In this work, we try to bridge the gap between predominantly single-method, single-source scientific studies and real-life brand and product reputation crisis scenarios, emerging and proliferating across various Internet channels. To this aim, we offer our own quantitative analysis of 15 reputation crises, as well as an overview of popular quantitative and qualitative methods of reputation crisis detection; build a dataset of crisis-related emotions and a tagset encompassing both explicit and implicit emotions, contrary to certain scientific works; build a rule-based language to elaborate domain-specific and domain-independent crisis detection models, predefined by brand owners based on domain-specific risks. In reaching our objective, we selected (1) heuristic crisis detection models for product-harm and brand reputation crises, both branch-specific and branch-independent, alongside a metadata-based scoring technique evaluating the company’s crisis vulnerability at a given time, (2) an emotion detection ML-based algorithm using hierarchical neural networks (HAN), (3) a statistical mention peak analysis tool, backed by an ML-based algorithm summarizing the mentions in the corpus. Having devised and tested the aforementioned approaches, we finally analyze their suitability for detecting early signs of crisis in real-life reputation crisis scenarios. We argue that a complete brand image monitoring system should go beyond sentiment or emotion detection and feature a hybrid design. We understand hybrid as covering different types of crises and various internet sources and monitoring possible warning signs in near real-time, adjusted for the specific problems of a given business branch. We contend that a reputation crisis should be defined as one ante- and not post-factum, and thus coin a definition of a reputation crisis based on the linguistic theory of computer-mediated communication as asynchronous online debates. The above allowed us to conclude that if a reputation crisis is a polylogue in linguistics, and its subtopics are based on a schema of affliction, detecting vocabulary related to the crisis schema gives a greater guarantee of crisis detection with its early signs than with emotion- or sentiment-based models. The statistical models detect the growing number of contributions or participants in the discussion; the topic detection models indicate the topics related to the discussion; the anomaly detection models capture their outbursts; the heuristic domain-related models ensure the topics and subtopics of the discussion are crisis-related, and the emotion detection models confirm whether the emotions expressed toward the topics and subtopics are negative. Therefore, heuristic models are good cause-indicators and early crisis detectors, while the statistical and emotion detection-based methods—effect and scale analyzers. Crises differ with respect to their domains or branches. Thus, the evolution of the reputation crisis is a function of three key factors: the domain, the crisis type, and the popularity of the given brand.
