TwitMoji

TwitMoji Sample Predictions

Inspiration

In the day and age of Twitter, messenger, and rampant texting, wouldn't it be amazing to know the perfect emoji to use every time? We were inspired by this desire, but more than that, we saw how many people, content creators, and business struggle with choosing the right emojis, often opting to use none. Past studies have shown that emojis can significantly increase user engagement. Therefore, misusing them, or failing to use them at all, can lead to missed business opportunities for local businesses looking to advertise their products. On a more personal level, it can also mean older generations choosing to avoid reaching out to others in the community over Twitter, or elsewhere, because they don't feel comfortable with their emoji game. This tool makes it possible to get real time emoji recommendations on a user's message to ameliorate all of these issues!

What it does

From our command line tool, users can type in the text of their tweet. The text is then passed through our Transformer-based model and a prediction of the best emoji to go along with that text will be outputted!

How we built it

We leveraged Hugging Face Transformers library and PyTorch to train our model off a pre-trained BERT base. We particularly leveraged a RoBERTa variant that was further pre-trained on sentiment-laden Twitter text, which made the model intrinsically more domain adapted.

Challenges we ran into

We didn't have sufficient time in this hackathon to train our predictive model for as long as we wanted to. Additionally, it would have been great to leverage Scale's products and technologies to get more labeled data for this task.

Accomplishments that we're proud of

We're proud that we got a reasonable model trained during this hackathon period, especially as we are both busy graduate students.

What we learned

We learned that emoji prediction, especially in this difficult multi-class setup, can be quite challenging. Large volumes of clean, labelled data are not readily available, which makes this task more difficult.