DeepQA is currently pre-alpha. There are no stable APIs, and there are plenty of rough edges. This grew out of some research code, and we think it'll be pretty useful generally, but we're still working on making it easily usable by people outside of our group. Feel free to submit issues for problems that arise so that we're aware of them, but we're not to the point of having a supported release yet.

We will have our first supported release in September.

DeepQA is a library for doing high-level NLP tasks with deep learning, particularly focused on various kinds of question answering. DeepQA is built on top of Keras and TensorFlow, and can be thought of as an interface to these systems that makes NLP easier.

Specifically, this library provides the following benefits over plain Keras / TensorFlow:

• It is easy to get NLP right in DeepQA.
  • In Keras, there are a lot of issues around padding sequences and masking that are not handled well in the main Keras code, and we have well-tested code that does the right thing for, e.g., computing attentions over padded sequences, padding all training instances to the same lengths (possibly dynamically by batch, to minimize computation wasted on padding tokens), or distributing text encoders across several sentences or words.
  • DeepQA provides a nice, consistent API around building NLP models. This API has functionality around processing data instances, embedding words and/or characters, easily getting various kinds of sentence encoders, and so on. It makes building models for high-level NLP tasks easy.
• DeepQA provides a clean interface to training, validating, and debugging Keras models. It is easy to experiment with variants of a model family just by changing some parameters in a JSON file. For example, the particulars of how words are represented, either with fixed GloVe vectors, fine-tuned word2vec vectors, or a concatenation of those with a character-level CNN, are all specified by parameters in a JSON file, not in your actual code. This makes it trivial to switch the details of your model based on the data that you're working with.
• DeepQA contains a number of state-of-the-art models, particularly focused around question answering systems (though we've dabbled in models for other tasks, as well). The actual model code for these systems is typically 50 lines or less.

DeepQA is built using Python 3. The easiest way to set up a compatible environment is to use Conda. This will set up a virtual environment with the exact version of Python used for development along with all the dependencies needed to run DeepQA.

1. Download and install Conda.

2. Create a Conda environment with Python 3.

   conda create -n deep_qa python=3.5
   

3. Now activate the Conda environment.

   source activate deep_qa
   

4. Install the required dependencies.

   ./scripts/install_requirements.sh
   

5. Set the PYTHONHASHSEED for repeatable experiments.

   export PYTHONHASHSEED=2157
   

You should now be able to test your installation with pytest -v. Congratulations! You now have a development environment for deep_qa that uses TensorFlow with CPU support. (For GPU support, see requirements.txt for information on how to install tensorflow-gpu).

To train or evaluate a model using a clone of the DeepQA repository, the recommended entry point is to use the run_model.py script. The first argument to that script is a parameter file, described more below. The second argument determines the behavior, either training a model or evaluating a trained model against a test dataset. Current valid options for the second argument are train and test (omitting the argument is the same as passing train).

Parameter files specify the model class you're using, model hyperparameters, training details, data files, data generator details, and many other things. You can see example parameter files in the examples directory. You can get some notion of what parameters are available by looking through the documentation.

Actually training a model will require input files, which you need to provide. We have a companion library, DeepQA Experiments, which was originally designed to produce input files and run experiments, and can be used to generate required data files for most of the tasks we have models for. We're moving towards putting the data processing code directly into DeepQA, so that DeepQA Experiments is not necessary, but for now, getting training data files in the right format is most easily done with DeepQA Experiments.

If you are using DeepQA as a library in your own code, it is still straightforward to run your model. Instead of using the run_model.py script to do the training/evaluation, you can do it yourself as follows:

from deep_qa import run_model, evaluate_model, load_model, score_dataset

# Train a model given a json specification
run_model("/path/to/json/parameter/file")


# Load a model given a json specification
loaded_model = load_model("/path/to/json/parameter/file")
# Do some more exciting things with your model here!


# Get predictions from a pre-trained model on some test data specified in the json parameters.
predictions = score_dataset("/path/to/json/parameter/file")
# Compute your own metrics, or do beam search, or whatever you want with the predictions here.


# Compute Keras' metrics on a test dataset, using a pre-trained model.
evaluate_model("/path/to/json/parameter/file", ["/path/to/data/file"])

The rest of the usage guidelines, examples, etc., are the same as when working in a clone of the repository.

To implement a new model in DeepQA, you need to subclass TextTrainer. There is documentation on what is necessary for this; see in particular the Abstract methods section. For a simple example of a fully functional model, see the simple sequence tagger, which has about 20 lines of actual implementation code.

In order to train, load and evaluate models which you have written yourself, simply pass an additional argument to the functions above and remove the model_class parameter from your json specification. For example:

from deep_qa import run_model
from .local_project import MyGreatModel

# Train a model given a json specification (without a "model_class" attribute).
run_model("/path/to/json/parameter/file", model_class=MyGreatModel)

If you're doing a new task, or a new variant of a task with a different input/output specification, you probably also need to implement an Instance type. The Instance handles reading data from a file and converting it into numpy arrays that can be used for training and evaluation. This only needs to happen once for each input/output spec.

DeepQA has implementations of state-of-the-art methods for a variety of tasks. Here are a few of them:

• The attentive reader, from Teaching Machines to Read and Comprehend, by Hermann and others
• Gated Attention Reader from Gated Attention Readers for Text Comprehension,
• Bidirectional Attention Flow, from Bidirectional Attention Flow for Machine Comprehension,

• Decomposable Attention, from A Decomposable Attention Model for Natural Language Inference,

This code allows for easy experimentation with the following datasets:

• AI2 Elementary school science questions (no diagrams)
• The Facebook Children's Book Test dataset
• The Facebook bAbI dataset
• The NewsQA dataset
• The Stanford Question Answering Dataset (SQuAD)
• The Who Did What dataset

Note that the data processing code for most of this currently lives in DeepQA Experiments, however.

If you use this code and think something could be improved, pull requests are very welcome. Opening an issue is ok, too, but we can respond much more quickly to pull requests.

• Matt Gardner
• Mark Neumann
• Nelson Liu.
• Pradeep Dasigi (the initial author of this codebase)

This code is released under the terms of the Apache 2 license.