This study explores how optimism and pessimism are expressed in social media by combining psycholinguistic profiling with model interpretability. Using the OPT dataset, we fine-tune a RoBERTa-based classifier and apply LIME to examine both the most confident and the most ambiguous predictions. We analyze the influential tokens driving these decisions and identify lexical patterns linked to affective intensity, certainty, and social orientation. A complementary LIWC-based analysis of ground truth labels reveals systematic differences in emotional tone and cognitive style. PCA projections further show that optimism and pessimism occupy overlapping yet distinguishable regions in psycholinguistic space. Our findings demonstrate the value of linguistic interpretability in understanding dispositional sentiment.

This paper investigates machine translation between two linguistically distant languages, Arabic and Romanian, with a focus on translating from Arabic to Romanian. Dataset cleaning techniques are addressed, offering insights on the impact of translation for a language pair with limited resources. Using publicly available corpora (e.g., OPUS) and manually translated diplomatic texts, filtering methods are applied, such as duplicate removal, embedding similarity analysis (LEALLA), and Large Language Model (LLM)-based validation (Gemini-flash-002). Transformer models are trained and evaluated with diverse preprocessing pipelines that incorporate subword tokenization. Additionally, the performance of a fine-tuned LLM is assessed for this task and is compared to their pre-trained counterparts. Despite computational limitations, the results emphasize the importance of targeted preprocessing and model adaptation in improving Arabic-Romanian translation quality.

The following paper is a joint contribution for the 2025 ReproNLP shared task, part of the ReproHum project. We focused on reproducing the human evaluation based on one criterion, namely, factuality of Scientific Automated Generated Systems from August et al. (2022). In accordance to the ReproHum guidelines, we followed the original study as closely as possible, with two human raters who coded 300 ratings each. Moreover, we had an additional study on two subsets of the dataset based on domain (medicine and physics) in which we employed expert annotators. Our reproduction of the factuality assessment found similar overall rates of factual inaccuracies across models. However, variability and weak agreement with the original model rankings suggest challenges in reliably reproducing results, especially in such cases when results are close.