## Neural machine translation | Coursework 2

We have looked at a variety of probabilistic models of translation. Neural machine translation models are the latest in a long line of such models. Notwithstanding any hype you may have heard about them, they are still fundamentally probabilistic models of discrete input and output sequences. However, they differ from previous generations of probabilistic translation models in some important ways.

1. Rather than work with discrete distributions directly parameterized as tables of real numbers, conditional distributions are created by first using a function to encode the discrete input sequence into a vector of continuous values (real numbers). Output is then decoded by sampling from a probability distribution that is constructed as a function of this continuous vector.

2. The encoder and decoder are simply functions composed from simple matrix operations like addition, multiplication, and pointwise transformations such as tanh, exponentiation, and scalar division. For example, the softmax operation converts any vector to a probability distribution by exponentiating every element, summing the result, and pointwise dividing the exponentiated elements by the sum.

3. These functions are parameterized by matrices, which can be learned using gradient-based optimization algorithms.1 The gradients are computed by differentiating the composed functions using the chain rule for derivatives (aka backpropagation). This is quite easy to automate, so instead of computing the gradients by hand after defining the model, we use software libraries for automatic differentiation.2 This means that the translation function is implemented declaratively, and much of the difficult work is performed by libraries. The entire function is specified, and then it is learned end-to-end: all parameters are trained simultaneously using a single algorithm. This is quite different from phrase-based models, in which a sequence of separate models are trained individually and then combined into a final model, which must also be trained.

Your task is to implement a neural machine translation pipeline by extending a simple baseline model, closely related to the neural language model you worked with in lab 2. In each part of the coursework you will be asked to implement a different extension.

IMPORTANT: Each extension will require you to train a completely new neural machine translation model from scratch. While implementing these changes may only take you a few minutes or hours, training the new models will take you A LONG TIME. You might implement something in thirty minutes and leave it to train overnight. Imagine that you return the next morning to find it has a bug! If the next morning is the due date, then you'll be in a pickle, but if it's a week before the due date, you have time to recover. So, if you want to complete this coursework on time, start early. I will not take pity on you if you start too late.

## Getting started [30 marks]

If you haven’t yet completed lab 2, you should do so now. Don’t simply click through the notebook: familiarize yourself with the code and make sure you understand what it does. This coursework uses similar code, and also relies on the same environment mtenv that you set up for lab 2, so pay particular attention to the setup instructions in the readme.

Get the code.

git clone https://github.com/INFR11133/hw2.git


You’ll find a directory data containing English and Japanese parallel data (from a tutorial that you may find helpful), a model directory containing a pretrained Japanese-to-English neural translation model, and three python files:

1. nmt_config.py contains configuration parameters, some of which you will be asked to experiment with. In particular, pay close attention to the model and training parameters towards the end of the file. You may also adjust the gpuid parameter if you have access to a GPU, which will make training times faster (but they will still take considerable time to train, so you should give yourself plenty of time even if you have a GPU).

2. nmt_translate.py trains a translation model (if one is not already present) and translates test sentences. You will not need to modify this script.

3. enc_dec.py contains a simple translation model, specified using the chainer library used in lab 2. You will need to understand and modify this code as instructed below in order to complete the coursework. Doing so should give you a good idea of how a neural machine translation system works, at a level of detail that you can’t get from lectures or reading.

To get started, run the following commands:

cd translate
source activate mtenv

ipython
In [1]: %run nmt_translate.py


You should see a message indicating that a pre-trained model has been loaded. To translate using this model, do:

In [2]: _ = predict(s=10000, num=10)


This displays translations of the first 10 japanese sentences of the dev set. To view predictions on training set do:

In [3]: _ = predict(s=0, num=10)


Most of these translations will be poor. To find better translations from this model, we can add filters based on precision and recall of each translation with respect to a reference translation. The following statement will only display predictions with recall >= 0.5

In [4]: _ = predict(s=10000, num=10, r_filt=.5)


The following statement will only display predictions with precision >= 0.5

In [5]: _ = predict(s=10000, num=10, p_filt=.5)


This model is still quite basic and trained on a small dataset, so the quality of translations is poor. Your goal will be to see if you can improve it.

The current implementation in enc_dec.py encodes the sentence using a bidirectional LSTM: one passing over the input sentence from left-to-right, the other from right-to-left. The final states of these LSTMs are concatenated and fed into the decoder, an LSTM that generates the output sentence from left-to-right. The encoder is essentially the same as the encoder described in Section 3.2 of the 2014 paper that now forms the basis of most current neural MT models. The decoder is simpler than the one in the paper (it doesn’t include the context vector described in Section 3.1), but you’ll fix that in Part 3.

Before we go deeply into modifications to the translation model, it is important to understand the baseline implementation, the data we run it on, and some of the techniques that are used to make the model run on this data.

Q1. [10 marks] The file enc_dec.py contains explanatory comments to step you through the code. Five of these comments (A-E) are missing, but they are easy to find: search for the string __QUESTION in the file. A sixth comment (F) is missing from file nmt_translate.py. For each of these cases, please (1) add explanatory comments to the code, and (2) copy your comments to your answer file (we will mark the comments in your answer file, not the code, so it is vital that they appear there). If you aren’t certain what a a particular function does, refer to the chainer documentation. (However, explain the code in terms of its effect on the MT model; don’t simply copy and paste function descriptions from the documentation).

In preparing the training data, word types that appear only once are replaced by a special token, _UNK. This prevents the vocabulary from growing out of hand, and enables the model to handle unknown words in new test sentences (which may be addressed by postprocessing). But what effect does this replacement have on the properties of our language data?

Q2. [10 marks] Examine the parallel data and answer the following questions.

1. Plot the distribution of sentence lengths in the English and Japanese and there corrletion. What do you infer from this about translating between these languages?
2. How many word tokens are in the English data? In the Japanese?
3. How many word types are in the English data? In the Japanese data?
4. How many word tokens will be replaced by _UNK in English? In Japanese?
5. Given the observations above, how do you think the NMT system will be affected by differences in sentence length, type/ token ratios, and unknown word handling?

Q3. [10 marks] What language phenomena might influence the severity of these effects you observed above? Any claims you make must be supported by evidence, in the form of statistics, known facts about language, and/ or examples. Unsupported answers will receive a mark of zero!

In answering question 3, you may find it useful to refer to The world atlas of language structures.

## Part 2: Exploring the model [30 marks]

Let’s first explore the decoder. It makes predictions one word at a time from left-to-right, as you can see by examining the function decoder_predict in the file enc_dec.py. Prediction works by first computing a distribution over all possible words conditioned on the input sentence. We then choose the most probable word, output it, add it to the conditioning context, and repeat until the predicted word is an end-of-sentence token (_EOS).

Q4. [10 marks] Modify the select_word function in the decoder to sample from the probability distribution at each time step, rather than returning the most probable word (this change should only require a few lines of code). Then sample a few translations for the dev data. These are alternatives to the one the decoder chooses.

1. What conclusions can you draw about the translation model based on this sample? Remember to support your claims with examples.
2. Return the decoder to its original state of always choosing the maximum-probability word at each time step. This is a greedy decoder. How would you modify this decoder to do beam search—that is, to consider multiple possible translations at each time step—as you did for a phrase-based decoder in coursework 1. NOTE: You needn’t implement beam search. The purpose of this question is simply for you to think through and clearly explain how you would do it.
3. Could you implement dynamic programming for this model? Why or why not? Again, you needn’t implement this.

The next two questions ask you to modify the model and retrain it. Implementing the modifications will not take you very long, but retraining the model will.

NOTE. I recommend that test your modifications by retraining on a small subset of the data (e.g. a thousand sentences). To do that, you should change the USE_ALL_DATA setting in nmt_config.py file to False. The results will not be very good; your goal is simply to confirm that the change does not break the code and that it appears to behave sensibly. This is simply a sanity check, and a useful time-saving engineering test when you’re working with computationally expensive models like neural MT. For your final models, you should train on the entire training set.

Q5. [10 marks] Experiment with one of the following changes to the model, and explain how it affects the perplexity, BLEU, and the actual translations your system produces, compared to the baseline model you were given. (In explaining the changes, be sure to include examples). Your answer should precisely specify how you changed the model—for example, if you change the number of layers, state the number you used.

1. Change the number of layers in the encoder, decoder, or both.
2. Change the number of hidden units by a substantial amount (e.g. by halving or doubling the number, not adding or subtracting one).

To train a new model, you have to modify nmt_config.py with your required settings - the number of layers you wish to use, layer width, number of epochs and a name for your experiment.

As an example, let’s define a new model with the size of hidden units in the LSTM(s) as 100, and 2 layers for both the encoder and the decoder:

# number of LSTM layers for encoder
num_layers_enc = 2
# number of LSTM layers for decoder
num_layers_dec = 2
# number of hidden units per LSTM
# both encoder, decoder are similarly structured
hidden_units = 100


And set the number of epochs equal 1 or more (otherwise the model will not train):

# Training EPOCHS
NUM_EPOCHS = 10


To start training a model with updated parameters execute the bash script:

./run_exp.bat


After each epoch, the model file is saved to disk. The model file name includes the parameters used for training. As an example, with the above settings, the model and the log file names will be:

model/seq2seq_10000sen_2-2layers_100units_{EXP_NAME}_NO_ATTN.model
model/train_10000sen_2-2layers_100units_{EXP_NAME}_NO_ATTN.log


Q6. [10 marks] An important but simple technique for working with neural models is dropout, which must be applied in a particular way to our model. Implement the method of dropout described in this paper. This change should require no more than one or two lines, but will test your understanding of the code (because you need to identify where it should go). Retrain you model. How does dropout affect the results, compared to the baseline? As in the previous question, your answer should explain the changes to perplexity, BLEU, and the translations themselves. You should also explain where you added dropout to the code and what parameters you used for it.

## Part 3: Attention [40 marks]

The last change you will implement is to augment the encoder-decoder with an attention mechanism. For this, we expect you to use a very simple model of attention, global attention with dot product, as described in this paper. This is the simplest model of attention, and reasonably effective in many settings. As a practical matter, at each time step it requires you to take the dot product of the decoder hidden state with the hidden state of each input word (itself the concatentation of forward and backward encoder LSTM hidden states). The results should be passed through the softmax function (i.e. exponentiated and normalized) and the resulting distribution should be used to interpolate the input hidden states to produce a context vector used as additional input to the decoder.

Q7. [20 marks] Implement the attention model described above. You will find placeholders in the code to save the hidden states of the encoder and return an array of attention weights. Using this API will help ensure that your code works correctly.

Q8. [10 marks] Retrain your decoder, and again explain how the change affects results compared to the baseline in terms of perplexity, BLEU, and the actual translations.

Q9. [10 marks] Visualize the evolution of the attention vectors for five decoded sentences, using the provided code. Do they seem reasonable? Why or why not? Base your argument in evidence from the data. You’ll need to understand the Japanese words to do this effectively (use Google Translate).

We provide a function, plot_attention to plot attention vectors. A plot will be generated and stored in the model directory. To do this set plot=True in the predict function.

_ = predict(s=10000, num=1, plot=True)


This will output a heatmap of the attention vectors and save the plot as model/sample_10000_plot.png

## Possible Extensions

Neural machine translation is an extremely active area of research, and this coursework has only introduced the basic ideas. Now that you have a working encoder-decoder model with attention, you may want to experiment with it further. Here are some ideas.

Minibatches
In lab 2 we arbitrarily picked 32 characters to be the length of the sequences we feed into the model during training. In the current task, however, we are not free to do so. The sequence length is fixed for each example and equals to the length of the source sentence plus the length of the translation plus one (for the end-of-sentence symbol). The sentences in our corpus are not of equal lengths, and therefore the length of the input sequences is variable. Variable-length input is, in theory, not a problem for a recurrent neural network, and both the encoder and the decoder are RNNs. During training, the learning sequences can be processed individually, and the weights updates after each one. In practice, training on just one sequence at a time in not efficient, and it’s preferable to train on batches of examples. All sequences in a batch are processed in parallel, and the weights are updated using loss information from the whole batch. All inputs in a batch have to be of the same length, to take advantage of operationalizing feedforward and feedback computations as matrix multiplications. As we established, the lengths of our sequences are not constant, so something needs to be done if we want to use a batch training approach. When implementing your model you will need to decide how to deal with this problem. Possible solutions include: finding the longest training example in the corpus and pad all shorter examples to its length deciding on a maximum example length, perhaps after inspecting the distribution of lengths in the corpus, and truncating any examples longer that that creating batches by bundling sentences of the same length, to eliminate or limit the extent of padding or truncating. An effective minibatch design will make training substantially faster, making it possible to do more experiments.
Understanding word embeddings
As a byproduct of training the translation models, we also learn embeddings. This question is about exploring them to assess how well they do at capturing lexical meaning, for example by measuring morphosyntactic or semantic similarity of words that have similar embeddings in continuous space.
Out-of-vocabulary words
I hope you found our approach to dealing with out-of-vocabulary words inelegant (replacing them all with _UNK). Consider some ways to improve this aspect of the model.

You can find many more ideas in the recent research literature. There are also many unsolved problems! To get some idea where they are, look for examples of incorrect translations. For each step in the decoder, explore the probability distribution over the English vocabulary. Extract the top k most probable translations and see if the correct word is amongst them. How far dow the list is it? This is a way of assessing just how wrong your model is. Do the same for several correctly translated sentences. In this case, you want to see how peaked the probability distribution is at each step in the decoder. The more peaky a distribution, the more certain the model is about the choice of next output word.

## Ground Rules

• You must work individually. If you submit work from someone other than yourself you will receive a mark of zero for the assignment. Your code and report must be your own work. You can safely assume that your instructor has software to accurately compute the probability that one piece of code is a modified copy of the other. On the other hand, sharing questions, clarifications, and ideas that stop short of actual answers is fine and encouraged, especially through the forum, since articulating your questions is often a good way to figure out what you do and don’t know.

• You must submit these files and only these files.
1. answers.pdf: A file containing your answers to Questions 1 through 9 in an A4 PDF. Your file must be written in LaTeX using the overleaf template, which you should clone and edit to provide your answers. Answers provided in any other format will receive a mark of zero. Your answers must not exceed three pages, so be concise. You are permitted to include graphs or tables on an unlimited number of additional pages. They should be readable. They should also be numbered and the text should refer to these numbers.
2. attention.py: Your modified version of enc_dec.py including both dropout and attention (or whichever of these you complete, if you don’t complete the assignment).
3. translations.txt: The output of your final model on the test set. Your answers to questions 8 and 9 should refer to translations in this file.
• Your name must not appear in any of the submitted files. If your name appears in the code or pdf (or output) you will receive a mark of zero.

To submit your files on dice, run:

submit mt 2 answers.pdf attention.py translations.txt


### Credits

This assignment was developed by Sameer Bansal and Ida Szubert, with occasional meddling from Adam Lopez.

### Footnotes

1. The current favored optimizer is the excellently named Adam.

2. This may seem obvious in 2017. But less than a decade ago, automatic differentiation was so uncommon in machine learning that one researcher called it the most criminally underused tool in the potential machine learning toolbox

Informatics Forum, 10 Crichton Street, Edinburgh, EH8 9AB, Scotland, UK
Tel: +44 131 651 5661, Fax: +44 131 651 1426, E-mail: school-office@inf.ed.ac.uk