Methods

We will be analyzing comments from Reddit, a popular social media platform. Reddit has a sizable, though not particularly diverse, user base, a significant portion of which is human. Additionally, it has data stretching back to 2007, all of which can be accessed for free. We will request data from 2010 to 2023.

Once we collect our data, we will pick some commonly used mental health related words to study, such as “trauma” and “depression.” Then, we will conduct three different analyses on it.

The first will be a quali-quantitative co-word analysis, which will help us gain an understanding of the semantic change in the words we are studying.

The second will be a quantitative analysis based on Word2vec embeddings of the words we are studying. We will slice up the time range under study. Within each time slice, we will generate Word2vec embeddings of the words under study. These embeddings will be generated in a way that embeddings for the same word from two different time slices can be compared. The comparison will be simple cosine similarity between vectors, and will allow us to quantify semantic change in each word between any two time slices.

The third will be a quali-quantiative analysis based on BERT embeddings of the words we are studying. For each of the words under study, we will cluster all of its embeddings within each time slice. Each cluster should represent a different sense of the word, allowing us to track how each individual sense of the word has changed over time.

Since this would be entirely unsupervised, we need to ensure that humans are in the loop. Threfore, we will then manually check that these clusters make sense and regenerate them using different hyperparamets if necessary.

Finally, we will put together all of the information we have gleaned through our different analyses to determine how the terms we are studying have changed over time.