Understanding the complex rules and mechanisms behind viral evolution is crucial for developing better preventive treatments, yet predicting immune-evading mutations remains challenging. Recent advances in protein language models have led to novel approaches for in silico analysis of viral escape. We introduce CoV-SNN, unifying variant classification and escape prediction within an efficient contrastive learning framework. CoV-SNN, built on a Siamese neural network architecture, classifies previously unseen variants by modeling sequence-level similarities and differences through embeddings from CoV-RoBERTa, a lightweight protein language model trained on high-quality SARS-CoV-2 Spike sequences. It prioritizes escape sequences using an enhanced Constrained Semantic Change Search (CSCS) function that maps antigenic variation to semantic change and viral fitness to sequence probability. We evaluate CoV-SNN on novel sequences containing both wet-lab-verified and computationally generated escape mutations. CoV-SNN achieves 98.8% accuracy in multi-class variant classification, an AUC of 0.909 in zero-shot variant classification, and 97.7% escape precision at Top-10%, while providing up to a 125-fold inference speedup. These results suggest that contrastive learning can support scalable in silico surveillance and prioritization of immune-evasive mutations.

Local citation recommendation (LCR) suggests a set of papers for a citation placeholder within a given context.This paper introduces Cite-BART, citation-specific pre-training within an encoder-decoder architecture, where authordate citation tokens are masked to learn to reconstruct them to fulfill LCR.The global version (CiteBART-Global) extends the local context with the citing paper's title and abstract to enrich the learning signal.CiteBART-Global achieves state-of-the-art performance on LCR benchmarks except for the FullTextPeerRead dataset, which is quite small to see the advantage of generative pre-training.The effect is significant in the larger benchmarks, e.g., Refseer and ArXiv., with the Refseer pre-trained model emerging as the best-performing model.We perform comprehensive experiments, including an ablation study, a qualitative analysis, and a taxonomy of hallucinations with detailed statistics.Our analyses confirm that CiteBART-Global has a cross-dataset generalization capability; the macro hallucination rate (MaHR) at the top-3 predictions is 4%, and when the ground-truth is in the top-k prediction list, the hallucination tendency in the other predictions drops significantly.We publicly share our code 1 , base datasets 2 , global datasets 3 , and pre-trained models 4 to support reproducibility.

Counterfactual statements are examples of causal reasoning as they describe events that did not happen and, optionally, those events’ consequences if they happened. SemEval-2020 introduces the counterfactual detection (CFD) task and shares an English dataset. Since then, a set of datasets has been released in English, German, and Japanese as part of Amazon product reviews. This work releases the first Turkish corpus of counterfactuals (TRCD). The data collection process is driven by a clue phrase list of counterfactuals, mainly in the form of verb inflections in Turkish. We use clue phrase-based filtering to collect sentences from the Turkish National Corpus (TNC). On the other hand, half of the collection is subject to random word filtering to avoid selection bias due to clue phrases. After the human annotation process with an Inter Annotator Agreement of 0.65, we have 5000 sentences, of which 12.8% contain counterfactual statements. Furthermore, we provide a comprehensive baseline of transformer-based models by testing the effect of clue phrases, cross-lingual performance comparisons using the available CFD datasets, and zero-shot cross-lingual classification experiments using fine-tuning on the different combinations of the existing datasets. The results confirm that TRCD is compatible with the other CFD datasets. Moreover, fine-tuning a Turkish-specific model (BERTurk) performs better than the multilingual alternatives (mBERT and XLM-R). BERTurk is more robust to clue phrase masking. This result emphasizes the importance of a language-specific tokenizer for contextual understanding, especially for low-resource languages. Finally, our qualitative analysis gives insights into errors by different models.

In the field of software engineering, practitioners’ share in the constructed knowledge cannot be underestimated and is mostly in the form of grey literature (GL). GL is a valuable resource though it is subjective and lacks an objective quality assurance methodology. In this paper, a quality assessment scheme is proposed for question and answer (Q&A) sites. In particular, we target stack overflow (SO) and stack exchange (SE) sites. We model the problem of author reputation measurement as a classification task on the author-provided answers. The authors’ mean, median, and total answer scores are used as inputs for class labeling. State-of-the-art language models (BERT and DistilBERT) with a softmax layer on top are utilized as classifiers and compared to SVM and random baselines. Our best model achieves [Formula: see text] accuracy in binary classification in SO design patterns tag and [Formula: see text] accuracy in SE software engineering category. Superior performance in SE software engineering can be explained by its larger dataset size. In addition to quantitative evaluation, we provide qualitative evidence, which supports that the system’s predicted reputation labels match the quality of provided answers.

While traditional distance measures are often capable of properly describing similarity between objects, in some application areas there is still potential to fine-tune these measures with additional information provided in the data sets. In this work we combine such traditional distance measures for document analysis with link information between documents to improve clustering results. In particular, we test the effectiveness of geodesic distances as similarity measures under the space assumption of spherical geometry in a 0-sphere. Our proposed distance measure is thus a combination of the cosine distance of the term-document matrix and some curvature values in the geodesic distance formula. To estimate these curvature values, we calculate clustering coefficient values for every document from the link graph of the data set and increase their distinctiveness by means of a heuristic as these clustering coefficient values are rough estimates of the curvatures. To evaluate our work, we perform clustering tests with the k-means algorithm on the English Wikipedia hyperlinked data set with both traditional cosine distance and our proposed geodesic distance. The effectiveness of our approach is measured by computing micro-precision values of the clusters based on the provided categorical information of each article.