How to select samples for active learning? Document clustering with active learning methodology
Norbert Ropiak, Mateusz Gniewkowski, Michał Swędrowski, Michał Pogoda, Karol Gawron, Bartłomiej Bojanowski, Tomasz Walkowiak · 2023
In this paper, we investigate the applicability of the Active Learning technique to text clustering and topic modeling tasks. The aforementioned problems are often a non-trivial task due to the ambiguity of text similarity meaning. In our experiments, we implemented the Active Learning approach using automatic annotation from datasets with prepared labels. In a simulated study conducted on Polish and English datasets, we show how labeling a relatively small carefully selected number of examples can improve the quality of clustering relative to approaches based on a general notion of text similarity. We compare a number of techniques for selecting samples for labeling, dimensionality reduction and training approaches in order to compare and obtain the best quality of the resulting clusters with a minimum number of annotations. The obtained results show that with a relatively simple approach it is possible to obtain good quality clusters and thus develop classification ontologies in a data-centric approach.