Research Assistant to Dr. Lyle Ungar, Professor in Computer and Information Science at Penn Engineering and Psychology at the School of Arts & Sciences, and, Dr. Sharath Chandra Guntuku, an Assistant Professor (Research) in the Department of Computer and Information Science at the University of Pennsylvania.

Research Overview

Professor Lyle Ungar is a renowned expert in the field of Artificial Intelligence and his research group actively develops explainable machine learning, deep learning, and natural language processing methods for psychology and medical research. We are currently working on leveraging Natural Language Processing and Multi-Task Learning to predict depression symptoms from Reddit posts, and analyze the language use of users with different depression symptoms.

My Contribution

1. Shortlisted a set of 30 subreddits based on depression research and questionnaires like PHQ9, BDI-2, CESD-R
2. Mined the data from the subreddits that possibly resonate with the 10 depression symptoms and the corresponding control data of the same set of users using the pushshift API
3. Cleaned the data containing approximately 100k posts by removing bots, empty posts, replacing URLs, etc
4. Leveraged DLATK package for Python to create n-grams, identify latent topics present in the posts and comments through Latent Dirichlet Allocation (LDA; Blei et al., 2003) and then extract LDA, Linguistic Inquiry and Word Count (LIWC) and Bidirectional Encoder Representations from Transformers (BERT) embeddings.
5. Comparing the performance of models like Logistic Regression with regularization, Random Forest Classifier, Extra trees Classifier, Multi-task Learning to predict the depression symptom of a reddit post.
6. Further, perform differential language analysis to understand the language use of people varies across different depression symptoms

Differential Language Analysis ToolKit (DLATK)

Citation

DLATK is an end to end human text analysis package, specifically suited for social media and social scientific applications. It is written in Python 3 and developed by the World Well-Being Project (WWBP) at the University of Pennsylvania and Stony Brook University. It contains:

1. feature extraction
2. part-of-speech tagging
3. correlation
4. prediction and classification
5. mediation
6. dimensionality reduction and clustering
7. wordcloud visualization

DLATK can utilize:

1. Mallet for creating LDA topics
2. Stanford Parser
3. CMU’s TweetNLP
4. pandas dataframe output