This repository maintains the code from our experiments where we used Long Short-Term Memory(LSTM) networks to understand the video content from the transcripts in order to identify if a video can trigger depression. Additionally, we propose a novel method of validating the results by analysing the CESD score of comments. You can find more details about this experiment in our paper / poster that we presented at NeurIPS 2019 AI for Social Good Workshop. Our major contributions in this work are:
- We construct a classifier that can help understand the content of a video by classifying it as Depressive/NonDepressive.
- To provide a real-life validation of the classification results in above step, we propose a methodology to evaluate the comments posted for a video and determine a potential score that would have been obtained on a CES-D scale and use it as a real life proxy to judge the accuracy of the classification.
- YouTube Data API
- Empath
- urllib
- pandas
- numpy
The process of data collection has been explained in detail in Section 2.1. The code for the same can be found here
For evaluation of the comments, we introduce a scoring method (Section 3), called CES-D score, for each video to analyze how depressive the video is. The score is the density of the terms, derived from the various symptoms (Insomnia, Self-hate, Appetite, etc.) considered in the CES-D scale, present in a negative connotation within a given text. Code for the same can be found here
We compare the performance of various models on their ability to classify a video as depressive/non-depressive by processing its transcripts.
An extension of this project involves analysing audio-visual features of the video to predict the arousal-valense response it would generate. More details can be found on the project website