Inspired by CharSCNN [1,3], Dracula combines character-level embeddings with two levels of deep LSTM representation for compelling performance, with small model sizes.

Dracula's performance (with 128-size word embeddings) on GATE's TweetIE T-Eval dataset is competitive with Derczynski et. al [2], but Dracula achieves this performance without word lists, frequency filtering, or any regular expressions.

Technical details

Dracula is implemented using Theano, and originates from Pierre Luc Carrier and Kyunghyun Cho's LSTM Networks for Sentiment Analysis tutorial. It's also readily portable to other frameworks, such as Google's TensorFlow.

Further reading

1. C. Santos and B. Zadrozny, "Learning character-level representations for part-of-speech tagging", Proceedings of the 31st International Conference on Machine Learning (ICML-14), pp. 1818--1826, 2014.
2. L. Derczynski, A. Ritter, S. Clark and K. Bontcheva, "Twitter Part-of-Speech Tagging for All: Overcoming Sparse and Noisy Data.", RANLP, pp. 198--206, 2013.
3. C. D. Santos and M. Gatti, "Deep Convolutional Neural Networks for Sentiment Analysis of Short Texts", COLING, pp. 69-78, 2014.