10th International Congress on Information and Communication Technology

Authors - Marisol Roldan-Palacios, Aurelio Lopez-Lopez
Abstract - Limited available data becomes a critical problem in specific machine learning tasks, where approaches, such as large language models, turn impractical. Reaching solutions in such situations requires alternative methods, especially whether the object of study contributes to data scarcity while preventing using techniques such as data augmentation. This scenario led us to formulate the research question on how to squeeze hidden information from small data. In this work, we propose a data processing and evaluation technique to increase information extraction from scarce data. Attributes expressed as trajectories are further pair-related by proximity and assessed by customary learning algorithms. The efficacy of the proposed approach is tested and validated in language samples from individuals affected by a brain injury. Direct classification on raw and normalized data from three sets of lexical attributes works as a baseline. Here, we report two learning algorithms out of five explored, showing consistent behavior and demonstrating satisfactory discriminatory capabilities of the approach in most cases, with encouraging percentages of improvement in terms of f1-measure. We are in the process of testing the approach in language data sets of syntactic and fluency features, but other fields can take advantage of the technique.