Re-implementation of the paper Real-Time Emotion Recognition via Attention Gated Hierarchical Memory Network as part of Advanced Deep Leaning Course.
Following modifications are made in architecture.
Utterance Reader : Instead of taking only max of hidden units, A weighted sum of hidden units is added. The weights are calculated using projection to a learnable vector followed by softmax.
Attention GRU To compute the importance of context the proposed model use dot product of context and query. It is possible that the importance of a context is not proportional to the cosine similarity with the query. So a new method is suggested in which both query and context is transformed to other dimension using separate learnable parameters and the dot product is calculated to get weights.
And Added information to distinguish speakers in a conversation