End to end knowledge-based, recommendation dialog system.
Recommendation system: CF(collaborative filtering), neural networks, etc.
In this paper, item embedding and user embedding is used.
Item embedding is computed on structure information of relational graph of knowledge DB, to incorporate varieties of relations between items. R-GCN(Relational Graph Convolutional Networks) is used to do so.
R-GCN:
level l+1 hidden vector for item i is computed from level l hidden vectors of items in neighborhood of item i, which is an implementation of convolutional method, with convolutional kernals (filters) to cover neighborhood.
User embedding is computed by AM(attention model), which transform variable-length item-embedding vectors into fixed-size embedding. Hu is concatenated vector of dialog-mentioned-item vectors.
In recommendation, probability of items to be chosen by user u is as below, where mask(.) means assigning low probability to non-entity(which may be from dialog, but NOT in recommending item list) items by means of masking.(Ref to BERT implementation). The mask operation is to assure recommended items on shelf.
Dialog system
Transformer encoder-decoder network is used to modal dialogue content.
A vocabulary bias bu is added to the top layer of the decoder, to introduce more consistency with user interests. bu is computed based on the recommender system’s hidden representation of user u. F(tu) is a feedforward network, tu is user embedding as above depicted.
Knowledge graph:
KG is used as knowledge source to acquire item(entity) relationship embedding. In this paper, DBpedia is used.
Probability-based switching mechanism:
o is the hidden representation in the final layer of the dialog system.