Neural machine translation (NMT) is a software that uses neural network techniques to translate text from one language to another. However, one of the most famous NMT models—Google Translate—failed to give an accurate English translation of a famous Korean nursery rhyme, "Airplane" (비행기). The authors fine-tuned a pre-trained model first with a dataset from the lyrics domain, and then with a smaller dataset containing the rhythmical properties, to teach the model to translate rhythmically accurate lyrics. This stacked fine-tuning method resulted in an NMT model that could maintain the rhythmical characteristics of lyrics during translation while single fine-tuned models failed to do so.
Catalase is a biocatalyst used to break down toxic hydrogen peroxide into water and oxygen in industries such as cheese and textiles. Improving the efficiency of catalase would help us to make some industrial products, such as cheese, less expensively. The best way to maintain catalase’s conformation, and thus enhance its activity, is to immobilize it. The primary goal of this study was to find a new way of immobilizing catalase.
Here, the authors wanted to explore mathematical paradoxes in which there are multiple contradictory interpretations or analyses for a problem. They used ChatGPT to generate a novel dataset of fairy tales. They found statistical differences between the artificially generated text and human produced text based on the distribution of parts of speech elements.
The authors looked at using machine learning to identify skills needed to apply for certain jobs, specifically looking at different techniques to parse apart the text. They found that Bidirectional Encoder Representation of Transforms (BERT) performed best.
Although there has been great progress in the field of Natural language processing (NLP) over the last few years, particularly with the development of attention-based models, less research has contributed towards modeling keystroke log data. State of the art methods handle textual data directly and while this has produced excellent results, the time complexity and resource usage are quite high for such methods. Additionally, these methods fail to incorporate the actual writing process when assessing text and instead solely focus on the content. Therefore, we proposed a framework for modeling textual data using keystroke-based features. Such methods pay attention to how a document or response was written, rather than the final text that was produced. These features are vastly different from the kind of features extracted from raw text but reveal information that is otherwise hidden. We hypothesized that pairing efficient machine learning techniques with keystroke log information should produce results comparable to transformer techniques, models which pay more or less attention to the different components of a text sequence in a far quicker time. Transformer-based methods dominate the field of NLP currently due to the strong understanding they display of natural language. We showed that models trained on keystroke log data are capable of effectively evaluating the quality of writing and do it in a significantly shorter amount of time compared to traditional methods. This is significant as it provides a necessary fast and cheap alternative to increasingly larger and slower LLMs.
Here, the authors investigated the effects of an interventional psychology on the study habits of high school students specifically related to the use of electronic distractions such as social media or texting, listening to music, or watching TV. They reported varying degrees of success between the control and intervention groups, suggesting that the methods of habit-breaking for students merits further study.
Researchers query whether reading comprehension is the same, worse, or better when using e-books as compared with standard paper texts. This study evaluated this question in the elementary school population. Our hypothesis was that information would be retained equally whether read from paper or from an electronic device. Each participant read four stories, alternating between electronic and paper media types. After each reading, the participants completed a five-question test covering the information read. The study participants correctly answered 167 out of 200 comprehension questions when reading from an electronic device. These same participants correctly answered 145 out of 200 comprehension questions when reading from paper. At a significance level of p < 0.05, the results showed that there was a statistically significant difference in reading comprehension between the two media, demonstrating better comprehension when using electronic media. The unexpected results of this study demonstrate a shift in children’s performance and desirability of using electronic media as a reading source.
Recent advances in generative AI have made it increasingly hard to distinguish real images from AI-generated ones. Traditional detection models using CNNs or U-net architectures lack precision because they overlook key spatial and frequency domain details. This study introduced a hybrid model combining Convolutional Neural Networks (CNN) with Fast Fourier Transform (FFT) to better capture subtle edge and texture patterns.
The authors use MoralBERT, a BERT-based machine learning model, to analyze the prevalence and evolution of moral language in art critique journals from 1960 to 2015. They draw on Moral Foundations Theory to examine changes in five virtue and five vice foundations over time. They discover slight increases in degradation and care and a decrease in purity among otherwise stable patterns. The study is a novel application of computational moral analysis in a domain not typically associated with such tools. Although the moral signals were relatively faint, the study offers insights into shifting art world discourses and models an exciting method for future inquiry.
