Ivan Koychev

Momchil Hardalov, Pepa Atanasova, Todor Mihaylov et al. · berkeley

We present bgGLUE(Bulgarian General Language Understanding Evaluation), a benchmark for evaluating language models on Natural Language Understanding (NLU) tasks in Bulgarian. Our benchmark includes NLU tasks targeting a variety of NLP problems (e.g., natural language inference, fact-checking, named entity recognition, sentiment analysis, question answering, etc.) and machine learning tasks (sequence labeling, document-level classification, and regression). We run the first systematic evaluation of pre-trained language models for Bulgarian, comparing and contrasting results across the nine tasks in the benchmark. The evaluation results show strong performance on sequence labeling tasks, but there is a lot of room for improvement for tasks that require more complex reasoning. We make bgGLUE publicly available together with the fine-tuning and the evaluation code, as well as a public leaderboard at https://bgglue.github.io/, and we hope that it will enable further advancements in developing NLU models for Bulgarian.

Ensiye Kiyamousavi, Boris Kraychev, Ivan Koychev

Federated learning (FL) is a decentralized machine learning approach where independent learners process data privately. Its goal is to create a robust and accurate model by aggregating and retraining local models over multiple rounds. However, FL faces challenges regarding data heterogeneity and model aggregation effectiveness. In order to simulate real-world data, researchers use methods for data partitioning that transform a dataset designated for centralized learning into a group of sub-datasets suitable for distributed machine learning with different data heterogeneity. In this paper, we study the currently popular data partitioning techniques and visualize their main disadvantages: the lack of precision in the data diversity, which leads to unreliable heterogeneity indexes, and the inability to incrementally challenge the FL algorithms. To resolve this problem, we propose a method that leverages entropy and symmetry to construct 'the most challenging' and controllable data distributions with gradual difficulty. We introduce a metric to measure data heterogeneity among the learning agents and a transformation technique that divides any dataset into splits with precise data diversity. Through a comparative study, we demonstrate the superiority of our method over existing FL data partitioning approaches, showcasing its potential to challenge model aggregation algorithms. Experimental results indicate that our approach gradually challenges the FL strategies, and the models trained on FedSym distributions are more distinct.

Petar Ivanov, Ivan Koychev, Momchil Hardalov et al.

Developing tools to automatically detect check-worthy claims in political debates and speeches can greatly help moderators of debates, journalists, and fact-checkers. While previous work on this problem has focused exclusively on the text modality, here we explore the utility of the audio modality as an additional input. We create a new multimodal dataset (text and audio in English) containing 48 hours of speech from past political debates in the USA. We then experimentally demonstrate that, in the case of multiple speakers, adding the audio modality yields sizable improvements over using the text modality alone; moreover, an audio-only model could outperform a text-only one for a single speaker. With the aim to enable future research, we make all our data and code publicly available at https://github.com/petar-iv/audio-checkworthiness-detection.

Rocktim Jyoti Das, Simeon Emilov Hristov, Haonan Li et al.

We introduce EXAMS-V, a new challenging multi-discipline multimodal multilingual exam benchmark for evaluating vision language models. It consists of 20,932 multiple-choice questions across 20 school disciplines covering natural science, social science, and other miscellaneous studies, e.g., religion, fine arts, business, etc. EXAMS-V includes a variety of multimodal features such as text, images, tables, figures, diagrams, maps, scientific symbols, and equations. The questions come in 11 languages from 7 language families. Unlike existing benchmarks, EXAMS-V is uniquely curated by gathering school exam questions from various countries, with a variety of education systems. This distinctive approach calls for intricate reasoning across diverse languages and relies on region-specific knowledge. Solving the problems in the dataset requires advanced perception and joint reasoning over the text and the visual content of the image. Our evaluation results demonstrate that this is a challenging dataset, which is difficult even for advanced vision-text models such as GPT-4V and Gemini; this underscores the inherent complexity of the dataset and its significance as a future benchmark.

Zhuohan Xie, Daniil Orel, Rushil Thareja et al.

Multi-step symbolic reasoning is essential for robust financial analysis; yet, current benchmarks largely overlook this capability. Existing datasets such as FinQA and ConvFinQA emphasize final numerical answers while neglecting the intermediate reasoning required for transparency and verification. To address this gap, we introduce FinChain, the first benchmark specifically designed for verifiable Chain-of-Thought (CoT) evaluation in finance. FinChain spans 58 topics across 12 financial domains, each represented by parameterized symbolic templates with executable Python traces that enable fully machine-verifiable reasoning and scalable, contamination-free data generation. To assess reasoning capacity, we propose ChainEval, a dynamic alignment metric that jointly evaluates both the final-answer correctness and the step-level reasoning consistency. Evaluating 26 leading LLMs reveals that even frontier proprietary systems exhibit clear limitations in symbolic financial reasoning, while domain-adapted and math-enhanced fine-tuned models substantially narrow this gap. Overall, FinChain exposes persistent weaknesses in multi-step financial reasoning and provides a foundation for developing trustworthy, interpretable, and verifiable financial AI.

Sylvia Vassileva, Ivan Koychev, Svetla Boytcheva

The linking of clinical entities is a crucial part of extracting structured information from clinical texts. It is the process of assigning a code from a medical ontology or classification to a phrase in the text. The International Classification of Diseases - 10th revision (ICD-10) is an international standard for classifying diseases for statistical and insurance purposes. Automatically assigning the correct ICD-10 code to terms in discharge summaries will simplify the work of healthcare professionals and ensure consistent coding in hospitals. Our paper proposes an approach for linking clinical terms to ICD-10 codes in different languages using Large Language Models (LLMs). The approach consists of a multistage pipeline that uses clinical dictionaries to match unambiguous terms in the text and then applies in-context learning with GPT-4.1 to predict the ICD-10 code for the terms that do not match the dictionary. Our system shows promising results in predicting ICD-10 codes on different benchmark datasets in Spanish - 0.89 F1 for categories and 0.78 F1 on subcategories on CodiEsp, and Greek - 0.85 F1 on ElCardioCC.

Tsvetomila Mihaylova, Pepa Gencheva, Martin Boyanov et al.

We present the system we built for participating in SemEval-2016 Task 3 on Community Question Answering. We achieved the best results on subtask C, and strong results on subtasks A and B, by combining a rich set of various types of features: semantic, lexical, metadata, and user-related. The most important group turned out to be the metadata for the question and for the comment, semantic vectors trained on QatarLiving data and similarities between the question and the comment for subtasks A and C, and between the original and the related question for Subtask B.

Todor Mihaylov, Ivan Koychev, Georgi Georgiev et al.

Recently, Web forums have been invaded by opinion manipulation trolls. Some trolls try to influence the other users driven by their own convictions, while in other cases they can be organized and paid, e.g., by a political party or a PR agency that gives them specific instructions what to write. Finding paid trolls automatically using machine learning is a hard task, as there is no enough training data to train a classifier; yet some test data is possible to obtain, as these trolls are sometimes caught and widely exposed. In this paper, we solve the training data problem by assuming that a user who is called a troll by several different people is likely to be such, and one who has never been called a troll is unlikely to be such. We compare the profiles of (i) paid trolls vs. (ii)"mentioned" trolls vs. (iii) non-trolls, and we further show that a classifier trained to distinguish (ii) from (iii) does quite well also at telling apart (i) from (iii).

Kristiyan Vachev, Momchil Hardalov, Georgi Karadzhov et al.

In education, open-ended quiz questions have become an important tool for assessing the knowledge of students. Yet, manually preparing such questions is a tedious task, and thus automatic question generation has been proposed as a possible alternative. So far, the vast majority of research has focused on generating the question text, relying on question answering datasets with readily picked answers, and the problem of how to come up with answer candidates in the first place has been largely ignored. Here, we aim to bridge this gap. In particular, we propose a model that can generate a specified number of answer candidates for a given passage of text, which can then be used by instructors to write questions manually or can be passed as an input to automatic answer-aware question generators. Our experiments show that our proposed answer candidate generation model outperforms several baselines.

Krasimira Bozhanova, Yoan Dinkov, Ivan Koychev et al.

We propose a novel framework for predicting the factuality of reporting of news media outlets by studying the user attention cycles in their YouTube channels. In particular, we design a rich set of features derived from the temporal evolution of the number of views, likes, dislikes, and comments for a video, which we then aggregate to the channel level. We develop and release a dataset for the task, containing observations of user attention on YouTube channels for 489 news media. Our experiments demonstrate both complementarity and sizable improvements over state-of-the-art textual representations.

Alex Nikolov, Giovanni Da San Martino, Ivan Koychev et al.

While misinformation and disinformation have been thriving in social media for years, with the emergence of the COVID-19 pandemic, the political and the health misinformation merged, thus elevating the problem to a whole new level and giving rise to the first global infodemic. The fight against this infodemic has many aspects, with fact-checking and debunking false and misleading claims being among the most important ones. Unfortunately, manual fact-checking is time-consuming and automatic fact-checking is resource-intense, which means that we need to pre-filter the input social media posts and to throw out those that do not appear to be check-worthy. With this in mind, here we propose a model for detecting check-worthy tweets about COVID-19, which combines deep contextualized text representations with modeling the social context of the tweet. We further describe a number of additional experiments and comparisons, which we believe should be useful for future research as they provide some indication about what techniques are effective for the task. Our official submission to the English version of CLEF-2020 CheckThat! Task 1, system Team_Alex, was ranked second with a MAP score of 0.8034, which is almost tied with the wining system, lagging behind by just 0.003 MAP points absolute.

Pepa Gencheva, Ivan Koychev, Lluís Màrquez et al.

In the context of investigative journalism, we address the problem of automatically identifying which claims in a given document are most worthy and should be prioritized for fact-checking. Despite its importance, this is a relatively understudied problem. Thus, we create a new dataset of political debates, containing statements that have been fact-checked by nine reputable sources, and we train machine learning models to predict which claims should be prioritized for fact-checking, i.e., we model the problem as a ranking task. Unlike previous work, which has looked primarily at sentences in isolation, in this paper we focus on a rich input representation modeling the context: relationship between the target statement and the larger context of the debate, interaction between the opponents, and reaction by the moderator and by the public. Our experiments show state-of-the-art results, outperforming a strong rivaling system by a margin, while also confirming the importance of the contextual information.

Momchil Hardalov, Ivan Koychev, Preslav Nakov

We study the problem of finding fake online news. This is an important problem as news of questionable credibility have recently been proliferating in social media at an alarming scale. As this is an understudied problem, especially for languages other than English, we first collect and release to the research community three new balanced credible vs. fake news datasets derived from four online sources. We then propose a language-independent approach for automatically distinguishing credible from fake news, based on a rich feature set. In particular, we use linguistic (n-gram), credibility-related (capitalization, punctuation, pronoun use, sentiment polarity), and semantic (embeddings and DBPedia data) features. Our experiments on three different testsets show that our model can distinguish credible from fake news with very high accuracy.

Yoan Dinkov, Ahmed Ali, Ivan Koychev et al.

We address the problem of predicting the leading political ideology, i.e., left-center-right bias, for YouTube channels of news media. Previous work on the problem has focused exclusively on text and on analysis of the language used, topics discussed, sentiment, and the like. In contrast, here we study videos, which yields an interesting multimodal setup. Starting with gold annotations about the leading political ideology of major world news media from Media Bias/Fact Check, we searched on YouTube to find their corresponding channels, and we downloaded a recent sample of videos from each channel. We crawled more than 1,000 YouTube hours along with the corresponding subtitles and metadata, thus producing a new multimodal dataset. We further developed a multimodal deep-learning architecture for the task. Our analysis shows that the use of acoustic signal helped to improve bias detection by more than 6% absolute over using text and metadata only. We release the dataset to the research community, hoping to help advance the field of multi-modal political bias detection.

Dimitrina Zlatkova, Preslav Nakov, Ivan Koychev

The recent explosion of false claims in social media and on the Web in general has given rise to a lot of manual fact-checking initiatives. Unfortunately, the number of claims that need to be fact-checked is several orders of magnitude larger than what humans can handle manually. Thus, there has been a lot of research aiming at automating the process. Interestingly, previous work has largely ignored the growing number of claims about images. This is despite the fact that visual imagery is more influential than text and naturally appears alongside fake news. Here we aim at bridging this gap. In particular, we create a new dataset for this problem, and we explore a variety of features modeling the claim, the image, and the relationship between the claim and the image. The evaluation results show sizable improvements over the baseline. We release our dataset, hoping to enable further research on fact-checking claims about images.

Yoan Dinkov, Ivan Koychev, Preslav Nakov

Online media aim for reaching ever bigger audience and for attracting ever longer attention span. This competition creates an environment that rewards sensational, fake, and toxic news. To help limit their spread and impact, we propose and develop a news toxicity detector that can recognize various types of toxic content. While previous research primarily focused on English, here we target Bulgarian. We created a new dataset by crawling a website that for five years has been collecting Bulgarian news articles that were manually categorized into eight toxicity groups. Then we trained a multi-class classifier with nine categories: eight toxic and one non-toxic. We experimented with different representations based on ElMo, BERT, and XLM, as well as with a variety of domain-specific features. Due to the small size of our dataset, we created a separate model for each feature type, and we ultimately combined these models into a meta-classifier. The evaluation results show an accuracy of 59.0% and a macro-F1 score of 39.7%, which represent sizable improvements over the majority-class baseline (Acc=30.3%, macro-F1=5.2%).

Momchil Hardalov, Ivan Koychev, Preslav Nakov

Recently, reading comprehension models achieved near-human performance on large-scale datasets such as SQuAD, CoQA, MS Macro, RACE, etc. This is largely due to the release of pre-trained contextualized representations such as BERT and ELMo, which can be fine-tuned for the target task. Despite those advances and the creation of more challenging datasets, most of the work is still done for English. Here, we study the effectiveness of multilingual BERT fine-tuned on large-scale English datasets for reading comprehension (e.g., for RACE), and we apply it to Bulgarian multiple-choice reading comprehension. We propose a new dataset containing 2,221 questions from matriculation exams for twelfth grade in various subjects -history, biology, geography and philosophy-, and 412 additional questions from online quizzes in history. While the quiz authors gave no relevant context, we incorporate knowledge from Wikipedia, retrieving documents matching the combination of question + each answer option. Moreover, we experiment with different indexing and pre-training strategies. The evaluation results show accuracy of 42.23%, which is well above the baseline of 24.89%.

Momchil Hardalov, Ivan Koychev, Preslav Nakov

Recent years have seen growing interest in conversational agents, such as chatbots, which are a very good fit for automated customer support because the domain in which they need to operate is narrow. This interest was in part inspired by recent advances in neural machine translation, esp. the rise of sequence-to-sequence (seq2seq) and attention-based models such as the Transformer, which have been applied to various other tasks and have opened new research directions in question answering, chatbots, and conversational systems. Still, in many cases, it might be feasible and even preferable to use simple information retrieval techniques. Thus, here we compare three different models:(i) a retrieval model, (ii) a sequence-to-sequence model with attention, and (iii) Transformer. Our experiments with the Twitter Customer Support Dataset, which contains over two million posts from customer support services of twenty major brands, show that the seq2seq model outperforms the other two in terms of semantics and word overlap.

Valentin Zmiycharov, Dimitar Alexandrov, Preslav Nakov et al.

We describe the system that our FMI@SU student's team built for participating in the Triple Scoring task at the WSDM Cup 2017. Given a triple from a "type-like" relation, profession or nationality, the goal is to produce a score, on a scale from 0 to 7, that measures the relevance of the statement expressed by the triple: e.g., how well does the profession of an Actor fit for Quentin Tarantino? We propose a distant supervision approach using information crawled from Wikipedia, DeletionPedia, and DBpedia, together with task-specific word embeddings, TF-IDF weights, and role occurrence order, which we combine in a linear regression model. The official evaluation ranked our submission 1st on Kendall's Tau, 7th on Average score difference, and 9th on Accuracy, out of 21 participating teams.

Martin Boyanov, Ivan Koychev, Preslav Nakov et al.

We propose to use question answering (QA) data from Web forums to train chatbots from scratch, i.e., without dialog training data. First, we extract pairs of question and answer sentences from the typically much longer texts of questions and answers in a forum. We then use these shorter texts to train seq2seq models in a more efficient way. We further improve the parameter optimization using a new model selection strategy based on QA measures. Finally, we propose to use extrinsic evaluation with respect to a QA task as an automatic evaluation method for chatbots. The evaluation shows that the model achieves a MAP of 63.5% on the extrinsic task. Moreover, it can answer correctly 49.5% of the questions when they are similar to questions asked in the forum, and 47.3% of the questions when they are more conversational in style.

Georgi Karadzhov, Preslav Nakov, Lluis Marquez et al.

Given the constantly growing proliferation of false claims online in recent years, there has been also a growing research interest in automatically distinguishing false rumors from factually true claims. Here, we propose a general-purpose framework for fully-automatic fact checking using external sources, tapping the potential of the entire Web as a knowledge source to confirm or reject a claim. Our framework uses a deep neural network with LSTM text encoding to combine semantic kernels with task-specific embeddings that encode a claim together with pieces of potentially-relevant text fragments from the Web, taking the source reliability into account. The evaluation results show good performance on two different tasks and datasets: (i) rumor detection and (ii) fact checking of the answers to a question in community question answering forums.

Todor Mihaylov, Daniel Belchev, Yasen Kiprov et al.

We transfer a key idea from the field of sentiment analysis to a new domain: community question answering (cQA). The cQA task we are interested in is the following: given a question and a thread of comments, we want to re-rank the comments so that the ones that are good answers to the question would be ranked higher than the bad ones. We notice that good vs. bad comments use specific vocabulary and that one can often predict the goodness/badness of a comment even ignoring the question, based on the comment contents only. This leads us to the idea to build a good/bad polarity lexicon as an analogy to the positive/negative sentiment polarity lexicons, commonly used in sentiment analysis. In particular, we use pointwise mutual information in order to build large-scale goodness polarity lexicons in a semi-supervised manner starting with a small number of initial seeds. The evaluation results show an improvement of 0.7 MAP points absolute over a very strong baseline and state-of-the art performance on SemEval-2016 Task 3.