search

PhilippChr / CONVEX

Code for our CIKM 2019 paper. As far as we know, CONVEX is the first unsupervised method for conversational question answering over knowledge graphs. A demo and our benchmark (and more) can be found at
https://convex.mpi-inf.mpg.de/
MIT License
28 stars 9 forks source link

Node Embedding #14

Closed SRL94 closed 2 years ago

SRL94 commented 2 years ago

Hi, in Equation (2), CONVEX initialises the word embeddings of a question by using word2vec but how does CONVEX initialise node embeddings?

PhilippChr commented 2 years ago

Hello,

within each node you would have a phrase (label of the KB-item). So node embeddings are the averaged word embeddings.

We do not materialize these initial node embeddings, but directly compute the matching similarity with the question: https://github.com/PhilippChr/CONVEX/blob/260d02933748abc74b0fc1d317ad46181960c2ce/convex.py#L226-L236

This function implements Equation (1), using cosine similarity.

Hope this answers your question.

Regards, Philipp

SRL94 commented 2 years ago

How about unknow words?

PhilippChr commented 2 years ago

Unknown words would get a 0 vector. A similarity of 0 is returned. https://github.com/PhilippChr/CONVEX/blob/260d02933748abc74b0fc1d317ad46181960c2ce/library/glove_similarity.py#L34-L35