Assignment 1.3: Naive word2vec (40 points)

This task can be formulated very simply. Follow this paper and implement word2vec like a two-layer neural network with matrices $W$ and $W'$. One matrix projects words to low-dimensional 'hidden' space and the other - back to high-dimensional vocabulary space.


You can use TensorFlow/PyTorch and code from your previous task.

Results of this task: (30 points)

  • trained word vectors (mention somewhere, how long it took to train)
  • plotted loss (so we can see that it has converged)
  • function to map token to corresponding word vector
  • beautiful visualizations (PCE, T-SNE), you can use TensorBoard and play with your vectors in 3D (don't forget to add screenshots to the task)

Extra questions: (10 points)

  • Intrinsic evaluation: you can find datasets here
  • Extrinsic evaluation: you can use these

Also, you can find any other datasets for quantitative evaluation.

Again. It is highly recommended to read this paper

Example of visualization in tensorboard:

Example of 2D visualisation: