Detecting Complaints From Multilingual Twitter Data

This project involved scraping data from the Twitter handles of government departments, such as the Railway department, and then using a transformer based method on them to detect complaints and grievances. In contrast to previous methods, our method aimed to solve the problem of handling code-mixed and multilingual text, which is the majority of all tweets in India. In addition to classifying between complaints and non-complaints, we also classify into six classes based on the type of complaint. We achieve an accuracy of 85% on the binary classifier and 98% on the multiclass classifier. This work has been submitted to EMNLP 2020, and is awaiting review.