byte2vec: a flexible embedding model con...

In today's fragmented, globalized world, supporting multiple languages in NLU and NLP applications is more important than ever. The inherent language dependence in classical Machine Learning and rule-based NLP systems has traditionally been a barrier to scaling said systems to new languages. This dependence typically manifests itself in feature extraction, as well as in pre-processing steps. In this talk, we present byte2vec as an extension to the well-known word2vec embedding model to facilitate dealing with multiple languages and unknown words. We explore its efficacy in a multilingual setting for tasks such as Twitter Sentiment Analysis and ABSA. Byte2vec is an embedding model that is constructed directly from the rawest forms of input: bytes, and is: i. truly language-independent; ii. particularly apt for synthetic languages through the use of morphological information; iii. intrinsically able to deal with unknown words; and iv. directly pluggable into state-of-the-art NN architectures. Pre-trained embeddings generated with byte2vec can be fed into state-of-the-art models; byte2vec can also be directly integrated and fine-tuned as a general-purpose feature extractor, similar to VGGNet's current role for computer vision.