Skip to main content
This assignment is due on Friday, February 14, 2020 before 01:30PM.
You can download the materials for this assignment here:

Homework 3: Natural Language Understanding

Instructions

In this homework, you will be experiment with the ROCSTories dataset, which consists of 5-sentence long stories. You will build a few classifiers for the Story Cloze task. The task involves predicting which of two candidate 5th sentences best ends a story.

You should walk through the IPython Notebook here. It walks you through building a sentiment-based system for predicting the correct next ending as well as two neural network-based approaches that learn a classifier for the task.

You should create a writeup describing the experiments you conduct. We will be primarily grading you on this writeup, although you should also submit your code. Your writeup should include

  • Your accuracy on the two validation sets and the 2016 test set using…
    • A sentiment-based classifier
    • A neural network trained only on the train set, as well as 2 variants.
    • A neural network trained in a supervised way on the validation set, as well as 2 variants.
  • Descriptions of your approaches for each of the above methods/variants. Try to make your descriptions detailed enough that another student could reimplement your approach from them.
  • An error analysis:
    • Find a couple examples your sentiment classifier gets wrong and discuss them.
    • Are there any examples that all of your classifiers get wrong? What do you notice about them?

What parameters can your vary?

Here are some ideas:

  • The learning rate and optimizer
  • The number of training steps
  • The number of layers and types of activation functions in the model architecture
  • The batch size
  • For the supervised model that trains on the validation set, can you also incorporate the unlabeled training data?

What to submit

  • A file containing your writeup, as either report.pdf or report.md.
  • Your modified IPython notebook as rocstories.ipynb.