers and evaluate them on synthetic, learner, and more general-domain Slovene datasets. The first spell-checker is a traditional, fast, word-based approach, based on a morphological lexicon with a significantly larger word list compared to existing spell-checkers. The second approach uses a language model trained on a large corpus with synthetically inserted errors. We present the training data construction strategies, which turn out to be a crucial component of neural spell-checkers. Further, the proposed neural model significantly outperforms all existing spell-checkers for Slovene in both precision and recall.

Aleš Žagar, Marko Robnik-Šikonja

We present a Slovene combined machine-human translated SuperGLUE benchmark. We describe the translation process and problems arising due to differences in morphology and grammar. We evaluate the translated datasets in several modes: monolingual, cross-lingual, and multilingual, taking into account differences between machine and human translated training sets. The results show that the monolingual Slovene SloBERTa model is superior to massively multilingual and trilingual BERT models, but these also show a good cross-lingual performance on certain tasks. The performance of Slovene models still lags behind the best English models.

Matej Ulčar, Marko Robnik-Šikonja

Large pretrained masked language models have become state-of-the-art solutions for many NLP problems. While studies have shown that monolingual models produce better results than multilingual models, the training datasets must be sufficiently large. We trained a trilingual LitLat BERT-like model for Lithuanian, Latvian, and English, and a monolingual Est-RoBERTa model for Estonian. We evaluate their performance on four downstream tasks: named entity recognition, dependency parsing, part-of-speech tagging, and word analogy. To analyze the importance of focusing on a single language and the importance of a large training set, we compare created models with existing monolingual and multilingual BERT models for Estonian, Latvian, and Lithuanian. The results show that the newly created LitLat BERT and Est-RoBERTa models improve the results of existing models on all tested tasks in most situations.

Matej Klemen, Marko Robnik-Šikonja

Paraphrasing is a useful natural language processing task that can contribute to more diverse generated or translated texts. Natural language inference (NLI) and paraphrasing share some similarities and can benefit from a joint approach. We propose a novel methodology for the extraction of paraphrasing datasets from NLI datasets and cleaning existing paraphrasing datasets. Our approach is based on bidirectional entailment; namely, if two sentences can be mutually entailed, they are paraphrases. We evaluate our approach using several large pretrained transformer language models in the monolingual and cross-lingual setting. The results show high quality of extracted paraphrasing datasets and surprisingly high noise levels in two existing paraphrasing datasets.

Boshko Koloski, Timen Stepišnik-Perdih, Marko Robnik-Šikonja et al.

Increasing amounts of freely available data both in textual and relational form offers exploration of richer document representations, potentially improving the model performance and robustness. An emerging problem in the modern era is fake news detection -- many easily available pieces of information are not necessarily factually correct, and can lead to wrong conclusions or are used for manipulation. In this work we explore how different document representations, ranging from simple symbolic bag-of-words, to contextual, neural language model-based ones can be used for efficient fake news identification. One of the key contributions is a set of novel document representation learning methods based solely on knowledge graphs, i.e. extensive collections of (grounded) subject-predicate-object triplets. We demonstrate that knowledge graph-based representations already achieve competitive performance to conventionally accepted representation learners. Furthermore, when combined with existing, contextual representations, knowledge graph-based document representations can achieve state-of-the-art performance. To our knowledge this is the first larger-scale evaluation of how knowledge graph-based representations can be systematically incorporated into the process of fake news classification.

Matej Ulčar, Aleš Žagar, Carlos S. Armendariz et al.

The current dominance of deep neural networks in natural language processing is based on contextual embeddings such as ELMo, BERT, and BERT derivatives. Most existing work focuses on English; in contrast, we present here the first multilingual empirical comparison of two ELMo and several monolingual and multilingual BERT models using 14 tasks in nine languages. In monolingual settings, our analysis shows that monolingual BERT models generally dominate, with a few exceptions such as the dependency parsing task, where they are not competitive with ELMo models trained on large corpora. In cross-lingual settings, BERT models trained on only a few languages mostly do best, closely followed by massively multilingual BERT models.

Matej Ulčar, Marko Robnik-Šikonja

Building machine learning prediction models for a specific NLP task requires sufficient training data, which can be difficult to obtain for less-resourced languages. Cross-lingual embeddings map word embeddings from a less-resourced language to a resource-rich language so that a prediction model trained on data from the resource-rich language can also be used in the less-resourced language. To produce cross-lingual mappings of recent contextual embeddings, anchor points between the embedding spaces have to be words in the same context. We address this issue with a novel method for creating cross-lingual contextual alignment datasets. Based on that, we propose several cross-lingual mapping methods for ELMo embeddings. The proposed linear mapping methods use existing Vecmap and MUSE alignments on contextual ELMo embeddings. Novel nonlinear ELMoGAN mapping methods are based on GANs and do not assume isomorphic embedding spaces. We evaluate the proposed mapping methods on nine languages, using four downstream tasks: named entity recognition (NER), dependency parsing (DP), terminology alignment, and sentiment analysis. The ELMoGAN methods perform very well on the NER and terminology alignment tasks, with a lower cross-lingual loss for NER compared to the direct training on some languages. In DP and sentiment analysis, linear contextual alignment variants are more successful.

Domen Vreš, Marko Robnik Šikonja

Machine learning models are used in many sensitive areas where besides predictive accuracy their comprehensibility is also important. Interpretability of prediction models is necessary to determine their biases and causes of errors, and is a necessary prerequisite for users' confidence. For complex state-of-the-art black-box models post-hoc model-independent explanation techniques are an established solution. Popular and effective techniques, such as IME, LIME, and SHAP, use perturbation of instance features to explain individual predictions. Recently, Slack et al. (2020) put their robustness into question by showing that their outcomes can be manipulated due to poor perturbation sampling employed. This weakness would allow dieselgate type cheating of owners of sensitive models who could deceive inspection and hide potentially unethical or illegal biases existing in their predictive models. This could undermine public trust in machine learning models and give rise to legal restrictions on their use. We show that better sampling in these explanation methods prevents malicious manipulations. The proposed sampling uses data generators that learn the training set distribution and generate new perturbation instances much more similar to the training set. We show that the improved sampling increases the robustness of the LIME and SHAP, while previously untested method IME is already the most robust of all.

Aleš Žagar, Marko Robnik-Šikonja

Automatic text summarization extracts important information from texts and presents the information in the form of a summary. Abstractive summarization approaches progressed significantly by switching to deep neural networks, but results are not yet satisfactory, especially for languages where large training sets do not exist. In several natural language processing tasks, a cross-lingual model transfer is successfully applied in less-resource languages. For summarization, the cross-lingual model transfer was not attempted due to a non-reusable decoder side of neural models that cannot correct target language generation. In our work, we use a pre-trained English summarization model based on deep neural networks and sequence-to-sequence architecture to summarize Slovene news articles. We address the problem of inadequate decoder by using an additional language model for the evaluation of the generated text in target language. We test several cross-lingual summarization models with different amounts of target data for fine-tuning. We assess the models with automatic evaluation measures and conduct a small-scale human evaluation. Automatic evaluation shows that the summaries of our best cross-lingual model are useful and of quality similar to the model trained only in the target language. Human evaluation shows that our best model generates summaries with high accuracy and acceptable readability. However, similar to other abstractive models, our models are not perfect and may occasionally produce misleading or absurd content.

Matej Klemen, Luka Krsnik, Marko Robnik-Šikonja

Deep learning approaches are superior in NLP due to their ability to extract informative features and patterns from languages. The two most successful neural architectures are LSTM and transformers, used in large pretrained language models such as BERT. While cross-lingual approaches are on the rise, most current NLP techniques are designed and applied to English, and less-resourced languages are lagging behind. In morphologically rich languages, information is conveyed through morphology, e.g., through affixes modifying stems of words. Existing neural approaches do not explicitly use the information on word morphology. We analyse the effect of adding morphological features to LSTM and BERT models. As a testbed, we use three tasks available in many less-resourced languages: named entity recognition (NER), dependency parsing (DP), and comment filtering (CF). We construct baselines involving LSTM and BERT models, which we adjust by adding additional input in the form of part of speech (POS) tags and universal features. We compare models across several languages from different language families. Our results suggest that adding morphological features has mixed effects depending on the quality of features and the task. The features improve the performance of LSTM-based models on the NER and DP tasks, while they do not benefit the performance on the CF task. For BERT-based models, the morphological features only improve the performance on DP when they are of high quality while not showing practical improvement when they are predicted. Even for high-quality features, the improvements are less pronounced in language-specific BERT variants compared to massively multilingual BERT models. As in NER and CF datasets manually checked features are not available, we only experiment with predicted features and find that they do not cause any practical improvement in performance.

Kristian Miok, Gregor Pirs, Marko Robnik-Sikonja

Human annotations are an important source of information in the development of natural language understanding approaches. As under the pressure of productivity annotators can assign different labels to a given text, the quality of produced annotations frequently varies. This is especially the case if decisions are difficult, with high cognitive load, requires awareness of broader context, or careful consideration of background knowledge. To alleviate the problem, we propose two semi-supervised methods to guide the annotation process: a Bayesian deep learning model and a Bayesian ensemble method. Using a Bayesian deep learning method, we can discover annotations that cannot be trusted and might require reannotation. A recently proposed Bayesian ensemble method helps us to combine the annotators' labels with predictions of trained models. According to the results obtained from three hate speech detection experiments, the proposed Bayesian methods can improve the annotations and prediction performance of BERT models.

Tadej Škvorc, Polona Gantar, Marko Robnik-Šikonja

Idiomatic expressions can be problematic for natural language processing applications as their meaning cannot be inferred from their constituting words. A lack of successful methodological approaches and sufficiently large datasets prevents the development of machine learning approaches for detecting idioms, especially for expressions that do not occur in the training set. We present an approach, called MICE, that uses contextual embeddings for that purpose. We present a new dataset of multi-word expressions with literal and idiomatic meanings and use it to train a classifier based on two state-of-the-art contextual word embeddings: ELMo and BERT. We show that deep neural networks using both embeddings perform much better than existing approaches, and are capable of detecting idiomatic word use, even for expressions that were not present in the training set. We demonstrate cross-lingual transfer of developed models and analyze the size of the required dataset.

Kristian Miok, Blaz Skrlj, Daniela Zaharie et al.

Hate speech is an important problem in the management of user-generated content. To remove offensive content or ban misbehaving users, content moderators need reliable hate speech detectors. Recently, deep neural networks based on the transformer architecture, such as the (multilingual) BERT model, achieve superior performance in many natural language classification tasks, including hate speech detection. So far, these methods have not been able to quantify their output in terms of reliability. We propose a Bayesian method using Monte Carlo dropout within the attention layers of the transformer models to provide well-calibrated reliability estimates. We evaluate and visualize the results of the proposed approach on hate speech detection problems in several languages. Additionally, we test if affective dimensions can enhance the information extracted by the BERT model in hate speech classification. Our experiments show that Monte Carlo dropout provides a viable mechanism for reliability estimation in transformer networks. Used within the BERT model, it ofers state-of-the-art classification performance and can detect less trusted predictions. Also, it was observed that affective dimensions extracted using sentic computing methods can provide insights toward interpretation of emotions involved in hate speech. Our approach not only improves the classification performance of the state-of-the-art multilingual BERT model but the computed reliability scores also significantly reduce the workload in an inspection of ofending cases and reannotation campaigns. The provided visualization helps to understand the borderline outcomes.

Matej Ulčar, Marko Robnik-Šikonja

Large pretrained masked language models have become state-of-the-art solutions for many NLP problems. The research has been mostly focused on English language, though. While massively multilingual models exist, studies have shown that monolingual models produce much better results. We train two trilingual BERT-like models, one for Finnish, Estonian, and English, the other for Croatian, Slovenian, and English. We evaluate their performance on several downstream tasks, NER, POS-tagging, and dependency parsing, using the multilingual BERT and XLM-R as baselines. The newly created FinEst BERT and CroSloEngual BERT improve the results on all tasks in most monolingual and cross-lingual situations

Marko Robnik-Sikonja, Kristjan Reba, Igor Mozetic

Word embeddings represent words in a numeric space so that semantic relations between words are represented as distances and directions in the vector space. Cross-lingual word embeddings transform vector spaces of different languages so that similar words are aligned. This is done by constructing a mapping between vector spaces of two languages or learning a joint vector space for multiple languages. Cross-lingual embeddings can be used to transfer machine learning models between languages, thereby compensating for insufficient data in less-resourced languages. We use cross-lingual word embeddings to transfer machine learning prediction models for Twitter sentiment between 13 languages. We focus on two transfer mechanisms that recently show superior transfer performance. The first mechanism uses the trained models whose input is the joint numerical space for many languages as implemented in the LASER library. The second mechanism uses large pretrained multilingual BERT language models. Our experiments show that the transfer of models between similar languages is sensible, even with no target language data. The performance of cross-lingual models obtained with the multilingual BERT and LASER library is comparable, and the differences are language-dependent. The transfer with CroSloEngual BERT, pretrained on only three languages, is superior on these and some closely related languages.

Kristian Miok, Dong Nguyen-Doan, Marko Robnik-Šikonja et al.

Due to complex experimental settings, missing values are common in biomedical data. To handle this issue, many methods have been proposed, from ignoring incomplete instances to various data imputation approaches. With the recent rise of deep neural networks, the field of missing data imputation has oriented towards modelling of the data distribution. This paper presents an approach based on Monte Carlo dropout within (Variational) Autoencoders which offers not only very good adaptation to the distribution of the data but also allows generation of new data, adapted to each specific instance. The evaluation shows that the imputation error and predictive similarity can be improved with the proposed approach.

Blaž Škrlj, Nika Eržen, Shane Sheehan et al.

Neural language models are becoming the prevailing methodology for the tasks of query answering, text classification, disambiguation, completion and translation. Commonly comprised of hundreds of millions of parameters, these neural network models offer state-of-the-art performance at the cost of interpretability; humans are no longer capable of tracing and understanding how decisions are being made. The attention mechanism, introduced initially for the task of translation, has been successfully adopted for other language-related tasks. We propose AttViz, an online toolkit for exploration of self-attention---real values associated with individual text tokens. We show how existing deep learning pipelines can produce outputs suitable for AttViz, offering novel visualizations of the attention heads and their aggregations with minimal effort, online. We show on examples of news segments how the proposed system can be used to inspect and potentially better understand what a model has learned (or emphasized).

Carlos Santos Armendariz, Matthew Purver, Matej Ulčar et al.

State of the art natural language processing tools are built on context-dependent word embeddings, but no direct method for evaluating these representations currently exists. Standard tasks and datasets for intrinsic evaluation of embeddings are based on judgements of similarity, but ignore context; standard tasks for word sense disambiguation take account of context but do not provide continuous measures of meaning similarity. This paper describes an effort to build a new dataset, CoSimLex, intended to fill this gap. Building on the standard pairwise similarity task of SimLex-999, it provides context-dependent similarity measures; covers not only discrete differences in word sense but more subtle, graded changes in meaning; and covers not only a well-resourced language (English) but a number of less-resourced languages. We define the task and evaluation metrics, outline the dataset collection methodology, and describe the status of the dataset so far.

Matej Ulčar, Marko Robnik-Šikonja

Recent results show that deep neural networks using contextual embeddings significantly outperform non-contextual embeddings on a majority of text classification task. We offer precomputed embeddings from popular contextual ELMo model for seven languages: Croatian, Estonian, Finnish, Latvian, Lithuanian, Slovenian, and Swedish. We demonstrate that the quality of embeddings strongly depends on the size of training set and show that existing publicly available ELMo embeddings for listed languages shall be improved. We train new ELMo embeddings on much larger training sets and show their advantage over baseline non-contextual FastText embeddings. In evaluation, we use two benchmarks, the analogy task and the NER task.

Matej Ulčar, Kristiina Vaik, Jessica Lindström et al.

In text processing, deep neural networks mostly use word embeddings as an input. Embeddings have to ensure that relations between words are reflected through distances in a high-dimensional numeric space. To compare the quality of different text embeddings, typically, we use benchmark datasets. We present a collection of such datasets for the word analogy task in nine languages: Croatian, English, Estonian, Finnish, Latvian, Lithuanian, Russian, Slovenian, and Swedish. We redesigned the original monolingual analogy task to be much more culturally independent and also constructed cross-lingual analogy datasets for the involved languages. We present basic statistics of the created datasets and their initial evaluation using fastText embeddings.

Kristian Miok, Dong Nguyen-Doan, Daniela Zaharie et al.

For many analytical problems the challenge is to handle huge amounts of available data. However, there are data science application areas where collecting information is difficult and costly, e.g., in the study of geological phenomena, rare diseases, faults in complex systems, insurance frauds, etc. In many such cases, generators of synthetic data with the same statistical and predictive properties as the actual data allow efficient simulations and development of tools and applications. In this work, we propose the incorporation of Monte Carlo Dropout method within Autoencoder (MCD-AE) and Variational Autoencoder (MCD-VAE) as efficient generators of synthetic data sets. As the Variational Autoencoder (VAE) is one of the most popular generator techniques, we explore its similarities and differences to the proposed methods. We compare the generated data sets with the original data based on statistical properties, structural similarity, and predictive similarity. The results obtained show a strong similarity between the results of VAE, MCD-VAE and MCD-AE; however, the proposed methods are faster and can generate values similar to specific selected initial instances.

Damjan Vavpotič, Marko Robnik-Šikonja, Tomaž Hovelja

Software development enterprises are under consistent pressure to improve their management techniques and development processes. These are comprised of several disciplines like requirements acquisition, design, coding, testing, etc. that must be continuously improved and individually tailored to suit specific software development project. This paper presents an evaluation approach that enables the enterprises to increase development process net benefits by improving disciplines' quality and increasing developers' satisfaction. Our approach builds on Kano's model of quality. Based on an empirical study of top 1000 enterprises from Slovenia we find that application of software development methodologies in individual development disciplines significantly relates to net benefits of IT projects. The results show that different types of Kano quality are present in individual disciplines. Enterprises should be cautious when altering must-be quality disciplines like testing or deployment as they can significantly disrupt the established routines, cause great dissatisfaction between developers and significantly reduce benefits. On the other hand, changing the attractive quality disciplines like requirements acquisition can notably increase developers' satisfaction and benefits but is less likely to disrupt the established routines.

Matej Martinc, Senja Pollak, Marko Robnik-Šikonja

We present a set of novel neural supervised and unsupervised approaches for determining the readability of documents. In the unsupervised setting, we leverage neural language models, whereas in the supervised setting, three different neural classification architectures are tested. We show that the proposed neural unsupervised approach is robust, transferable across languages and allows adaptation to a specific readability task and data set. By systematic comparison of several neural architectures on a number of benchmark and new labelled readability datasets in two languages, this study also offers a comprehensive analysis of different neural approaches to readability classification. We expose their strengths and weaknesses, compare their performance to current state-of-the-art classification approaches to readability, which in most cases still rely on extensive feature engineering, and propose possibilities for improvements.

Andreja Čufar, Aleš Mrhar, Marko Robnik-Šikonja

The survey data sets are important sources of data and their successful exploitation is of key importance for informed policy-decision making. We present how a survey analysis approach initially developed for customer satisfaction research in marketing can be adapted for the introduction of clinical pharmacy services into hospital. We use two analytical approaches to extract relevant managerial consequences. With OrdEval algorithm we first evaluate the importance of competences for the users of clinical pharmacy and extract their nature according to the users expectations. Next, we build a model for predicting a successful introduction of clinical pharmacy to the clinical departments. We the wards with the highest probability of successful cooperation with a clinical pharmacist. We obtain useful managerially relevant information from a relatively small sample of highly relevant respondents. We show how the OrdEval algorithm exploits the information hidden in the ordering of class and attribute values and their inherent correlation. Its output can be effectively visualized and complemented with confidence intervals.

Marko Robnik-Šikonja

There are plenty of problems where the data available is scarce and expensive. We propose a generator of semi-artificial data with similar properties to the original data which enables development and testing of different data mining algorithms and optimization of their parameters. The generated data allow a large scale experimentation and simulations without danger of overfitting. The proposed generator is based on RBF networks, which learn sets of Gaussian kernels. These Gaussian kernels can be used in a generative mode to generate new data from the same distributions. To assess quality of the generated data we evaluated the statistical properties of the generated data, structural similarity and predictive similarity using supervised and unsupervised learning techniques. To determine usability of the proposed generator we conducted a large scale evaluation using 51 UCI data sets. The results show a considerable similarity between the original and generated data and indicate that the method can be useful in several development and simulation scenarios. We analyze possible improvements in classification performance by adding different amounts of generated data to the training set, performance on high dimensional data sets, and conditions when the proposed approach is successful.