Michał Czuba

Several studies in the sociology of immigration have focussed on informal networks as the primary source of migrants’ social capital. However, the literature has largely eschewed the potential afforded by the computational analytical tools of network science that permit the reconstruction and mapping of community sociograms and the calculation of the impact of centrality metrics. In addition, studies are still scarce that would combine network-analytic approaches with rigorous investigations of refugees’ acquisition of host-country language skills, despite the proven import thereof for functioning in the new destination. We analyse the peer interaction networks of 251 Ukrainian refugees participating in an intensive Polish language course. Employing a custom-designed name-interpreter survey in conjunction with a pre-/post-test design, we leverage computational social network analysis to i) identify patterns of participants’ informal communication beyond the classroom, with particular attention to interactions within their co-national group, and ii) examine how these patterns, alongside individual background characteristics, affect their language development. Speaking Ukrainian correlated with greater centrality in the contact network. Russian-dominant speakers often concealed their use of this language, possibly because they were frequently found at the network periphery. We discuss the affective, motivational, and interactive factors identified as predictors of progress in Polish.

Using computational Social Network Analysis (SNA), this longitudinal study investigates the development of the interaction network and its influence on the second language (L2) gains of a complete cohort of 41 U.S. sojourners enrolled in a 3‐month intensive study‐abroad Arabic program in Jordan. Unlike extant research, our study focuses on students’ interactions with alma mater classmates, reconstructing their complete network, tracing the impact of individual students’ positions in the social graph using centrality metrics, and incorporating a developmental perspective with three measurement points. Objective proficiency gains were influenced by predeparture proficiency (negatively), multilingualism, perceived integration of the peer learner group (negatively), and the number of fellow learners speaking to the student. Analyses reveal relatively stable same‐gender cliques, but with changes in the patterns and strength of interaction. We also discuss interesting divergent trajectories of centrality metrics, L2 use, and progress; predictors of self‐perceived progress across skills; and the interplay of context and gender.

Lameness in dairy cows poses a significant challenge to improving animal well-being and optimizing economic efficiency in the dairy industry. To address this, employing automated animal surveillance for early lameness detection and prevention through activity sensors proves to be a promising strategy. In this study, we analyzed activity (accelerometer) data and additional cow-individual and farm-related data from a longitudinal study involving 4860 Holstein dairy cows on six farms in Germany during 2015–2016. We designed and investigated various statistical models and chose a logistic regression model with mixed effects capable of detecting lameness with a sensitivity of 77%. Our results demonstrate the potential of automated animal surveillance and hold the promise of significantly improving lameness detection approaches in dairy livestock.

With a recent influx of voice generation methods, the threat introduced by audio DeepFake (DF) is ever-increasing. Several different detection methods have been presented as a countermeasure. Many methods are based on so-called front-ends, which, by transforming the raw audio, emphasize features crucial for assessing the genuineness of the audio sample. Our contribution contains investigating the influence of the state-of-the-art Whisper automatic speech recognition model as a DF detection front-end. We compare various combinations of Whisper and well-established front-ends by training 3 detection models (LCNN, SpecRNet, and MesoNet) on a widely used ASVspoof 2021 DF dataset and later evaluating them on the DF In-The-Wild dataset. We show that using Whisper-based features improves the detection for each model and outperforms recent results on the In-The-Wild dataset by reducing Equal Error Rate by 21%.