Page 297 - ISC PROCEEDINGS 21.4

2.3. Sentiment analysis in technology product evaluation
Computational sentiment analysis has emerged as a powerful methodology for
extracting consumer insights from unstructured text data. Unlike traditional survey
approaches that impose predetermined categories on respondents, sentiment analysis of
user reviews captures organic, unprompted feedback reflecting actual usage experiences.
Recent advances in natural language processing enable granular feature-level sentiment
extraction, allowing researchers to map satisfaction across specific product attributes
rather than relying on aggregate ratings alone.
Machine learning approaches to sentiment classification have demonstrated high
accuracy in distinguishing positive, negative, and neutral expressions, with Python
libraries such as NLTK, TextBlob, and VADER providing accessible implementations for
researchers. The application of these techniques to technology products enables
identification of pain points and satisfaction drivers that may not emerge through
structured research methods. Visualization techniques including word clouds, n-gram
frequency analysis, and temporal sentiment tracking transform complex sentiment data
into actionable insights for product development teams.
3. Research methodology
3.1. Data collection
User reviews of Meta Glasses were collected from major e-commerce platforms and
technology review websites between March 2024 and January 2026. The dataset
comprises comprehensive user feedback including review text, star ratings, review dates,
and verified purchase indicators. To ensure data quality and representativeness, only
English-language reviews from verified purchasers were included in the analysis. The final
dataset contains approximately 10,000 reviews, providing sufficient statistical power for
sentiment pattern identification across temporal and demographic segments.
3.2. Text preprocessing and sentiment analysis
The sentiment analysis pipeline implemented multiple preprocessing stages to
ensure data quality and analytical accuracy. First, review text underwent cleaning
procedures including removal of HTML tags, special characters, and non-alphabetic
symbols while preserving essential punctuation for sentiment context. Text normalization
was applied through lowercasing and tokenization, converting review strings into
analyzable word sequences.
Sentiment classification employed the VADER (Valence Aware Dictionary and
Sentiment Reasoner) algorithm, specifically designed for social media and product review
analysis. VADER provides polarity scores ranging from -1 (most negative) to +1 (most
positive), with additional compound scores enabling three-class classification into positive,
neutral, and negative sentiment categories. VADER was selected over alternative models
(e.g., TextBlob or basic Transformer models) because of its lightweight computational
footprint, lack of training data requirements, and superior out-of-the-box accuracy when
handling informal language, emoticons, and slang heavily present in consumer reviews.
The algorithm demonstrates particular strength in handling informal language, emoticons,
and emphasis markers common in user reviews.
Feature-specific sentiment extraction utilized keyword matching and context
window analysis to associate sentiment scores with product attributes. Keywords related
to battery, audio, camera, connectivity, comfort, and AI features were identified through
frequency analysis and domain knowledge, with sentiment scores calculated for text

296