A0475
Title: Monitoring (social) media narratives combining retrospective few-shot classification with continuous topic modeling
Authors: Jonas Rieger - TU Dortmund University (Germany) [presenting]
Abstract: A variety of topic models and text classification models exist, each with distinct characteristics. However, in truly interdisciplinary applications, many of such models prove unsuitable due to specific individual limitations such as continuously growing text corpora or (huge) unbalancedness of classification labels. These issues are addressed by meshing suitable supervised and un-/semi-supervised learning approaches to efficiently visualize thematic trends in texts in real-time using dynamic topic models while enabling diachronic quantification of argumentative shifts and their uncertainties using static few-shot learning techniques to enable parameter-efficient fine-tuning of transformer-based pre-trained language models. This hybrid approach facilitates rapid insights into thematic trends and allows for retrospective quantification with the need for only small-scale manual coding experiments. Presenting different performance and reliability measures, straightforward usability and state-of-the-art performances are demonstrated for argumentative example datasets covering different thematic areas.
