Comparing emotion feature extraction approaches for predicting depression and anxiety

For example, pride may be impacted by depression in a unique way. Gruber et al. (2011) showed that pride, a positive emotion relating to the self, is inversely correlated with depression, which is often associated with a poor self-image.

new opportunities to assist practitioners in quantifying depression and anxiety severity by assessing emotion in patient-generated text. Linguistic Inquiry and Word Count (LIWC) (Pennebaker et al., 2007; Tausczik and Pennebaker, 2010) is a software package designed to count words belonging to pre-defined categories with an extensive track record of validation for the detection of linguistic indicators of mental state (Tausczik and Pennebaker, 2010). It is commonly used to measure positive and negative affect, a limited set of specific emotions (sadness, anxiety, and anger), and other linguistic dimensions related to style and topic. Several LIWC categories have established relationships with depression, including the affect category sadness (e.g. “sad”, “cry”, “suffer“), the topic category health (e.g. “alcohol”, “rash”, “self-care”), and the syntactic category firstperson pronouns (e.g. “I”, “me”, “my”). LIWC has been used to measure depression levels in social media posts (Coppersmith et al., 2014; De Choudhury et al., 2014, 2013a,b), therapy conversations (Burkhardt et al., 2021; Sonnenschein et al., 2018), and other written texts (Rude et al., 2004; Wiltsey Stirman and Pennebaker, 2001).

However, word counting methods cannot address linguistic phenomena such as negation (“not bad”), sarcasm, and context-dependence (for example, in the case of polysemy, words have multiple meanings that can only be disambiguated in context), and manually defined dictionaries may omit synonyms for terms they encode.

Shen and Rudzicz found that the performance of machine learning models identifying whether or not Reddit posts were drawn from anxiety-related subreddits improved when these models included neural word embeddings rather than LIWC-derived features (2017).

contemporary transformer-based NN language models offer advantages over neural word embeddings in their ability to leverage proximal cues (such as "not")