A semi-supervised method for topic extraction from micro postings
Social networking services have become a major channel for the digital society to share content, opinions, experiences on activities or events, as well as on products, services and brands. Evaluating digital feedback on the latter can be a valuable asset for companies seeking product and consumer insights. However, the analysis of short, noisy, fragmented, and often subjective textual data still remains a challenge. Typically, the human analyst needs to be actively involved during extraction and modeling to resolve ambiguities that will inevitable arise in such data and to put the model into context. This paper proposes a visual analytics approach that enables a first intuition and exploration of topics appearing in the text corpus, and facilitates the interactive-iterative refinement of the overall topic model describing the stream of tweets. A second contribution is the discussion of efficient graph community detection algorithms to extract initial topics as the starting point of interactive analysis that complement approaches such as LDA. The applicability and utility of the proposed approach is shown for a real-world use case: the analysis of product insights and topic-driven social networks analysis for a specific product line for an international hair styling and cosmetics company.