Natural Language Inference (NLI) evaluation is crucial for assessing language understanding models; however, popular datasets suffer from systematic spurious correlations that artificially inflate actual model performance. To address this, we propose a method for the automated creation of a challenging test set without relying on the manual construction of artificial and unrealistic examples. We categorize the test set of popular NLI datasets into three difficulty levels by leveraging methods that exploit training dynamics. This categorization significantly reduces spurious correlation measures, with examples labeled as having the highest difficulty showing markedly decreased performance and encompassing more realistic and diverse linguistic phenomena. When our characterization method is applied to the training set, models trained with only a fraction of the data achieve comparable performance to those trained on the full dataset, surpassing other dataset characterization techniques. Our research addresses limitations in NLI dataset construction, providing a more authentic evaluation of model performance with implications for diverse NLU applications.

In recent years, Large Language Models (LLMs) have achieved almost human-like performance on various tasks. While some LLMs have been trained on multilingual data, most of the training data is in English; hence, their performance in English greatly exceeds other languages. To our knowledge, we are the first to collect and translate a large collection of texts, instructions, and benchmarks and train, evaluate, and release open-source LLMs tailored for Romanian. We evaluate our methods on four different categories, including academic benchmarks, MT-Bench (manually translated), and a professionally built historical, cultural, and social benchmark adapted to Romanian. We argue for the usefulness and high performance of RoLLMs by obtaining state-of-the-art results across the board. We publicly release all resources (i.e., data, training and evaluation code, models) with the goal of supporting and encouraging research on Romanian LLMs while concurrently creating a generalizable recipe adequate for other low or less-resourced languages.

This paper introduces the Common Lit Augmented Student Summary Evaluation (CLASSE) corpus. The corpus comprises 11,213 summaries written over six prompts by students in grades 3-12 while using the CommonLit website. Each summary was scored by expert human raters on analytic features related to main points, details, organization, voice, paraphrasing, and language beyond the source text. The human scores were aggregated into two component scores related to content and wording. The final corpus was the focus of a Kaggle competition hosted in late 2022 and completed in 2023 in which over 2,000 teams participated. The paper includes a baseline scoring model for the corpus based on a Large Language Model (Longformer model). The paper also provides an overview of the winning models from the Kaggle competition.

Readability is a crucial characteristic of texts, greatly influencing comprehension and reading efficacy. Unfortunately, limited research is available for less-resourced languages, especially for young populations where its impact is even higher. This paper introduces a new readability tool for children’s literature in the Romanian language, explicitly targeting primary school students aged 7-11. The tool consists of a digital repository of school reading texts (self-compiled corpus) and a text analysis interface that generates automatic readability reports for uploaded short texts. The methodology involves extracting, testing, and calibrating a readability formula for Romanian using the children’s literature corpus. Related work on readability and readability tools is discussed, followed by a description of the children’s literature corpus and the platform functionalities. The first steps are presented towards validating the readability formula for children’s literature in Romanian using the ReaderBench framework, while calibration variables relevant to the Romanian language and children’s literature are examined. Currently, no existing platform integrates a research-based readability formula for the Romanian language, making this tool unique. Overall, this research contributes to applied corpus linguistics and Digital Humanities studies and offers a valuable resource for educators, parents, and children in accessing age-appropriate and readable texts.

Complex word identification (CWI) is a cornerstone process towards proper text simplification. CWI is highly dependent on context, whereas its difficulty is augmented by the scarcity of available datasets which vary greatly in terms of domains and languages. As such, it becomes increasingly more difficult to develop a robust model that generalizes across a wide array of input examples. In this paper, we propose a novel training technique for the CWI task based on domain adaptation to improve the target character and context representations. This technique addresses the problem of working with multiple domains, inasmuch as it creates a way of smoothing the differences between the explored datasets. Moreover, we also propose a similar auxiliary task, namely text simplification, that can be used to complement lexical complexity prediction. Our model obtains a boost of up to 2.42% in terms of Pearson Correlation Coefficients in contrast to vanilla training techniques, when considering the CompLex from the Lexical Complexity Prediction 2021 dataset. At the same time, we obtain an increase of 3% in Pearson scores, while considering a cross-lingual setup relying on the Complex Word Identification 2018 dataset. In addition, our model yields state-of-the-art results in terms of Mean Absolute Error.

Running large-scale pre-trained language models in computationally constrained environments remains a challenging problem yet to be addressed, while transfer learning from these models has become prevalent in Natural Language Processing tasks. Several solutions, including knowledge distillation, network quantization, or network pruning have been previously proposed; however, these approaches focus mostly on the English language, thus widening the gap when considering low-resource languages. In this work, we introduce three light and fast versions of distilled BERT models for the Romanian language: Distil-BERT-base-ro, Distil-RoBERT-base, and DistilMulti-BERT-base-ro. The first two models resulted from the individual distillation of knowledge from two base versions of Romanian BERTs available in literature, while the last one was obtained by distilling their ensemble. To our knowledge, this is the first attempt to create publicly available Romanian distilled BERT models, which were thoroughly evaluated on five tasks: part-of-speech tagging, named entity recognition, sentiment analysis, semantic textual similarity, and dialect identification. Our experimental results argue that the three distilled models offer performance comparable to their teachers, while being twice as fast on a GPU and ~35% smaller. In addition, we further test the similarity between the predictions of our students versus their teachers by measuring their label and probability loyalty, together with regression loyalty - a new metric introduced in this work.

In recent times, the detection of hate-speech, offensive, or abusive language in online media has become an important topic in NLP research due to the exponential growth of social media and the propagation of such messages, as well as their impact. Misogyny detection, even though it plays an important part in hate-speech detection, has not received the same attention. In this paper, we describe our classification systems submitted to the SemEval-2022 Task 5: MAMI - Multimedia Automatic Misogyny Identification. The shared task aimed to identify misogynous content in a multi-modal setting by analysing meme images together with their textual captions. To this end, we propose two models based on the pre-trained UNITER model, one enhanced with an image sentiment classifier, whereas the second leverages a Vocabulary Graph Convolutional Network (VGCN). Additionally, we explore an ensemble using the aforementioned models. Our best model reaches an F1-score of 71.4% in Sub-task A and 67.3% for Sub-task B positioning our team in the upper third of the leaderboard. We release the code and experiments for our models on GitHub.

The real-world impact of polarization and toxicity in the online sphere marked the end of 2020 and the beginning of this year in a negative way. Semeval-2021, Task 5 - Toxic Spans Detection is based on a novel annotation of a subset of the Jigsaw Unintended Bias dataset and is the first language toxicity detection task dedicated to identifying the toxicity-level spans. For this task, participants had to automatically detect character spans in short comments that render the message as toxic. Our model considers applying Virtual Adversarial Training in a semi-supervised setting during the fine-tuning process of several Transformer-based models (i.e., BERT and RoBERTa), in combination with Conditional Random Fields. Our approach leads to performance improvements and more robust models, enabling us to achieve an F1-score of 65.73% in the official submission and an F1-score of 66.13% after further tuning during post-evaluation.

Extracting semantic information on measurements and counts is an important topic in terms of analyzing scientific discourses. The 8th task of SemEval-2021: Counts and Measurements (MeasEval) aimed to boost research in this direction by providing a new dataset on which participants train their models to extract meaningful information on measurements from scientific texts. The competition is composed of five subtasks that build on top of each other: (1) quantity span identification, (2) unit extraction from the identified quantities and their value modifier classification, (3) span identification for measured entities and measured properties, (4) qualifier span identification, and (5) relation extraction between the identified quantities, measured entities, measured properties, and qualifiers. We approached these challenges by first identifying the quantities, extracting their units of measurement, classifying them with corresponding modifiers, and afterwards using them to jointly solve the last three subtasks in a multi-turn question answering manner. Our best performing model obtained an overlapping F1-score of 36.91% on the test set.

Reading is a complex process which requires proper understanding of texts in order to create coherent mental representations. However, comprehension problems may arise due to hard-to-understand sections, which can prove troublesome for readers, while accounting for their specific language skills. As such, steps towards simplifying these sections can be performed, by accurately identifying and evaluating difficult structures. In this paper, we describe our approach for the SemEval-2021 Task 1: Lexical Complexity Prediction competition that consists of a mixture of advanced NLP techniques, namely Transformer-based language models, pre-trained word embeddings, Graph Convolutional Networks, Capsule Networks, as well as a series of hand-crafted textual complexity features. Our models are applicable on both subtasks and achieve good performance results, with a MAE below 0.07 and a Person correlation of .73 for single word identification, as well as a MAE below 0.08 and a Person correlation of .79 for multiple word targets. Our results are just 5.46% and 6.5% lower than the top scores obtained in the competition on the first and the second subtasks, respectively.

Detecting humor is a challenging task since words might share multiple valences and, depending on the context, the same words can be even used in offensive expressions. Neural network architectures based on Transformer obtain state-of-the-art results on several Natural Language Processing tasks, especially text classification. Adversarial learning, combined with other techniques such as multi-task learning, aids neural models learn the intrinsic properties of data. In this work, we describe our adversarial multi-task network, AMTL-Humor, used to detect and rate humor and offensive texts from Task 7 at SemEval-2021. Each branch from the model is focused on solving a related task, and consists of a BiLSTM layer followed by Capsule layers, on top of BERTweet used for generating contextualized embeddings. Our best model consists of an ensemble of all tested configurations, and achieves a 95.66% F1-score and 94.70% accuracy for Task 1a, while obtaining RMSE scores of 0.6200 and 0.5318 for Tasks 1b and 2, respectively.

This paper presents our contribution to the Social Media Mining for Health Applications Shared Task 2021. We addressed all the three subtasks of Task 1: Subtask A (classification of tweets containing adverse effects), Subtask B (extraction of text spans containing adverse effects) and Subtask C (adverse effects resolution). We explored various pre-trained transformer-based language models and we focused on a multi-task training architecture. For the first subtask, we also applied adversarial augmentation techniques and we formed model ensembles in order to improve the robustness of the prediction. Our system ranked first at Subtask B with 0.51 F1 score, 0.514 precision and 0.514 recall. For Subtask A we obtained 0.44 F1 score, 0.49 precision and 0.39 recall and for Subtask C we obtained 0.16 F1 score with 0.16 precision and 0.17 recall.

With the increasing adoption of technology, more and more systems become target to information security breaches. In terms of readily identifying zero-day vulnerabilities, a substantial number of news outlets and social media accounts reveal emerging vulnerabilities and threats. However, analysts often spend a lot of time looking through these decentralized sources of information in order to ensure up-to-date countermeasures and patches applicable to their organisation’s information systems. Various automated processing pipelines grounded in Natural Language Processing techniques for text classification were introduced for the early identification of vulnerabilities starting from Open-Source Intelligence (OSINT) data, including news websites, blogs, and social media. In this study, we consider a corpus of more than 1600 labeled news articles, and introduce an interpretable approach to the subject of cyberthreat early detection. In particular, an interpretable classification is performed using the Longformer architecture alongside prototypes from the ProSeNet structure, after performing a preliminary analysis on the Transformer’s encoding capabilities. The best interpretable architecture achieves an 88% F2-Score, arguing for the system’s applicability in real-life monitoring conditions of OSINT data.

Deep pre-trained language models tend to become ubiquitous in the field of Natural Language Processing (NLP). These models learn contextualized representations by using a huge amount of unlabeled text data and obtain state of the art results on a multitude of NLP tasks, by enabling efficient transfer learning. For other languages besides English, there are limited options of such models, most of which are trained only on multi-lingual corpora. In this paper we introduce a Romanian-only pre-trained BERT model – RoBERT – and compare it with different multi-lingual models on seven Romanian specific NLP tasks grouped into three categories, namely: sentiment analysis, dialect and cross-dialect topic identification, and diacritics restoration. Our model surpasses the multi-lingual models, as well as a another mono-lingual implementation of BERT, on all tasks.

Manipulative and misleading news have become a commodity for some online news outlets and these news have gained a significant impact on the global mindset of people. Propaganda is a frequently employed manipulation method having as goal to influence readers by spreading ideas meant to distort or manipulate their opinions. This paper describes our participation in the SemEval-2020, Task 11: Detection of PropagandaTechniques in News Articles competition. Our approach considers specializing a pre-trained BERT model on propagandistic and hyperpartisan news articles, enabling it to create more adequate representations for the two subtasks, namely propaganda Span Identification (SI) and propaganda Technique Classification (TC). Our proposed system achieved a F1-score of 46.060% in subtask SI, ranking 5th in the leaderboard from 36 teams and a micro-averaged F1 score of 54.302% for subtask TC, ranking 19th from 32 teams.

Financial causality detection is centered on identifying connections between different assets from financial news in order to improve trading strategies. FinCausal 2020 - Causality Identification in Financial Documents – is a competition targeting to boost results in financial causality by obtaining an explanation of how different individual events or chain of events interact and generate subsequent events in a financial environment. The competition is divided into two tasks: (a) a binary classification task for determining whether sentences are causal or not, and (b) a sequence labeling task aimed at identifying elements related to cause and effect. Various Transformer-based language models were fine-tuned for the first task and we obtained the second place in the competition with an F1-score of 97.55% using an ensemble of five such language models. Subsequently, a BERT model was fine-tuned for the second task and a Conditional Random Field model was used on top of the generated language features; the system managed to identify the cause and effect relationships with an F1-score of 73.10%. We open-sourced the code and made it available at: https://github.com/avramandrei/FinCausal2020.

Information on drug administration is obtained traditionally from doctors and pharmacists, as well as leaflets which provide in most cases cumbersome and hard-to-follow details. Thus, the need for medical knowledge bases emerges to provide access to concrete and well-structured information which can play an important role in informing patients. This paper introduces a Romanian medical knowledge base focused on drug-drug interactions, on representing relevant drug information, and on symptom-disease relations. The knowledge base was created by extracting and transforming information using Natural Language Processing techniques from both structured and unstructured sources, together with manual annotations. The resulting Romanian ontologies are aligned with larger English medical ontologies. Our knowledge base supports queries regarding drugs (e.g., active ingredients, concentration, expiration date), drug-drug interaction, symptom-disease relations, as well as drug-symptom relations.