LSTM for sentiment analysis and detection of domestic violence | LA Hacks 2019
This is a quick project that my roommates and I did for LA Hacks 2019. We wanted to use a smartphone's microphone to detect arguments and domestic violence. Lacking a dataset, we downloaded videos of Walmart fights from YouTube and trained and LSTM for sentiment analysis.
Adam Egyed built an Android app that connects to a GCP Cloud Function that runs the ML. Data analysis and model training was done by me in Python.
Sadly, the results weren't great. The LSTM only occassionally identified violent phrases successfully. We believe this is because the transcripts of the Walmart videos were too low quality -- arguments are difficult to convert into organized sentences. In any case, the potential for false positives is far too high for this app to be practical. Further development would require a better dataset.