CSc22100 Assignment 5

In this assignment, we'll explore a Java implementation of Peter Norvig's Toy Spell Checker.

Instructions

Run ./gradlew download to download the data sets. They'll download to a directory called ./data.
Implement the prepare method in the EvalSet class.
In WordFrequencyCounter.java implement countFrequencies.
Write some tests for (2) and (3) (Hint: Life will be easier if you make a new test class rather than writing the tests in SpellCheckerTest)
When you're satisfied with your implementation, run ./gradlew prepareTask
Next, implement train in LanguageModel to compute the probability of each token based on the word frequencies. (See Norvig's explanation or Wikipedia)
Write some tests for LanguageModel to convince yourself you've done it correctly. (Hint: Use one or two sentences as your 'corpus' and you can compute the expected probabilities by hand easily.)
Run ./gradlew train to train the language model
Implement each of the methods in SpellChecker.
There are some unit tests to check that SpellChecker is working correctly. Unfortunately, because they rely on the path to the dataset, you may need to set a system property in your IDE, or run them with gradle (./gradlew testAll).
Run ./gradlew evaluate to see how your spell checker does. You can have extra points if you outperform my implementation in either accuracy (on the "hidden" test set--it's on Norvig's site) or in words per second.

My Results

java -jar build/libs/assignment5-all.jar grade --data-dir ./data --working-dir ./data 
[main] INFO csc22100.spelling.SpellCheckerMain - Preprocessing corpus...
[main] INFO csc22100.spelling.SpellCheckerMain - Preprocessing dev examples...
Correct: 200
Incorrect: 55
Accuracy: 0.7843137254901961
Words per second: 127

Grading

Item	Points
Code compiles	1
Tests pass	1
Code cleanliness/readability (are your language model tests in `LanguageModelTest` etc?)	2
At least two test cases for `WordFrequencyCounter`	2
At least two test cases for `LanguageModel`	2
Correctly packaged	1

Extra Points:

2 for outperforming me on the test set. (Be creative--can you improve on this fairly simple approach?)
2 for having an implementation that runs faster than ~127 words per second on my laptop (disclosure: this was 85 wps before I removed some logging)

Tip If you're going to work on the extra points, I suggest completing the assignment first, saving a copy (or making a git commit) and then start trying to optimize your use of StringBuilder or whatever.

What to submit

As ever, submit two things: a zip file (run ./gradlew packageAssignment) and a PDF of your code (remember to put source2pdf.jar in the same directory as gradlew and then run ./gradlew source2pdf)

Other notes

Right now the build gradle task depends on the testAll task, which requires that you have trained the language model. If you want to run build without running those tests, remember you can skip them with ./gradlew build -xtestAll.

The project is set up to use the shadow plugin, meaning that it builds a runnable jar (complete with all the depedencies packaged together). If you'd rather run individual commands, you can run the jar itself:

java -jar build/libs/assignment5-all.jar        
Missing required subcommand
Usage: spell [-hV] [COMMAND]
  -h, --help      Show this help message and exit.
  -V, --version   Print version information and exit.
Commands:
  download
  evaluate
  grade
  prepare
  train

sacsar / csc-22100-assignment5