Original subtitles

In this lecture, we will begin looking at our question answering notebook.

As usual, we begin by installing transformers and data sets.

The next step is to input the function load data set.

We then call this function passing in the string squad.

We'll call the result of raw data sets.

So note that the data set already becomes split into two parts one for train and one for validation.

The train set has about 88,000 samples, while the validation set has about a 10,000.

Note that each data set it comes with the columns ID title, context, question and answers.

So far in this course, we haven't had much a need for the ID column, but in this section it will be

crucial.

On the other hand, we won't be using the title.

Let's look at one of the titles anyway just to see what it is.

As you can see, it says University of Notre Dame.

Now let's look at the context for the same sample.

As you can see, it's the same context I showed you in the previous lecture.

Now let's look at the corresponding question.

So it says what is in front of the Notre Dame main building.

Finally, let's look at the answers.

So the answer is a copper statue of Christ.

Now, as you've seen, the answers are stored in a list and the column is the plural answers indicating

that there can be multiple answers per input.

I stated that although this is always possible, it is not the case for the train set.

So in this line we simply check how many samples in the train set.

Have a list of answers with any length, not equal to one.

So as you can see, after we apply this filter, we find that zero samples have an answer list with

length and not equal to one.

This means that for the train set, every input has precisely one answer.

The next step is to check out one of the answers from the validation set.

As you can see, this particular sample has three answers.

So for the validation set, it is possible for one question to have multiple answers.

Let's now check the corresponding context to see why this question may have multiple answers.

So we find the text.

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

Clara, California.

This explains the three possible answers we saw above.

In fact, there are probably even more valid answers.

For instance, we could say San Francisco Bay Area or simply Bay Area.

Although these are not in the data set.

Now just for completion sake.

Let's check out the question.

As you can see, the question is, as expected, where did Super Bowl 50 take place?

Now, one weird aspect of this data set is that for some cases where there are multiple answers, the

answers are actually all the same.

Let's check the first sample to see an example.

So as you can see, we get Denver Broncos three times.

Now, you might think that they could be different answers if the same string shows up multiple times

in the context.

But in this case, we see that the answer is START index is the same in all three cases.