Topic Extraction from A Cancer Health Forum
Samuel Miles, Lixia Yao, Weilin Meng, Christopher Matthew Black, Zina Ben Miled · 2021
This paper presents a methodology for topic extraction from a corpus of unstructured posts submitted to the online health forum r/Cancer. Topic extraction is important in many fields. It can provide an understanding of how patients and their caregivers manage the disease and related treatments. Reduced vector embeddings are generated for each post using a combination of a pre-trained language model and dimensionality reduction. These embeddings are then clustered with particle swarm optimization (PSO). An embedding size of 300 produced a topic model with a quality of 0.44. This quality level was obtained without biasing the words in the vocabulary towards a specific topic as currently practiced.