Skewness in imbalanced datasets affects Automatic Text Classification (ATC), leading to classifier bias toward the majority classes. This work examines undersampling methods to mitigate such bias in Small and Large Language Model (SLMs and LLMs) classifiers. Based on the limitations found in existing solutions, we propose two novel undersampling methods inspired by state-of-the-art Instance Selection techniques, relying on calibrated confidences and semantic difficulty estimates. We compare them against 19 baselines across 13 datasets, evaluating: (i) effectiveness, (ii) class imbalance bias, (iii) efficiency, (iv) scalability, and (v) consistency. Results show our methods uniquely reduce classifier bias (up to 56%) across all datasets without effectiveness loss while improving efficiency (1.6x speedup), scalability and reducing carbon emissions (up to 50%).

Identifying subjective phenomena, such as irony in language, poses unique challenges, as these tasks involve subjective interpretation shaped by both cultural and individual perspectives. Unlike conventional models that rely on aggregated annotations, perspectivist approaches aim to capture the diversity of viewpoints by leveraging the knowledge of specific annotator groups, promoting fairness and representativeness. However, such models often incur substantial computational costs, particularly when fine-tuning large-scale pre-trained language models. We also observe that the fine-tuning process can negatively impact fairness, producing certain perspective models that are underrepresented and have limited influence on the outcome. To address these, we explore two complementary strategies: (i) the adoption of traditional machine learning algorithms—such as Support Vector Machines, Random Forests, and XGBoost—as lightweight alternatives; and (ii) the application of calibration techniques to reduce imbalances in inference generation across perspectives. Our results demonstrate up to 12× faster processing with no statistically significant drop in accuracy. Notably, calibration significantly enhances fairness, reducing inter-group bias and leading to more balanced predictions across diverse social perspectives.

We seek to explain the causes of the misclassification of the most challenging documents, namely those that no classifier using state-of-the-art, very semantically-separable contextual embedding representations managed to predict accurately. To do so, we propose a taxonomy of incorrect predictions, which we used to perform qualitative human evaluation. We posed two (research) questions, considering three sentiment datasets in two different domains – movie and product reviews. Evaluators with two different backgrounds evaluated documents by comparing the predominant sentiment assigned by the model to the label in the gold dataset in order to decide on a likely misclassification reason. Based on a high inter-evaluator agreement (81.7%), we observed significant differences between the product and movie review domains, such as the prevalence of ambivalence in product reviews and sarcasm in movie reviews. Our analysis also revealed an unexpectedly high rate of incorrect labeling in the gold dataset (up to 33%) and a significant amount of incorrect prediction by the model due to a series of linguistic phenomena (including amplified words, contrastive markers, comparative sentences, and references to world knowledge). Overall, our taxonomy and methodology allow us to explain between 80%-85% of the errors with high confidence (agreement) – enabling us to point out where future efforts to improve models should be concentrated.

This study describes the development of a Portuguese Community-Question Answering benchmark in the domain of Diabetes Mellitus using a Recognizing Question Entailment (RQE) approach. Given a premise question, RQE aims to retrieve semantically similar, already answered, archived questions. We build a new Portuguese benchmark corpus with 785 pairs between premise questions and archived answered questions marked with relevance judgments by medical experts. Based on the benchmark corpus, we leveraged and evaluated several RQE approaches ranging from traditional information retrieval methods to novel large pre-trained language models and ensemble techniques using learn-to-rank approaches. Our experimental results show that a supervised transformer-based method trained with multiple languages and for multiple tasks (MUSE) outperforms the alternatives. Our results also show that ensembles of methods (stacking) as well as a traditional (light) information retrieval method (BM25) can produce competitive results. Finally, among the tested strategies, those that exploit only the question (not the answer), provide the best effectiveness-efficiency trade-off. Code is publicly available.

In this paper, we describe our participation in two tasks organized by WOSP 2020, consisting of classifying the context of a citation (e.g., background, motivational, extension) and whether a citation is influential in the work (or not). Classifying the context of an article citation or its influence/importance in an automated way presents a challenge for machine learning algorithms due to the shortage of information and inherently ambiguity of the task. Its solution, on the other hand, may allow enhanced bibliometric studies. Several text representations have already been proposed in the literature, but their combination has been underexploited in the two tasks described above. Our solution relies exactly on combining different, potentially complementary, text representations in order to enhance the final obtained results. We evaluate the combination of various strategies for text representation, achieving the best results with a combination of TF-IDF (capturing statistical information), LDA (capturing topical information) and Glove word embeddings (capturing contextual information) for the task of classifying the context of the citation. Our solution ranked first in the task of classifying the citation context and third in classifying its influence.