COMP9444 Neural Networks and Deep Learning

Session 2, 2017

Project 2 - Recurrent Networks and Sentiment Classification

Due Dates:
Stage 1: Sunday  8 October, 23:59 pm
Stage 2: Sunday 15 October, 23:59 pm
Marks: 15% of final assessment

Introduction

Preliminaries

You should download the following files from the directory hw2/src

hw2.zip	Python source code for Stage 1
hw2sent.zip	Python source code for Stage 2
reviews.tar.gz	25000 plain text movie reviews, split into positive and negative sets
glove.6B.50d.txt	GloVe word embeddings for Stage 2

Dataset

You will need to extract these files, and load them into a datastructure you can feed into TensorFlow. It is essential to preform some level of preprocessing on this text prior to feeding it into your model. Because the glove embeddings are all in lowercase, you should convert all reviews to lowercase, and also strip punctuation. You may want to do additional preprocessing by stripping out unessesary words etc. You should examine any avenue you can think of to reduce superfluous data that you will feed into your model.

For the purposes of reducing training time, you MUST limit every review fed into the classifier at 40 words. This should occur after your preprocessing. The model must only accept input sequences of length 40. If a review is not 40 words it should be 0-padded. Some reviews are much longer than 40 words, but for this assignment we will assume the sentiment can be obtained from the first 40.

For evaluation, we will run your model against a test set that contains an additional 25000 reviews spit into positive and negative categories. For this reason you should be very careful about overfitting - your model could report 100% training accuracy but completely fail on unseen reviews. There are various ways to prevent this such as judicious use of dropout, splitting the data into a training and validation set, etc.

Tasks

Groups

Stage 1: Word Embeddings

A good introduction to word embeddings can be found in the TensorFlow word2vec tutorial. This section of the assignment builds on that tutorial.

The aim of this task is to modify the code here so that it uses the continuous bag of words (CBOW) model instead of the skip-gram model. This should produce better embeddings, particularly when less data is available. Furthermore, implementing this change should give you a better understanding of both models, and the differences between them.

CBOW vs Skip-gram

Input-Output

The main difference between the skip-gram and CBOW model, is the way training data is presented.

With the skip-gram model, the input is the word in the middle of the context window, and the target is to predict any context word (word that is skip_window words to the left or the right) for the given word.

With CBOW, the input is all the words in the context besides the middle word, and the target is to predict the middle word, that was omitted from the context window.

For example, given the sentence fragment "the cat sat on the", the following training examples would be used by skip-gram, with parameters skip_window=1, num_skips=2 - in the form: [words in context window]: (input, target)

[the cat sat]: (cat, the), (cat, sat)
[cat sat on]: (sat, cat), (sat, on),
[sat on the]: (on, sat), (on, the)

While for CBOW the input-output pairs are (note that the inputs now contain more than one word):

[the cat sat]: ([the sat], cat),
[cat sat on]: ([cat on], the),
[sat on the]: ([sat the], on)

Of course, as is explained in the tutorial, the words themselves aren't actually used, but rather their (integer) index into the vocabulary (dictionary) for the task.

CBOW Input: Mean of Context Words Embeddings

In the skip-gram model there is just a single word as the input, and this word's embedding is looked up, and passed to the predictor.

In the CBOW, since there's more than one word in the context we just take the mean (average) of the embeddings for all context words (hint tf.reduce_mean(?, axis=?)).

Task

word2vec_fns.py	skeleton code for your word2vec implementation
word2vec_cbow.py	code to train your word2vec model
imdb_sentiment_data.py	helper functions for loading the sentiment data, used by word2vec_cbow
plot_embeddings.py	to visualise embeddings

The file word2vec_fns.py is the one you will need to modify and submit. It contains two functions:

1. generate_batch(...) which is initially identical to the function in https://github.com/tensorflow/tensorflow/blob/r1.3/tensorflow/examples/tutorials/word2vec/word2vec_basic.py, with just one change, the num_skips parameter has been removed as it is not needed in the CBOW regime.
2. get_mean_context_embeds(...)

You are to:

1. modify generate_batch so batch is a vector of shape (batch_size, 2*skip_window), with each entry for the batch containing all the context words, with corresponding label being the word in the middle of the context
2. implement get_mean_context_embeds so that it returns mean_context_embeds - the mean of the embeddings for all context words for each entry in the batch, see the docstring in the function for more details.

You can run the code that does the embeddings with:

python3 word2vec_cbow.py

If this completes without error, you should see a file called CBOW_Embeddings.npy in the current directory, which you will need to submit along with your version of word2vec_fns.py

Additionally, if you run

python3 plot_embeddings.py

you should be able to see a low dimensional visualisation of the embeddings created with TSNE. Don't worry if you are unable to get the visualisation running. (Note: If you are working on the CSE Lab machines, you may need to use pip3 in order to get matplotlib installed correctly).

Hints:

1. [i for i in range(5) if i != 2] produces the list [0,1,3,4], something like this will be useful for extracting the CBOW input from the full context_window
2. generate_batch should be slightly simpler than the skip-gram version, since only one (input, output) pair needs to be created for each context window.
3. get_mean_context_embeds(...) should only require about 2 lines of TensorFlow code.

Submitting Stage 1

give cs9444 hw2 word2vec_fns.py CBOW_Embeddings.npy

The submission deadline for Stage 1 is Sunday 8 October.

Stage 2: Sentiment Classifier

implementation.py	this is a skeleton file for your RNN classifier implementation
train.py	file that calls implementation.py and trains your sentiment model

data = open("/home/cs9444/public_html/17s2/hw2/glove.6B.50d.txt",'r',encoding="utf-8")

In this assignment, unlike assignment 1, the network structure is not specified, and you will be assessed based on the performance of your final classifier.

There are very few constraints on your model - it must only use some form of recurrent unit (tanh, LSTM, GRU etc.) and be trained by the code provided.

So that no one has an advantage due to access to better hardware, we have provided the train.py file. On submission, you should train your model using an unedited version of train.py (this ensures the same number of data presentations for everyone's model). While an unchanged file must be used during submission, you are encouraged to make changes to this file during development. You may want to track additional tensors in tensorboard, or serialize training data so that the word embedding is not called on each run. It would also be highly advisable to split the data into a training and validation set to ensure your model does not overfit.

You must make use of recurrent network elements in your final solution. Aside from the fact that this is the type of network this assessment aims to assess, for text classification some recurrency will be important. Consider the review fragment; "I really thought this was a great example of how not to make a movie.". A naive classifier (e.g. a feed forward network trained on word counts) would be unable to correctly identify the sentiment as it depends on the tail end of the review being understood in the context of the "not" negation. Recurrent units allow us to preserve this dependency as we parse the review.

During testing, we will load your saved network from a TensorFlow checkpoint (see: the TensorFlow programmers guide to saved models). To allow our test code to find the correct path of the graph to connect to, the following naming requirements must be implemented.

1. Input placeholder: name="input_data"
2. labels placeholder: name="labels"
3. accuracy tensor: name="accuracy"
4. loss tensor: name="loss"

Code Structure

This file contains the training code for the model to be defined in implementation.py
It calls functions to load the data, convert it to embedded form, and define the model structure. It then runs 100000 training iterations, with whatever batch size is defined in implementation.py.
Accuracy and loss values are printed to the terminal every 50 iterations, and are also saved as tensorboard summaries in a created tensorboard directory. The model is saved every 10000 iterations. These model files are saved in a created checkpoints directory, and should consist of a checkpoint file, plus three files ending in the extentions .data-00000-of-00001, .index and .meta .

This is the model that must be submitted, and will be used for marking.

implemention.py

This is where you should implement your solution. This file must contain three functions: load_glove_embeddings(), load_data() and define_graph()

load_glove_embeddings() should load the word embedding vectors found in glove.6B.50d.txt and return the vectors in array form, and a dictionary matching words to index's in that array. The array should contain one vector per row, and the dictionary should have words in string form as keys, and index's as values. For example, in the provided file, the first word vector is for "the". On loading this first vector, the values themselves (0.418 0.24968 -0.41242 0.1217 etc.) should be put in the first row of the array. A new entry in the dictionary should then be added: {"the":0}. This tells us that the vector values for the word "the" are located in the first row of our embeddings array. There should also be a 0 vector as the first entry, with it's associated key being 'UNK' - this will be used for unknown words.

load_data(glove_dict) should load the training data found in reviews.tar.gz into a numpy array that can be used for training. Reviews should take up one row of the array, which must be capped at 40 colums. Each element should contain the index of the word from that review in the generated embeddings array. If reviews are shorter than 40 words they should be 0-padded.

define_graph(glove_embeddings_arr, batch_size) This is where you should define your model. The embedding array will be required to perform an embedding lookup (tf.nn.embedding_lookup()). You are required to make use of at least one recurrent unit - either an LSTM, GRU or tanh. To ensure your model is sufficiently general (so as to achieve the best test accuracy) you should experiment with regularization techniques such as dropout. This is where you must also provide the correct names for your placeholders and variables.

There is also a global variable, batch_size which you should experiment with changing. This defines the size of the batches that will be used to train the model in train.py and may have a significant effect on model performance.

Visualizing Your Progress

python3 -m tensorflow.tensorboard --logdir=./tensorboard

tensorboard --logdir=./tensorboard

1. open a Web browser and navigate to  http://localhost:6006
2. you should be able to see a plot of the loss and accuracies in TensorBoard under the "scalars" tab

Submission

give cs9444 hw2 word2vec_fns.py CBOW_Embeddings.npy

For Stage 2, you need to submit four files:

• Your complete implementation.py file
• the (final) checkpoint files generated by train.py

give cs9444 hw2sent implementation.py *.data* *.index *.meta

You can submit as many times as you like - later submissions will overwrite earlier ones, and submissions by either group member will overwrite those of the other. You can check that your submission has been received by using the following command:

9444 classrun -check

The submission deadline for Stage 1 is Sunday  8 October, 23:59.
The submission deadline for Stage 2 is Sunday 15 October, 23:59.
15% penalty will be applied to the (maximum) mark for every 24 hours late after the deadline.

Additional information may be found in the FAQ and will be considered as part of the specification for the project.

Questions relating to the project can also be posted to the Forums on the course Web page.

If you have a question that has not already been answered on the FAQ or the MessageBoard, you can email it to alex.long@student.unsw.edu.au

You should always adhere to good coding practices and style. In general, a program that attempts a substantial part of the job but does that part correctly will receive more marks than one attempting to do the entire job but with many errors.

Plagiarism Policy

Your program must be entirely your own work. Plagiarism detection software will be used to compare all submissions pairwise and serious penalties will be applied, particularly in the case of repeat offences.

DO NOT COPY FROM OTHERS; DO NOT ALLOW ANYONE TO SEE YOUR CODE

Please refer to the UNSW Policy on Academic Honesty and Plagiarism if you require further clarification on this matter.

Good luck!