Assessment of Word Embedding Techniques for Identification of Personal Experience Tweets Pertaining to Medication Uses

摘要

Twitter, a general purpose social media service, has seen growing interests as an active data source for possible use of post-market surveillance of medicinal products. Being able to identify Twitter posts of personal experience related to medication use is as important as being able to identify expressions of adverse medical events/reactions for the surveillance purpose. Identifying personal experience tweets is a challenging task, especially in the aspect of engineering features for classification. Word embedding has become a superior alternative to engineered features in many text classification applications. To investigate if word embedding-based methods can perform constantly better than conventional classification methods with engineered features, we assessed the classification performance of 4 word embedding techniques: GloVe, word2vec, fastText, and wordRank. Using a corpus of 22 million unlabeled tweets for learning of word embedding and a corpus of 12,331 annotated tweets for classification, we discovered that word embedding-based classification methods consistently outperform the engineered feature-based classification methods with statistical significance of p < 0.01, but there exist no significantly statistical differences among the 4 study word embedding methods (p < 0.05).