Master of Science in Applied Information and Data Science
Lucerne University of Applied Sciences and Arts
This dissertation evaluates the effectiveness of different vector embedding techniques for multilabel text classification. The primary goal is to understand which embedding spaces are more effective when applied to real world data.
This research uses data collected by IMPACT Initiatives, an organisation that gathers information on the needs and displacement patterns of crisis-affected populations. IMPACT’s data, primarily collected through discussion groups and interviews, is analysed to identify key talking points. These talking points form the foundation of IMPACT’s data products, which support humanitarian organizations in making informed crisis response decisions.
A range of models were used to generate embeddings, including Word2Vec, BERT, Sentence-BERT (sBERT), and Llama 3.1. Each embedding space was tested using three traditional classifiers - Logistic Regression, Random Forest, and Support Vector Classification.
Of all the embedding–classifier combinations evaluated, a fine-tuned sBERT model paired with a Random Forest classifier achieved the highest classification performance. This combination is also capable of running efficiently on limited hardware, such as an IMPACT laptop in the field.
Overall, the findings suggest that lightweight, task-specific embedding models—when paired with lightweight classifiers—can deliver strong performance, even in constrained computing environments.
Keywords: vector embeddings, multilabel classification, text classification, fine-tuning, Sentence-BERT