CSE 258: Homework 4

CSE 258: Homework 4
Instructions
Please submit your solution by the beginning of the week 9 lecture (Mar 6). Submissions should be
made on gradescope. Please complete homework individually.
Download the “50,000 beer reviews” data from the course webpage: http://jmcauley.ucsd.edu/cse258/
data/beer/beer_50000.json. Code is provided on the course webpage (week5.py) showing how to load and
perform simple processing on the data. Executing the code requires a working install of Python 2.7 with the
scipy packages installed.
Tasks
Using the code provided on the webpage, read the first 5000 reviews from the corpus, and read the reviews
without capitalization or punctuation.
1. How many unique bigrams are there amongst all of the reviews? List the 5 most-frequently-occurring
bigrams along with their number of occurrences in the corpus (1 mark).
2. The code provided performs least squares using the 1000 most common unigrams. Adapt it to use
the 1000 most common bigrams and report the MSE obtained using the new predictor (use bigrams
only, i.e., not unigrams+bigrams) (1 mark). Note that the code performs regularized regression with a
regularization parameter of 1.0.
3. Repeat the above experiment using unigrams and bigrams, still considering the 1000 most common.
That is, your model will still use 1000 features (plus an offset), but those 1000 features will be some
combination of unigrams and bigrams. Report the MSE obtained using the new predictor (1 mark).
4. Using the model from the previous questions which are the 5 unigrams/bigrams with the most positive
associated weights, and the 5 unigrams/bigrams with the most negative associated weights (1 mark)?
5. What is the inverse document frequency of the words ‘foam’, ‘smell’, ‘banana’, ‘lactic’, and ‘tart’? What
are their tf-idf scores in the first review (using log base 10) (1 mark)?
6. What is the cosine similarity between the first and the second review in terms of their tf-idf representations
(considering unigrams only) (1 mark)?
7. Which other review has the highest cosine similarity compared to the first review (provide the beerId
and profileName, or the text of the review) (1 mark)?
8. Adapt the original model that uses the 1000 most common unigrams, but replace the features with their
1000-dimensional tf-idf representations, and report the MSE obtained with the new model.
1