Using AI to Predict Public Relations Crises and Recommend Effective Responses (and How CapeStart Built the Solution).
It’s a difficulty that public relations (PR) teams and organizations in general have grappled with since practically the beginning of time – how to predict when a PR crisis will occur. It is also a challenge to determine the best way to manage that crisis so it causes the least possible damage.
Organizations have traditionally reacted to a potential PR crisis only when it’s gained enough steam to get on their radar. They then must decide the best way to respond to the problem, sometimes in a matter of hours, in most cases relying on a combination of experience and gut instinct to determine the appropriate steps. However, it’s often too late to mitigate at least some damage to their organization’s brand by that time.
There have been plenty of attempts in the past to apply automation, AI, and machine learning (ML) to this long-standing problem. But none of them have done an adequate job. Most fully automated crisis media monitoring solutions quickly become inundated with irrelevant results. And AI solutions designed to handle crisis media monitoring and engagement haven’t performed well enough to be commercially viable.
CapeStart’s innovative data scientists and engineers knew there was a better way to use large amounts of historical crisis-related data, combined with machine learning (ML), to predict a fast-moving crisis before it gains momentum.
CapeStart’s AI crisis communications solution
Texas A&M communications professor Timothy Coombs’s well-known situational crisis communication theory (SCCT) was the basis for developing an AI-based crisis media prediction solution. CapeStart developed the solution on behalf of our client, Fullintel, a Cambridge, MA-based media monitoring company specializing in high-touch media monitoring, media analysis, and executive news briefs.
The solution accurately performs several other crucial crisis communications functions and is designed to handle a PR crisis from end-to-end, including crisis discovery, prediction, monitoring, and management. These functions include early detection of crisis signals, diagnosis of impending trouble, recommendations for the right strategy to diffuse the crisis based on data, improved insights through ongoing monitoring of stakeholder commentary, and continuous learning by the AI algorithm to adjust itself based on the facts on the ground.
How does the AI crisis communications solution work?
CapeStart combined 40 years’ worth of media articles on various public crises, sourced from a leading content provider, with the different crisis types identified in SCCT, including:
- The Victim Cluster. All crises beyond the immediate control of the crisis victim, including natural disasters, workplace violence, or product tampering.
- The Accidental Cluster. Crises borne out of an accident, such as technical errors.
- The Intentional Cluster. Crises created by organizational misdeeds, management misconduct, and similar incidents.
CapeStart then combined these crisis types with the various possibilities of crisis response identified in SCCT, which include:
- Deny. Includes attacking the accuser, denying the crisis, or finding a scapegoat.
- Diminish. Includes finding a justification or excuse for the crisis.
- Rebuild. Includes providing compensation to mitigate the crisis.
- Bolster. Includes presenting the company as a victim and ingratiating itself with stakeholders.
Categorizing the various crisis and response types was crucial for training the AI solution the right way. By identifying the correct crisis type, the algorithm can compare the current situation with past crises of the same kind. Various other metrics enable the solution to evaluate the crisis, including severity, audience reaction, brand reputation, predicted sentiment, event impact score, crisis timespan, days to neutral (or the number of days for the crisis to reach a neutral sentiment), and comms team performance.
After combining these various crisis types and responses with the rich collection of articles contained in the articles database, CapeStart was able to create three separate but crucial databases to feed data to its AI algorithms:
- Global Events Database. All significant events referenced within our 40 years of media data, including non-crisis events.
- Media Crisis Database. All PR crisis-related events within the Global Events Database.
- Crisis Ontology. All crisis-related topics, categories, entities in the lexicon, and metadata labels.
These databases feed the 100-plus ML models built by CapeStart – including severity calculator, crisis detector, response recommender, and similarity finder algorithms – which then provide crisis predictions, recommendations, measurement, and crisis management help.
Natural language processing (NLP) techniques that went into the solution include text classification, summarization, topic modeling, and clustering. Algorithms to provide predictions ranged from the cutting-edge Bidirectional Encoder Representations from Transformers (BERT) model to the more conventional support vector machine (SVM) algorithm.
Improving crisis communications through AI
The dashboard-style solution automatically detects potential or emerging crises in real-time, predicting which events will likely turn into full-blown snafus if left unattended. Once an event is identified as a bona fide crisis, the system predicts how many days it will take for public and media sentiment to return to neutral. The system also displays various metrics and predicted metrics based on the recommended response options, to show the expected difference between an aggressive Deny-style response or a more laid-back Rebuild approach (for example).
The AI solution houses a slew of machine learning models that provide insights in real-time by digging into historical data troves to find similar crises. These are then fed to statistical models that make mission-critical suggestions for handling the crisis.
Algorithms and ML models used in the AI solution for crisis communications
The AI solution for crisis communications utilizes several AI algorithms and ML models, including:
- One vs. rest classifier
- Support vector machine (SVM)
- Bidirectional Encoder Representations from Transformers (BERT) for article and post summary generation
- Long short-term memory (LSTM)
- Latent Dirichlet allocation (LDA)
- K-means clustering
- Topic modeling
- Text classification
- Text summarization
- Statistical modeling
CapeStart machine learning engineers and data scientists also used several real-time and occasional prediction models, depending on requirements. These included:
Real-time prediction models
- Crisis or not classifier for traditional media content (print, online news, and broadcast coverage)
- Crisis type classifier for traditional media content
- Article content classifier to map specific media articles to a particular crisis
- Article content classifier to identify and classify events from non-crisis articles
- Crisis or not classifier for social media content
- Social media post content classifier to map posts to a particular crisis
Occasional prediction models
- Crisis similarity finder
- Finds crises similar to the identified ongoing crisis
- Severity predictor
- Predicts the severity of the crisis based on analysis of past crises
- Days to Neutral predictor
- Predicts the number of days for coverage and conversations around the crisis to move from negative to neutral
- Sentiment engine
- Predicts media and public sentiment based on various crisis responses