Brand Hate Detector: An R Shiny Application for Automated Detection and Multilevel Classification of Brand Hate in Consumer Reviews

1. Data Acquisition Module

The workflow begins with Brand Information. Here, the user provides a ConsumerAffairs brand URL, which the application uses to retrieve and display key metadata—such as the brand’s name, logo, and total number of available reviews. This initial step ensures the target context is correct before proceeding. Once confirmed, the process transitions seamlessly into Scraping Reviews. Leveraging the rvest package (Wickham & Bryan, 2023), the application automatically navigates paginated review pages, extracting the full review text, optional “reasons” metadata, and basic sentiment cues. The collected data are immediately presented in an interactive DT table, allowing users to inspect, search, and filter entries. For transparency and reproducibility, the raw review set can be exported at this stage as a CSV file for independent analysis.

2. Preprocessing Module

Following data acquisition, the workflow advances to Negative Review Filtering. In this step, each review is scored for sentiment polarity using the sentimentr package (Rinker, 2017). Only those with a negative polarity score are retained, narrowing the dataset to content most likely to contain brand hate signals in line with theoretical assumptions. With the filtered set-in place, the process transitions to Stopword Removal. Using the tidytext and stopwords packages (Wickham & Bryan, 2023), the retained reviews are tokenized and stripped of common stopwords, punctuation, and numerical characters. Both the cleaned text and its original counterpart are preserved, ensuring transparency and allowing for full traceability in subsequent analysis.

3. Emotion Profiling Module

Once the textual data are cleaned, the workflow moves into NRC Emotion Analysis. At this stage, tokens are matched to the NRC Emotion Lexicon (Mohammad & Turney, 2013), assigning each word to one or more of eight primary emotions—anger, fear, anticipation, trust, surprise, sadness, joy, and disgust—along with positive or negative polarity. For the purposes of brand hate detection, emphasis is placed on negative emotions most strongly associated with hostility, including anger, sadness, fear, disgust, and surprise (Kucuk, 2019; Sternberg & Sternberg, 2008; Zhang & Laroche, 2020). Counts are aggregated at the review level, producing a structured emotional profile for each entry. Building on these results, the process transitions into Emotion Visualization. Using ggplot2, the application generates interactive bar charts (Wickham & Bryan, 2023) that summarize the overall distribution of emotions within the dataset. These visualizations provide a high-level emotional “fingerprint” of the brand’s negative reviews, setting the stage for classification into distinct hate intensity levels.

4. Brand Hate Classification Module

With the emotional profiles established, the workflow advances to the Initial Rule-Based Classification stage. Here, the application combines aggregated sentiment statistics—mean and standard deviation of polarity scores—with the presence of specific negative emotion patterns to assign each review to one of three categories: Mild Hate, Moderate Hate, or Strong Hate. Reviews that do not meet the criteria for any of these categories are temporarily labeled as Unclassified. This step ensures that only reviews with clear emotional or sentiment-based signals are assigned an intensity level at this stage.

From this point, the process transitions to a Reclassification Step, where unclassified reviews are re-evaluated using a purely emotion-based approach. This secondary pass identifies complex cases in which multiple high-intensity emotions co-occur, even if sentiment thresholds were not met. Such cases are labeled as Hybrid Hate, reflecting the combined presence of diverse emotional triggers. This two-tiered approach ensures a comprehensive categorization of all reviews, capturing both direct hostility and more nuanced emotional blends.

5. Visualization & Reporting Module

With classification complete, the workflow transitions to the Visualization & Reporting stage, where results are synthesized into interactive, exportable outputs. The application provides multiple perspectives on the data to support both academic analysis and managerial decision-making.

The Sentiment Distribution plot shows the proportion of negative reviews relative to the total dataset, providing a quick gauge of overall consumer dissatisfaction. The Hate Category Distribution chart displays the relative prevalence of mild, moderate, strong, and hybrid hate, enabling easy comparison of intensity levels.

To explore underlying drivers, the Reasons Analysis module aggregates the “reasons” metadata—such as customer service, product quality, or pricing—by hate category. This view highlights which aspects of the brand are most associated with different levels of hostility.

Lexical patterns are further illustrated in the Emotion Word Clouds, where words are color-coded by emotion category. These clouds provide a visually intuitive summary of the most salient terms within each hate level.

Finally, the Hate Intensity Bubble Plot maps sentiment scores against hate categories, revealing the relationship between emotional diversity and perceived hostility. Outliers and clusters are easily spotted, aiding in the identification of unusual or particularly impactful reviews.

All visualizations can be downloaded in publication-quality PNG or TIFF formats, and all processed data tables can be exported as CSV files for further offline analysis or integration into other research workflows.