Using data from social media websites to inspire the design of assistive technology
Xing Yu · 2016 · Proceedings of the 13th International Web for All Conference (W4A) · doi:10.1145/2899475.2899501
Summary
This doctoral consortium paper proposes using social media data as a low-cost, scalable method to inform the design of assistive technology, addressing limitations of traditional user research approaches. The author argues that designing assistive technology faces unique recruitment challenges: target users with specific disabilities may represent a small portion of the population, be geographically dispersed, and be difficult to reach through conventional methods like campus-based studies. These challenges lead to insufficient sample sizes and unrepresentative data. Additionally, traditional qualitative methods (interviews, focus groups, observation) and quantitative methods (questionnaires, experiments) handle relatively small numbers of observations. The proposed solution leverages user-generated content on social media platforms, which can be collected at virtually no cost via APIs, provides far larger datasets than traditional methods, and captures perspectives that users may share more freely due to the anonymous nature of many platforms. The tool combines three computational techniques: sentiment analysis to identify topics that provoke negative emotional responses (which the author links to technology non-adoption), text mining using Latent Dirichlet Allocation (LDA) for topic extraction from large corpora, and machine learning (Random Forests) to identify causal relationships between discussion topics and sentiment.
Key findings
In a feasibility study, the author scraped 858 posts with 45,933 comments about prosthetics from Reddit, spanning February 2008 to November 2015. Using LDA for topic modeling and semantic compositionality-based sentiment analysis, a Random Forest classifier predicted sentiment responses from topic distributions with 93% overall accuracy. By extracting the most important predictors from the classifier, the tool identified six key topics associated with negative sentiment: flights/TSA security (reflecting travel difficulties with prosthetics), games (character representation), government/legal issues, surgery/pain, arms/hands (physical design), and human/body/technology (broader concerns about the relationship between technology and the body). These topics surface real user concerns that might not emerge in traditional design research—for instance, the prominence of airport security as a pain point for prosthetics users. The research was motivated by findings that 20% of prosthetics users abandon their devices due to appearance, functionality, or inconvenience issues, highlighting the need for better problem-space definition early in the design process.
Relevance
This work proposes a novel complement to traditional accessibility user research, particularly valuable for under-resourced designers and researchers working with hard-to-reach populations. The approach of mining social media for unsolicited user feedback on assistive technology could help surface concerns that people might not voice in formal research settings. For accessibility practitioners, the methodology offers a way to understand large-scale user sentiment at the early problem-definition stage of design, potentially reducing assistive technology abandonment rates. However, social media data has inherent biases—users of platforms like Reddit skew younger and more tech-savvy, and the anonymous nature of posts makes it difficult to verify disability status or demographics. The paper is at an early doctoral stage, and the tool had not yet been validated with designers or compared with insights from traditional methods. Nevertheless, the general approach of computational analysis of online disability communities for design insights remains a promising research direction.
Tags: assistive technology · social media · natural language processing · machine learning · prosthetics · user research · sentiment analysis · design methodology