CS 151: Programming Assignment 4 [PAIR OPTION]

Naive Bayes Nets [90 points +10 extra credit]
due Wednesday, March 9 11:55pm

This assignment was designed by Prof. Sara Sood.

Provided files for this assignment:
The purpose of this assignment is to program a simple Bayesian classifier (naive Bayes, which is a VERY simple Bayes net).

In this assignment, you will implement a Naïve Bayes Classifier that analyzes the sentiment conveyed in text. When given a selection of text as input, your system will return whether the text is positive, negative, or neutral. The system will use a corpus of movie reviews as training and testing data, using the number of stars assigned to the each review by its author as the truth data (1 star is negative, 5 stars is positive). Depending on time, we may hold a short tournament to test the accuracy of your systems.

Notice that I have provided a template file for this assignment, “bayes_template.py.” Copy this file to your machine and rename it “bayes.py” – where you will complete your work for parts 1, 2 and 3. Create a file called “bayes_best.py” for part 4 and put all of your reflection and evaluation from the questions in the assignment in a file called “reflection_evaluation.txt”.   You will submit all three files through the assignments page on Sakai.

Part 0: Overview

Recall that a Naive Bayes classifier is just a simple Bayes net that has one root node with some number of children.  The children are connected only to the root, not to each other, as in the network shown below:

Naive Bayes network

This is the Naive Bayes network we will begin with for this assignment.  In this network, all variables are Boolean.  The variable "Positive" represents whether the document is a positive review (Positive=false means the document is a negative review), and the variables "word_1", "word_2", ..., "word_n" represent the presence or absence of different words in the document.  The idea behind this network is that certain words are more likely to occur in postive documents, while others are more likely to occur in negatives ones.  However, the assumption we make is that once we know whether the document is positive or negative, the occurance of the words in the document become independent from one another.

Your task in this assignment will be to learn the necessary CPTs needed to perform inference with this network, i.e. P(Positive) and P(word_n|Positive).

Part 1: The Training Data

To train your system, you will be using data from a corpus of movie and product reviews collected from www.rateitall.com. The reviews come from a variety of domains, e.g., movies, music, books, and politicians and are stored in individual text files where the file name is of the form domainnumber_of_stars-id.txt. For example, the file “movies-1-32.txt” contains a movie review that was assigned one star by its author, and has an id of 32. Please note that the id number will be of no use to you for this assignment.

For your system, you will only be using reviews that had one or five stars, where the reviews with one star are the “negative” training data and the reviews with five stars are the “positive” training data. I’ve provided you with two Python functions (in the template file) to help you in using this training data.

Familiarize yourself with these functions by using them at the python interpreter to read reviews from the training data files. Take a moment to explore the code provided in the template file so that you don’t create extra work for yourself.  Note that you might need to use full paths on
filenames to get these functions to work.

Part 2: Training the Bayes Net [25 points]

Now that you’re familiar with the training data, your next task is to train the classifier. As you know from your reading and from class, a Naive Bayes classifier is just a simple Bayes Net, in this case the one shown above.  

We will use maximum likelihood parameter estimation to estimate each parameter in the Bayes net.  As a reminder, the parameters that we need to estimate are all of the conditional and prior probabilities we need to specify the network, which are:
Recall from class (on Monday, February 28) that the ML estimate for these parameters are:

P(positive) = #positive / #docs
P(word_i|positive) = (#word_i in positive documents) / (#words in positive documents)
P(word_i|negative) = (#word_i in negative documents) / (#words in negative documents)

Where #positive is the total number of positive documents, #docs is the total number of documents, #word_i is the total frequency of word_i (in positive or negative documents).

When training the system, instead of storing the actual parameter estimates, you will simply calculate the frequencies needed to calculate the parameters on the fly.  In calculating these parameter estimates, you will need to keep track of how often each word appears in positive documents and how often each word appears in negative documents.    There are two different types of tallies that are commonly used: presence refers to the number of documents in the training data that the feature occurs in and frequency refers to the number of times that the feature occurs in the training set (i.e. if it occurs three times in one document, that counts as three).  We will use frequency counts for this part of the assignment to estimate parameters, but you might experiment in part 4 with using presence counts instead.

Given all of the counts that are to be stored, Naïve Bayes Classifiers typically use a database. However, since you are writing this code in Python, we’ll instead use a handy Python utility called pickle. Using pickle, one can write a data structure to a file, and then load it back into memory later. In particular, you’ll have two dictionaries (like hashtables), one which holds all of the words that occur in positive documents and the frequencies at which they occurred, and the corresponding dictionary for negative documents. Since these dictionaries are time consuming to construct, it is useful to only calculate them once, pickle them, and then load them into memory the next time you need to use them.

Write a function called train that takes no parameters, and trains the system by:
    lFileList = []
    for fFileObj in os.walk("reviews/"):
        lFileList = fFileObj[2]
        break

Following the execution of the above code, the list of filenames will be stored in lFileList.

Part 3: Start Classifying! [30 points]

At this point, you now have all the data you need stored in memory, and can start classifying text!  Given a piece of text, the goal is for your system to output the correct document class (i.e. positive, negative, or neutral). In particular, you’ll calculate the conditional probability of each document class given the features in the target document (e.g. P(positive | word1, word2, …)) and return the document class of the highest probability.

To classify a document, you need to find:

P(positive|word_1, word_2, ..., word_k)

That is, the probability that a document is postive, given the words in that document.  Like any Bayes Net, the Naive Bayes classifier calculates this probability as the product of the probability of each node in the network, given the value for its parents.  In this case:

P(positive|word_1,...,word_k) = alpha * P(positive)*P(word_1|positive)*...*P(word_k|positive)

Note, however, that this probability ignores some information that we are given.  In particular, it only considers words that you HAVE seen in the document, and it does not factor in the words that appearn in the Bayes net that you HAVE NOT seen.  What we're really after is:

P(positive|word_1,...,word_k,¬word_k+1,...,¬word_n) =
alpha * P(positive)*P(word_1|positive)*...*P(word_k|positive)*P(¬word_k+1|positive)*...*P(¬word_n|positive)

where word_k+1...word_n are words that are in the full Bayes Net (i.e., words seen in some document), but do not happen to appear in this document.  Effectively the simplification above is saying that we have not observed word_k+1...word_n, i.e., we haven't seen them, but we haven't NOT seen them either.

Reflection question #1: In your "reflection_evaulation.txt" file, answer the following question:
Furthermore, we don't *really* care about alpha, since it is the same for P(positive|[words_in_doc]) and P(negative|[words_in_doc]).  Thus, if you want you can ignore it and just compare the unnormalized probabilities for positive and negative directly.

You will find that two problems will arise in building this basic classifier. I suggest that build the classifier first, and then address the problems, which are:
  1. Underflow – Because you are multiplying so many fractions, the product becomes too small to be represented. The standard solution to this problem is to calculate the sum of the logs of the probabilities, as opposed to the product of the probabilities themselves.
  2. Smoothing – This problem is more subtle as it won’t produce a bug, but will throw off your calculations. Suppose a feature (word) f does not occur at all in the training data for negative (or positive) documents.  Then our estimate for P(f|negative)=0.  This one missing word essentially negates the influence of all other words in the document!  One word should not be given that much power in our decisions.  Read online about “add one smoothing” as a solution this problem.  Feel free to implement any smoothing algorithm you choose.
Write a function classify that takes as input a string of text, and returns a string of the classification – either “positive,” “negative,” or “neutral.” Note that in building your classifier, you only have data to calculate the probability of a document being positive or negative, so you’ll have to come up with some way to determine if a document is neutral (keep it simple).

If you’ve followed my instructions, you should now have a classifier that can be used in the following manner:
>>> execfile("bayes.py")
>>> bc = Bayes_Classifier()
>>> result = bc.classify("I love my AI class!")
>>> print result
positive

Reflection Question #2: Take a moment to celebrate your success. Try out your classifier on a few sentences and see when it performs well and when it performs poorly. What are three examples of sentences (or any length of text) on which you think your classifier failed? For each example, why do you think it failed? What was hard about the target text? Write your answers to these questions in a file called “reflection_evaluation.txt.”

Part 4: Improve your Classifier [15 points + 5 extra credit]

Congratulations! Once you reach this point, you will have built your first Naïve Bayes Classifier! If the previous part, you probably realized some ways in which you can improve upon your system. In this part, I would like you to be creative and implement these ideas and build your best classifier. Time permitting; we’ll hold a competition in class to see whose classifier does best on a set of documents that I’ll choose randomly. The outcome of this competition will not have an impact on your grade. For this part, make a copy of your python file and add the word “best” to the end of the filename. This will make it easier for me to grade these parts separately.

The most important research problem with respect to Naïve Bayes Classifiers is the choice of features used in classification (i.e., the children nodes in the Bayes net). Thus far, you’ve only used “unigrams” (i.e. single words) as features. Another common feature is “bigrams” (i.e. two word phrases). Another feature might be the length of the document, or amount of capitalization or punctuation used. For your best classifier, you must include at least one other feature in your classification.  For particularly creative or ambitious classifiers, you can earn up to +5 extra credit points.

Part 5: Evaluate Your System [20 points +5 extra credit]

I’d now like you to evaluate how well your system works. In particular, I’d like you to compare your base system (that you completed in part 3) to your best classifier (that you completed in part 4).

You should first use the standard measures of performace for text classification: precision, recall and f-measure.  Why three different measures?  As you've likely noticed, there is a tradeoff between performing well on positive documents and performing well on negative documents.  That is, you could classify all positive documents correctly simply by saying that all documents are positive, but then you would perform very poorly on negative documents.  Precision, recall and f-measure are ways to quantify this tradeoff.

Precision is a measure of how many irrelevant documents you included in a class.  In this context it is defined as:

# of positive (or negative) documents classified as positive (negative) / total # of documents classified as positive (negative)

Recall, on the other hand, is a measure of how many of the documents you were supposed to include in a class actually got included in that class.  In our context it is defined as:

# of positive (or negative) documents classified as positive (negative) /  total number of positive (negative) documents

For example, a Precision of 1.0 on positive documents means that all the documents you classified as positive were actually positive (though you might have missed some).  A Recall of 1.0 on positive documents means that you classified all positive documents as positive (though you might have also classified some negative documents as positive too).

Finally, f-measure is simply a way to combine recall and precision into a single number:

F = 2 * (precision * recall) / (precision+recall)

Not only should you compare the two systems based on their precision, recall, and f-measure, but also reflect on why you think the systems performed well (or poorly). Outline ways that you would extend the system to improve future performance. Consider and present other ways that you might approach this task of classifying sentiment in text. Include your careful evaluation and write your answers to these questions in a file called “reflection_evaluation.txt.”

You should think carefully about your evaluation.  Consider testing your classifier on documents not included in the set I provided.  How will you tease out the specific strengths and weaknesses of your classifier (beyond precision and recall).  Particularly complete, creative or ambitious evaluations (e.g., comparing more than just the two classifiers) will earn up to +5 extra credit points.