Original subtitles

So in this lecture we will discuss the squad data set, which is the dataset we'll be using for this

task.

Squad simply stands for Stanford.

Question Answering Data Set.

Now, as you already know, the task of question answering is still constrained in a few ways.

We can't yet give a neural network a big database of knowledge and just ask it any question we want.

Instead, we do what is called extractive.

Question Answering.

What this means is that we're going to give the network a pair of texts, namely the question and the

context which contains the answer to the question.

In that way, the answer is always a substring of the context.

Note also that because of this, the network never has to actually generate any text, so we don't require

an encoder decoder setup.

Now there are some details that will become important when you want to actually write the code.

Firstly, let's look at how we will load in the data.

As you can see, we just call the standard function load data set passing in the string squad.

The data set comes with five columns which are ID title, context, question and answers.

Note that the title is pretty much irrelevant.

Interestingly, ID is not irrelevant, which may seem strange since we've ignored it up until this point.

We'll discuss this more when the time comes.

So let's look at some examples.

Here's a context.

It says, Architecturally, the school has a Catholic character.

Atop the main building is Gold Dome, is a golden statue of the Virgin Mary, and so on and so forth.

The question is what is in front of the Notre Dame main building?

And the corresponding answer is a copper statue of Christ.

Note that the answer seems to have a funny format.

Firstly, the text is stored in a list.

We'll see why that makes sense shortly.

Furthermore, we see that in addition to the text, we also get the position of the start of the answer

in terms of characters.

As you recall, a string is simply an array of characters.

So if you think of the context as an array of characters, this would be the index of the start of the

answer.

So for this example, the corresponding title happens to be University of Notre Dame.

As you can see, this is irrelevant for finding the answer to the question.

What should strike you as interesting is that the answer column is plural.

This implies that there can potentially be multiple answers to the same question.

This also explains why the answer data is stored in lists.

Now, how can this be?

Well, consider the question where did Super Bowl 50 take place?

One possible answer is Santa Clara, California.

This is a true fact.

But another possible answer is Levi's Stadium.

This is also a true fact.

So how can one question have multiple answers?

Well, here's the context where this came from.

It says The game was played on February seven, 2016 at Levi's Stadium in the San Francisco Bay area

at Santa Clara, California.

Depending on how you interpret this question, both answers would be valid.

So this is an example of where the same question can have multiple valid answers.

Now, oddly, this data set is built such that for some questions, the exact same answer can appear

multiple times.

I'm not sure why that is.

Finally, note that this only happens for the validation set for the train set.

Although the column is called Answers.

There is only one answer per sample.

Since our neural networks loss function is only built for one target per input, this is a good thing

since it means we don't have to do any extra work to split up multiple answers into separate training

samples.