Augmenting Semantic Representation of Depressive Language: From Forums to Microblogs

We discuss and analyze the process of creating word embedding feature representations specifically designed for a learning task when annotated data is scarce, like depressive language detection from Tweets. We start from rich word embedding pre-trained from a general dataset, then enhance it with embedding learned from a domain specific but relatively much smaller dataset. Our strengthened representation portrays better the domain of depression we are interested in as it combines the semantics learned from the specific domain and word coverage from the general language. We present a comparative analyses of our word embedding representations with a simple bag-of-words model, a well known sentiment lexicon, a psycholinguistic lexicon, and a general pre-trained word embedding, based on their efficacy in accurately identifying depressive Tweets. We show that our representations achieve a significantly better F1 score than the others when applied to a high quality dataset.