Reviews: Hash Embeddings for Efficient Word Representations

This work describes an extension of the standard word hashing trick for embedding representation by using a weighted combination of several vectors indexed by different hash functions to represent each word. This can be done using a predefined dictionary or during online training. The approach has the benefit of being easy to understand and implement and greatly reduces the number of embedding parameters. The results are generally good and the approach is quite elegant. Additionally, the hash embeddings seem to act as an effective regularizer.