DATA 311 - Lab 4: Text Normalization and Natural Language Processing

Overview

This pre-lab and lab serve to give hands-on experience with text normalization and natural language processing. In Part I, you will practice normalizing text using the ntlk package. In Part II, you will use the Spacy NLP package to analyze a dataset of fake and real news articles.

Submission and Development Environment

Details for submission and development environment are the same as for Lab 1; please see the Lab 1 handout if you need a refresher. As a reminder, if you choose to use Google Colab or any other alternative notebook hosting service, you must disable any built-in generative AI features.

Collaboration Policy

You must complete the pre-lab individually, but complete the lab in pairs. Your TA will help facilitating pairing; you should pair up with someone you haven’t worked with on a prior lab. If an odd number of students are in attendance at lab, your TA will arrange one group of three or one person working alone. The lab must be done synchronously and collaboratively with your partner, with no free-loaders and no divide-and-conquer; every answer should reflect the understanding and work of both members. If your group does not finish during lab time, please arrange to meet as a pair outside of class time to finish up any work.

The Data

There are four files you will need for this lab. You can find them on our class’s shared storage on the cluster at the following paths:

1. /cluster/academic/DATA311/202620/lab4/lab4.txt

   • This will be used in Part I when practicing text normalization.

   • Credit to the New York Times Magazine.

2. /cluster/academic/DATA311/202620/lab4/contractions.csv

   • This will also be used in Part I when practicing text normalization.

   • Credit to Andrew Tucker, San Jose State University Writing Center.

3. /cluster/academic/DATA311/202620/lab4/Fake.csv

   • This will be used in the Pre-Lab and Part II for Spacy NLP analysis.

   • Credit to the clmentbisaillon’s Kaggle Dataset (Kaggle dataset page)

4. /cluster/academic/DATA311/202620/lab4/True.csv

   • This will be used in the Pre-Lab and Part II for Spacy NLP analysis.

   • Credit to the clmentbisaillon’s Kaggle Dataset (Kaggle dataset page)

Pre-Lab

• Please carefully read “Natural Language Processing With Python’s NLTK Package”

• In a text editor of your choice that is capable of exporting to pdf, answer these questions:

  1. How many tokens are in the following sentence if you call word_tokenize on it? “Muad’Dib learned rapidly because his first training was in how to learn.”

  2. How many sentence tokens are in the same sentence if you instead call sent_tokenize?

  3. What are stemming and lemmatizing and how do they differ?

• Please carefully read “spaCy 101: Everything you need to know”

• In the same editor, answer these questions:

  4. How many tokens are in the following sentence after it is tokenized by Spacy’s small model (en_core_web_sm)?

     Apple is looking at buying U.K. startup for $1 billion

  5. What is a Named Entity? List two example entity labels.

  6. What is a Doc object?

• Please download nlp_spacy_and_counters.ipynb, as well as True.csv and Fake.csv linked above, and upload them to JupyterHub. Go through all the cells in the notebook and answer these questions in the same editor as your previous questions:

  7. How long does it take to run the Spacy model on the short_fake data?

  8. How long does it take to run the Spacy model on the short_real data?

  9. What does token.tag_ == "NNPS" mean in the last cell?

• You can return to this notebook for examples of how to use Spacy to analyze news data in Part II of the Lab.

Lab

Part I

JupyterLab Resources For this lab, you’ll be using a somewhat beefy Spacy language model. When starting your server on JupyterHub, it is recommended that you request 2 CPUs and 6GB of memory; in our experience, this is comfortably sufficient. If the memory usage (displayed in the status bar at the bottom) goes into the red, please restart your server with greater resources.

Create a fresh notebook, lab4.ipynb; title it and include the names of both partners in a Markdown cell at the top. In this notebook, load the contents of lab4.txt (linked above) document into a string. Then implement (and explain your processing in comments and/or text boxes as needed) the following text normalization steps:

1. Replace any instances of successive spaces or tabs with a single space character.

2. Use the contraction-expansion pairs in contractions.csv to find and replace any instances of the contractions anywhere in the document. Hint: don’t overthink this one - use Python’s string processing features.

3. Use nltk’s sentence tokenization to break the document into a list of sentences. Count and print the number of sentences; compute and print the longest sentence (most characters).

4. Use nltk’s word tokenization to break each sentence into a list of tokens. (You will produce a list of lists, where inner lists are lists of tokens.)

5. Lowercase every token.

6. Create a dictionary that maps each unique token to the number of times it appears. Which token appears most often? List the tokens that only appear once.

If you encounter any unexpected challenges, make sure to describe them and how you decided to deal with them.

Part II

In the same lab4.ipynb notebook as Part I, perform the following analyses separately on (1) the True.csv data and (2) the Fake.csv data. For analyses 1-5, use the en_core_web_trf Spacy model - this one is more expensive to run, but also powerful and better at recognizing named entities, among other things.

1. Using news article titles, provide a list of all the countries, cities, and states that are mentioned in the month of March 2017. Briefly comment on your findings.

2. Using the news article titles, calculate the top five people mentioned on election day November 8th, 2016? Briefly comment on your findings.

3. Create histograms of real and fake article sentiment mentioning the top person (from the previous analysis). Briefly comment on your findings.

4. Plot histograms for real and fake news article sentiment (by title) for a 2 month period of your choosing. Briefly comment on your findings.

5. Perform 1 additional analysis of your choosing on some date range/date of your choosing. Justify your analysis choice, and comment on your findings.

6. Time running each Spacy model (en_core_web_sm, en_core_web_md, en_core_web_lg,
   and en_core_web_trf), on the first 5 full text news articles from True.csv. Display your findings in a bar chart with the bar heights equal to the time elapsed in each of your timing experiments and a bar for each Spacy model. Briefly discuss your findings.

Rubric

Pre-Lab (9 points)

• 1 point per question

Lab (48 points)

Part 1 (24 points)

• 4 points per task

Part 2 (24 points)

• 4 points per task

Acknowledgements

Thanks to Brian Hutchinson and Aaron Tuor for developing and iterating on prior versions of this lab.