Introduction
Public health officials require granular insights into how different demographic groups perceive vaccines to design effective outreach and education programs. While traditional surveys provide high-quality data, they are often resource-intensive, slow, and difficult to scale in real-time. Social media platforms like Twitter offer a massive alternative for sentiment analysis; however, they lack explicit demographic metadata, such as gender, which is critical for epidemiological research.
This paper addresses the challenge of inferring gender from linguistic patterns on social media and applies these inferences to analyze gender-specific attitudes toward the Human Papillomavirus (HPV) vaccine.
Methodology: CNN with Embedding Fusion
To overcome the limitations of existing gender-inference tools—which often rely on unreliable usernames—the researchers developed a Convolutional Neural Network (CNN) with embedding fusion. This architecture is designed to capture the nuances of digital communication by combining three distinct types of information:
- Character Embeddings: These allow the model to process individual characters, making it robust against typos, slang, and "out-of-vocabulary" terms common in tweets.
- Word Embeddings: The model utilizes pre-trained GloVe vectors (trained on 2 billion tweets) to capture the semantic meaning of words.
- Part-of-Speech (POS) Tags: By identifying grammatical roles (e.g., nouns, verbs, adjectives), the model captures stylistic differences in how different genders structure their sentences.
These inputs are fused into a single word-level layer, processed through convolutional and max-pooling layers, and finalized through a dense layer with batch normalization. To ensure stability and maximize accuracy, the authors employed an ensemble approach, utilizing a majority vote from five model iterations.
Data Sources
The study leveraged two primary datasets:
- PAN 2018 Author Profiling Task: A benchmark dataset (3,000 training users, 1,900 testing users) used to train and validate the gender prediction model.
- HPV Vaccine Corpus: A custom dataset of 1,489,272 English tweets collected between 2014 and 2018. A subset of 6,000 tweets was manually annotated to train a classifier for health-related constructs based on the Health Belief Model (HBM).
Key Results and Insights
Model Performance
The ensemble model (CNN_char_pos) achieved a state-of-the-art accuracy of 0.8237 on the PAN 2018 benchmark, outperforming previous benchmarks. This success suggests that gendered "voice" on social media is encoded not just in what is said (keywords), but in the underlying grammatical and character-level patterns.
Gendered Perceptions of HPV Vaccination
Applying the model to the HPV corpus revealed that approximately 52.5% of the conversation was driven by female users. More importantly, the study identified significant differences in how genders perceive the vaccine:
- The "Barriers" Gap: Men were significantly more likely to discuss barriers to vaccination. As shown in Figure 2, the odds ratio for men discussing barriers remained consistently high (above 1.7) across the five-year study period.
- Benefits and Severity: Conversely, women were more likely to discuss the benefits of the vaccine and the severity of HPV-related health issues.
These findings were statistically significant (p < 0.001) and align with smaller-scale traditional sociological surveys, validating the use of deep learning for large-scale social media mining in public health.
Limitations and Future Work
While promising, the study notes two primary limitations. First, misclassification bias is a risk; because gender labels are predicted, errors in the CNN could skew the resulting health analysis. Second, representativeness remains a challenge, as Twitter users tend to be younger and more tech-savvy than the general population.
Future research will aim to apply these deep learning techniques to other demographic factors, such as age and geographic location, to provide an even more nuanced view of public health sentiments.
Summary updated