L’émergence de modèles de langage très puissants tels que GPT-3 a sensibilisé les chercheurs à la problématique de la détection de textes académiques générés automatiquement, principalement dans un souci de prévention de plagiat. Plusieurs études ont montré que les modèles de détection actuels ont une précision élevée, en donnant l’impression que la tâche soit résolue. Cependant, nous avons observé que les ensembles de données utilisés pour ces expériences contiennent des textes générés automatiquement à partir de modèles pré-entraînés. Une utilisation plus réaliste des modèles de langage consisterait à effectuer un fine-tuning sur un texte écrit par un humain pour compléter les parties manquantes. Ainsi, nous avons constitué un corpus de textes générés de manière plus réaliste et mené des expériences avec plusieurs modèles de classification. Nos résultats montrent que lorsque les ensembles de données sont générés de manière réaliste pour simuler l’utilisation de modèles de langage par les chercheurs, la détection de ces textes devient une tâche assez difficile.

Thanks to the state-of-the-art Large Language Models (LLMs), language generation has reached outstanding levels. These models are capable of generating high quality content, thus making it a challenging task to detect generated text from human-written content. Despite the advantages provided by Natural Language Generation, the inability to distinguish automatically generated text can raise ethical concerns in terms of authenticity. Consequently, it is important to design and develop methodologies to detect artificial content. In our work, we present some classification models constructed by ensembling transformer models such as Sci-BERT, DeBERTa and XLNet, with Convolutional Neural Networks (CNNs). Our experiments demonstrate that the considered ensemble architectures surpass the performance of the individual transformer models for classification. Furthermore, the proposed SciBERT-CNN ensemble model produced an F1-score of 98.36% on the ALTA shared task 2023 data.

Automatic text generation based on neural language models has achieved performance levels that make the generated text almost indistinguishable from those written by humans. Despite the value that text generation can have in various applications, it can also be employed for malicious tasks. The diffusion of such practices represent a threat to the quality of academic publishing. To address these problems, we propose in this paper two datasets comprised of artificially generated research content: a completely synthetic dataset and a partial text substitution dataset. In the first case, the content is completely generated by the GPT-2 model after a short prompt extracted from original papers. The partial or hybrid dataset is created by replacing several sentences of abstracts with sentences that are generated by the Arxiv-NLP model. We evaluate the quality of the datasets comparing the generated texts to aligned original texts using fluency metrics such as BLEU and ROUGE. The more natural the artificial texts seem, the more difficult they are to detect and the better is the benchmark. We also evaluate the difficulty of the task of distinguishing original from generated text by using state-of-the-art classification models.

A Mathematical Word Problem (MWP) differs from a general textual representation due to the fact that it is comprised of numerical quantities and units, in addition to text. Therefore, MWP generation should be carefully handled. When it comes to multi-lingual MWP generation, language specific morphological and syntactic features become additional constraints. Standard template-based MWP generation techniques are incapable of identifying these language specific constraints, particularly in morphologically rich yet low resource languages such as Sinhala and Tamil. This paper presents the use of a Long Short Term Memory (LSTM) network that is capable of generating elementary level MWPs, while satisfying the aforementioned constraints. Our approach feeds a combination of character embeddings, word embeddings, and Part of Speech (POS) tag embeddings to the LSTM, in which attention is provided for numerical values and units. We trained our model for three languages, English, Sinhala and Tamil using separate MWP datasets. Irrespective of the language and the type of the MWP, our model could generate accurate single sentenced and multi sentenced problems. Accuracy reported in terms of average BLEU score for English, Sinhala and Tamil languages were 22.97%, 24.49% and 20.74%, respectively.