Abstract
Manufacturer-provided age recommendations for children's products are often missing, inaccurate, or overly optimistic. This paper introduces a text-mining framework that extracts real-world age appropriateness from Amazon reviews. By analyzing how parents actually use products, the system provides more accurate age recommendations than official specifications, achieving a median error of three months or less.
Introduction
When shopping for children online, parents struggle to judge if a product is truly suitable for their child's developmental stage. Manufacturer-provided age ranges frequently fail to reflect reality—for example, a toy labeled "0-2 years" might be too complex for an infant or quickly outgrown. While customer reviews are a goldmine of "ground truth" data, manually reading hundreds of reviews to find mentions of a child's age is time-consuming. This research aims to automate this process by extracting age-related features from text to create a "recommended age" profile based on actual user experience.
Methodology
The researchers developed a six-phase NLP pipeline to process unstructured review data from the Amazon "Baby" category:
- Term Extraction: Identifying concepts and quantities.
- Review Selection: Filtering for reviews discussing age or weight appropriateness.
- Entity Value Normalization: Converting messy text (e.g., "3 yrs") into a standardized numerical format.
- Relevance Classification: Using a Naive Bayes classifier to distinguish between age mentions (e.g., "my 5-month-old") and irrelevant time mentions (e.g., "it broke after 1 month").
- Outlier Detection: Applying Tukey’s range test to remove extreme data points.
- Sentiment Analysis: Distinguishing between the sentiment of the entire review versus the specific sentence mentioning the age.
The study compared four estimation strategies: Accumulation-based, Rating-based, Overall Sentiment, and Term Sentiment.
Manufacturer vs. Reality
A significant contribution of this work is highlighting the gap between official marketing and actual usage. The data revealed that manufacturer specs often do not align with how parents use products in the real world.
As seen in the table above, products like "close and secure" sleepers are marketed for up to 2 years, but review data shows parents stop using them around 6 months—a finding that aligns better with pediatric health recommendations than the manufacturer's label.
Key Results and Evaluation
The system was evaluated using a dataset of 45,263 reviews for 6,154 products, with a manually annotated "ground truth" set of 200 products for accuracy testing.
The median error for all estimation strategies was remarkably low—3 months or less. With an error tolerance of just one age division, the system achieved 54% accuracy for estimating the minimum age. Interestingly, strategies relying strictly on sentiment analysis performed slightly worse than simple accumulation, likely because the relevance classifier already effectively filters out negative or irrelevant experiences.
Conclusion
This framework demonstrates that crowdsourced wisdom from reviews can effectively supplement or replace missing manufacturer data. In user validation studies, 72% of participants agreed that the resulting age-distribution visualizations belonged directly on product pages. Future work could expand this method to other categories like toys and media, or extract other "hidden" product dimensions such as durability and material safety.
Summary updated