In this project, I design, implement and compare deep-learning siamese networks to learn a similarity measure between names and nicknames. Please see a thorough project description here.
In short, I explore both text-based models, implementing Neculoiu, Versteegh and Rotaru (2017) and modified versions of Gehring et. al (2017). Then, I also test the effect of converting texts to audio and using 2d-CNNs siamese networks on spectrograms instead of raw texts.
Converting to audio makes sense in the current use-case since, in many cases, nicknames involve changing letters while keeping the sounds of the nickname similar to the original name (for example, replacing "i" with "ea"). Results showed that converting to audio using Google's gTTS
and then using 2d-CNN models on spectrograms led to significant improvement in balanced accuracy, F1 and PR AUC, and for a much lower number of parameters.