Abstract
Public health surveillance systems leveraging social media data often struggle with the figurative use of medical terms, leading to false positives in disease tracking. This research proposes integrating a figurative language detection module into a neural network to enhance the accuracy of identifying real Personal Health Mentions (PHMs) on Twitter. By comparing a rigid pipeline filter against a soft feature augmentation approach, we demonstrate that the latter improves the F-score by an average of 2.21%.
Introduction
Epidemic intelligence relies on monitoring "Personal Health Mentions" (PHMs)—reports where a user or someone they know experiences a medical symptom. However, symptom words are frequently used figuratively (e.g., "When Paris sneezes, Europe catches cold"). Traditional models often misclassify these as literal health reports, creating significant noise in public health data. This study explores whether explicitly modeling figurative versus literal usage can improve the precision of public health tracking.
Methods
The researchers combined an unsupervised figurative usage detection module with a Convolutional Neural Network (CNN) for PHM detection. The figurative detector calculates a "literal usage score" by comparing tweet content to word2vec-based literal representations of symptom words and incorporating linguistic features like "domain-relatedness" (the presence of other medical terms).
Two integration strategies were tested:
- Pipeline Approach: A hard filter where tweets labeled as figurative are immediately discarded, bypassing the PHM classifier.
- Feature Augmentation (+FeatAug): A soft signal approach where figurative labels and linguistic features are fed into the CNN's dense layer as additional context, allowing the model to learn the appropriate weight for the figurative signal.
Datasets
The study utilized the PHM2017 Dataset, containing 5,837 tweets across six conditions: Alzheimer’s, heart attack, Parkinson’s, cancer, depression, and stroke. A manual analysis of a subset of tweets containing "cough" or "breath" revealed that 52.75% of mentions were figurative, highlighting the scale of the misclassification problem in raw social media data.
Results and Discussion
Experimental results revealed that the "hard filter" Pipeline approach actually degraded performance, dropping the F-score by 4.78%. This is likely due to the figurative detector's errors being irrecoverable in a serial architecture. In contrast, the Feature Augmentation approach provided the model with useful "hints," allowing it to weigh figurative signals against the broader context of the tweet.
As shown in the results, the +FeatAug method yielded an average 2.21% improvement in F-score. The most significant gains were observed when using GloVe embeddings retrofitted with WordNet. Disease-specific analysis showed improvements for Alzheimer’s (+3.15) and Depression (+0.86), though "Heart Attack" remained challenging due to highly nuanced hyperbolic usage that the unsupervised detector struggled to capture.
Conclusion
Integrating figurative language detection "softly" into deep learning models effectively filters noise in health surveillance. While challenges like sarcasm and indirect references remain, this approach marks a significant step toward more robust social media-based epidemic intelligence. Future work involving supervised figurative datasets for the health domain could further refine these results.
Summary updated