- Implementation of Meta-Word-Embeddings, a combination of word2vec, GloVe, and fassttext word embeddings using various types of autoencoders.
- Implementation based on the paper "Learning Word Meta-Embeddings by Autoencoding". [paper link]
- The meta-embeddings are generating using three different variants of autoencoders namely:
- Decopuled Autoencodeder (DAE)
- Concatenated Autoencoder (CAE)
- Averaged Autoencoder (AAE)
- Each of them is explained below.
- The figure above shows the architecture of the Decopuled Autoencoder.
- DAEME can be seen as an extension of naive concatenation that has non-linear neural networks applied on the raw model. Here, each encoder namely: E1, E2, and E3 can be seen as independently performing a transformation to the respective source embedding so that it can learn to retain essential information rather than simply concatenate features.
- The decoders D1, D2, and D3 independently reconstruct the source embeddings from the encoded versions of the source embeddings. Because of this behaviour we call this approach decoupled autoencoded meta-embedding.
- For the detailed explaination of DAEME, refer this link.
- The figure above shows the architecture of the Concatenated Autoencoder.
- Similar to DAEME, DAEME, the meta-embedding in CAEME is also constructed as the concatenation of the encoded source embeddings.
- However, instead of treating the meta-embedding as individual components, CAEME reconstructs the source embeddings from the same meta-embedding, thereby implicitly using both common and complementary information in the source embeddings.
- TCompared to DAEME, CAEME imposes a tighter integration between the two sources in their meta-embedding.
- For the detailed explaination of CAEME, refer this link.
- The figure above shows the architecture of the Averaged Autoencoder.
- AAEME can be seen as a special case of CAEME, where we compute the meta-embedding by averaging the encoded sources embeddings instead of their concatenation.
- This operation has the benefit that we can transform the sources such that they could be averaged in the same vector space, and also guarantees orthogonality between the encoded vectors.
- For the detailed explaination of AAEME, refer this link.