Table 5.

Illustration of Correlation Explanation (CorEx)Q9 algorithm.

Input:phq_lexicon, Stressed Tweets (geotagged) Output:topic sparse matrix S where row: tweetid and columns: PHQ Stress Level Index (1 to 9)

Procedure: 1. Shallow parsing each tweet into $tweet\_pharse$ using $spaCy$ 2. For each $word\_set$ in $phq\_lexicon$ do 3. Calculate average vector of $word\_set$ and $tweet\_pharse$ using GloVe 4. Match $word\_set$ with $tweet\_pharse$ set using cosine similarity measure 5. Append each matched $tweet\_pharse$ to $word\_set$ 6. Calculate Tf-Idf vector for all the tweets and transform the calculated value to a sparse matrix $X$ 7. Iteratively run CorEx function with initial random variables $V_{random}$ 8. Estimate marginals; calculate total correlation; update $V_{random}$ 9. For each $word\_set$ in $phq\_lexicon$ 10. Compare $V_{random}$ and $word\_set$ with bottleneck function 11. Until convergence