test.lua

require 'nn'

input = torch.DoubleTensor{4} --input word index
target = torch.DoubleTensor{6,5,4,2} --target word indices. So we have 1 true context ("6") and 3 negative contexts ("5,4,2")
label = torch.DoubleTensor({1,0,0,0}) --the first label is true sample, rest are neg samples

model1 = nn.Sequential()
model1:add(nn.LookupTable(1000,4)) -- word embedding layer, vocab size is 1000 and each word is 50 dimensional
model2 = nn.Sequential()
model2:add(nn.LookupTable(1000,4)) -- context embedding layer, this will be dot producted with an element of the word embedding layer
sig_layer = nn.SoftMax() -- SoftMax layer to get probabilities
criterion = nn.BCECriterion() -- simple binary case, so using BCE Criterion

dot_prod = model2:forward(input):cmul(model1:forward(input):view(4)) -- results in a tensor where each element is the log odds of being a true sample
probs = sig_layer:forward(dot_prod) --get probabilities

-- get errors
err = criterion:forward(probs, label)
dl_do = criterion:backward(probs, label)

--backpropagate through both lookuptables
model1:zeroGradParameters()
model1:backward(input, dl_do)
model1:updateParameters(0.01) --update word embeddings

model2:zeroGradParameters()
model2:backward(input, dl_do) --update context embeddings
model2:updateParameters(0.01)

Hey. Please let me know whether the update paramets be called just for the input tensors for the model1 lookup table or for all the weights in lookup table.

I think there is an error in line 14 and 27