Supervised sentiment: hand-built feature functions

Contents

1. Overview
2. Set-up
3. Feature functions
4. Building datasets for experiments
5. Basic optimization
   1. Wrapper for SGDClassifier
   2. Wrapper for LogisticRegression
   3. Other scikit-learn models
6. Experiments
   1. Experiment with default values
   2. A dev set run
   3. Assessing BasicSGDClassifier
   4. Comparison with the baselines from Socher et al. 2013
   5. A shallow neural network classifier
   6. A softmax classifier in PyTorch
7. Hyperparameter search
   1. utils.fit_classifier_with_crossvalidation
   2. Example using LogisticRegression
   3. Example using BasicSGDClassifier
8. Statistical comparison of classifier models
   1. Comparison with the Wilcoxon signed-rank test
   2. Comparison with McNemar's test

Overview

• The focus of this notebook is building feature representations for use with (mostly linear) classifiers (though you're encouraged to try out some non-linear ones as well!).

• The core characteristics of the feature functions we'll build here:

  • They represent examples in very large, very sparse feature spaces.
  • The individual feature functions can be highly refined, drawing on expert human knowledge of the domain.
  • Taken together, these representations don't comprehensively represent the input examples. They just identify aspects of the inputs that the classifier model can make good use of (we hope).

• These classifiers tend to be highly competitive. We'll look at more powerful deep learning models in the next notebook, and it will immediately become apparent that it is very difficult to get them to measure up to well-built classifiers based in sparse feature representations.

Set-up

See the previous notebook for set-up instructions.

Feature functions

• Feature representation is arguably the most important step in any machine learning task. As you experiment with the SST, you'll come to appreciate this fact, since your choice of feature function will have a far greater impact on the effectiveness of your models than any other choice you make.

• We will define our feature functions as dicts mapping feature names (which can be any object that can be a dict key) to their values (which must be bool, int, or float).

• To prepare for optimization, we will use sklearn's DictVectorizer class to turn these into matrices of features.

• The dict-based approach gives us a lot of flexibility and frees us from having to worry about the underlying feature matrix.

A typical baseline or default feature representation in NLP or NLU is built from unigrams. Here, those are the leaf nodes of the tree:

In the docstring for sst.sentiment_treebank_reader, I pointed out that the labels on the subtrees can be used in a way that feels like cheating. Here's the most dramatic instance of this: root_daughter_scores_phi uses just the labels on the daughters of the root to predict the root (label). This will result in performance well north of 90% F1, but that's hardly worth reporting. (Interestingly, using the labels on the leaf nodes is much less powerful.) Anyway, don't use this function!

It's generally good design to write lots of atomic feature functions and then bring them together into a single function when running experiments. This will lead to reusable parts that you can assess independently and in sub-groups as part of development.

Building datasets for experiments

The second major phase for our analysis is a kind of set-up phase. Ingredients:

• A reader like train_reader
• A feature function like unigrams_phi
• A class function like binary_class_func

The convenience function sst.build_dataset uses these to build a dataset for training and assessing a model. See its documentation for details on how it works. Much of this is about taking advantage of sklearn's many functions for model building.

Notice that sst.build_dataset has an optional argument vectorizer:

• If it is None, then a new vectorizer is used and returned as dataset['vectorizer']. This is the usual scenario when training.

• For evaluation, one wants to represent examples exactly as they were represented during training. To ensure that this happens, pass the training vectorizer to this function:

Basic optimization

We're now in a position to begin training supervised models!

For the most part, in this course, we will not study the theoretical aspects of machine learning optimization, concentrating instead on how to optimize systems effectively in practice. That is, this isn't a theory course, but rather an experimental, project-oriented one.

Nonetheless, we do want to avoid treating our optimizers as black boxes that work their magic and give us some assessment figures for whatever we feed into them. That seems irresponsible from a scientific and engineering perspective, and it also sends the false signal that the optimization process is inherently mysterious. So we do want to take a minute to demystify it with some simple code.

The module np_sgd_classifier contains a complete optimization framework, as BasicSGDClassifier. Well, it's complete in the sense that it achieves our full task of supervised learning. It's incomplete in the sense that it is very basic. You probably wouldn't want to use it in experiments. Rather, we're going to encourage you to rely on sklearn for your experiments (see below). Still, this is a good basic picture of what's happening under the hood.

So what is BasicSGDClassifier doing? The heart of it is the fit function (reflecting the usual sklearn naming system). This method implements a hinge-loss stochastic sub-gradient descent optimization. Intuitively, it works as follows:

1. Start by assuming that all the feature weights are 0.
2. Move through the dataset instance-by-instance in random order.
3. For each instance, classify it using the current weights.
4. If the classification is incorrect, move the weights in the direction of the correct classification

This process repeats for a user-specified number of iterations (default 10 below), and the weight movement is tempered by a learning-rate parameter eta (default 0.1). The output is a set of weights that can be used to make predictions about new (properly featurized) examples.

In more technical terms, the objective function is

$$ \min_{\mathbf{w} \in \mathbb{R}^{d}} \sum_{(x,y)\in\mathcal{D}} \max_{y'\in\mathbf{Y}} \left[\mathbf{Score}_{\textbf{w}, \phi}(x,y') + \mathbf{cost}(y,y')\right] - \mathbf{Score}_{\textbf{w}, \phi}(x,y) $$

where $\mathbf{w}$ is the set of weights to be learned, $\mathcal{D}$ is the training set of example–label pairs, $\mathbf{Y}$ is the set of labels, $\mathbf{cost}(y,y') = 0$ if $y=y'$, else $1$, and $\mathbf{Score}_{\textbf{w}, \phi}(x,y')$ is the inner product of the weights $\mathbf{w}$ and the example as featurized according to $\phi$.

The fit method is then calculating the sub-gradient of this objective. In succinct pseudo-code:

• Initialize $\mathbf{w} = \mathbf{0}$
• Repeat $T$ times:
  • for each $(x,y) \in \mathcal{D}$ (in random order):
    • $\tilde{y} = \text{argmax}_{y'\in \mathcal{Y}} \mathbf{Score}_{\textbf{w}, \phi}(x,y') + \mathbf{cost}(y,y')$
    • $\mathbf{w} = \mathbf{w} + \eta(\phi(x,y) - \phi(x,\tilde{y}))$

This is very intuitive – push the weights in the direction of the positive cases. It doesn't require any probability theory. And such loss functions have proven highly effective in many settings. For a more powerful version of this classifier, see sklearn.linear_model.SGDClassifier. With loss='hinge', it should behave much like BasicSGDClassifier (but faster!).

Wrapper for SGDClassifier

For the sake of our experimental framework, a simple wrapper for SGDClassifier:

Wrapper for LogisticRegression

As I said above, we likely don't want to rely on BasicSGDClassifier (though it does a good job with SST!). Instead, we want to rely on sklearn. Here's a simple wrapper for sklearn.linear.model.LogisticRegression using our build_dataset paradigm.

Other scikit-learn models

• The sklearn.linear_model package has a number of other classifier models that could be effective for SST.

• The sklearn.ensemble package contains powerful classifiers as well. The theme that runs through all of them is that one can get better results by averaging the predictions of a bunch of more basic classifiers. A RandomForestClassifier will bring some of the power of deep learning models without the optimization challenges (though see this blog post on some limitations of the current sklearn implementation).

• The sklearn.svm contains variations on Support Vector Machines (SVMs).

Experiments

We now have all the pieces needed to run experiments. And we're going to want to run a lot of experiments, trying out different feature functions, taking different perspectives on the data and labels, and using different models.

To make that process efficient and regimented, sst contains a function experiment. All it does is pull together these pieces and use them for training and assessment. It's complicated, but the flexibility will turn out to be an asset.

Experiment with default values

A few notes on this function call:

• Since assess_reader=None, the function reports performance on a random train–test split. Give sst.dev_reader as the argument to assess against the dev set.

• unigrams_phi is the function we defined above. By changing/expanding this function, you can start to improve on the above baseline, perhaps periodically seeing how you do on the dev set.

• fit_softmax_classifier is the wrapper we defined above. To assess new models, simply define more functions like this one. Such functions just need to consume an (X, y) constituting a dataset and return a model.

A dev set run

Assessing BasicSGDClassifier

Comparison with the baselines from Socher et al. 2013

Where does our default set-up sit with regard to published baselines for the binary problem? (Compare Socher et al., Table 1.)

A shallow neural network classifier

While we're at it, we might as well see whether adding a hidden layer to our softmax classifier yields any benefits. Whereas LogisticRegression is, at its core, computing

$$\begin{align*} y &= \textbf{softmax}(xW_{xy} + b_{y}) \end{align*}$$

the shallow neural network inserts a hidden layer with a non-linear activation applied to it:

$$\begin{align*} h &= \tanh(xW_{xh} + b_{h}) \\ y &= \textbf{softmax}(hW_{hy} + b_{y}) \end{align*}$$

It looks like, with enough iterations (and perhaps some fiddling with the activation function and hidden dimensionality), this classifier would meet or exceed the baseline set up by LogisticRegression.

A softmax classifier in PyTorch

Our PyTorch modules should support easy modification. For example, to turn TorchShallowNeuralClassifier into a TorchSoftmaxClassifier, one need only write a new define_graph method:

Hyperparameter search

The training process learns parameters — the weights. There are typically lots of other parameters that need to be set. For instance, our BasicSGDClassifier has a learning rate parameter and a training iteration parameter. These are called hyperparameters. The more powerful sklearn classifiers often have many more such hyperparameters. These are outside of the explicitly stated objective, hence the "hyper" part.

So far, we have just set the hyperparameters by hand. However, their optimal values can vary widely between datasets, and choices here can dramatically impact performance, so we would like to set them as part of the overall experimental framework.

utils.fit_classifier_with_crossvalidation

Luckily, sklearn provides a lot of functionality for setting hyperparameters via cross-validation. The function utils.fit_classifier_with_crossvalidation implements a basic framework for taking advantage of these options.

This method has the same basic shape as fit_softmax_classifier above: it takes a dataset as input and returns a trained model. However, to find its favored model, it explores a space of hyperparameters supplied by the user, seeking the optimal combination of settings.

Note: this kind of search seems not to have a large impact for SST as we're using it. However, it can matter a lot for other data sets, and it's also an important step to take when trying to publish, since reviewers are likely to want to check that your comparisons aren't based in part on opportunistic or ill-considered choices for the hyperparameters.

Example using LogisticRegression

Here's a fairly full-featured use of the above for the LogisticRegression model family:

Example using BasicSGDClassifier

The models written for this course are also compatible with this framework. They "duck type" the sklearn models by having methods fit, predict, get_params, and set_params, and an attribute params.

Statistical comparison of classifier models

Suppose two classifiers differ according to an effectiveness measure like F1 or accuracy. Are they meaningfully different?

• For very large datasets, the answer might be clear: if performance is very stable across different train/assess splits and the difference in terms of correct predictions has practical importance, then you can clearly say yes.

• With smaller datasets, or models whose performance is closer together, it can be harder to determine whether the two models are different. We can address this question in a basic way with repeated runs and basic null-hypothesis testing on the resulting score vectors.

In general, one wants to compare two feature functions against the same model, or one wants to compare two models with the same feature function used for both. If both are changed at the same time, then it will be hard to figure out what is causing any differences you see.

Comparison with the Wilcoxon signed-rank test

The function sst.compare_models is designed for such testing. The default set-up uses the non-parametric Wilcoxon signed-rank test to make the comparisons, which is relatively conservative and recommended by Demšar 2006 for cases where one can afford to do multiple assessments. For discussion, see the evaluation methods notebook.

Here's an example showing the default parameters values and comparing LogisticRegression and BasicSGDClassifier:

Comparison with McNemar's test

McNemar's test operates directly on the vectors of predictions for the two models being compared. As such, it doesn't require repeated runs, which is good where optimization is expensive. For discussion, see the evaluation methods notebook.