In this question, you will learn to implement the Abstractive Summarization task using the encoder$-$decoder model.

$1.$ Load the dataset cnn$-$dailymail $\backslash (\{\}^\{1\}\backslash ).$ Use version $3\mathrm{.}0\mathrm{.}0.$ Use train set for training and test set for testing purposes. Use columns article which is the news article published in CNN and Daily Mail, and highlights which is the summary of the given news article.

$2.$ Preprocess the article column and the highlights column with NLTK$.$ The preprocessing steps include converting the text to lower case, removing special characters and punctuations. Add and tokens at the start and the end of each row of the highlights text.

$3.$ Tokenize the text in both columns and build a separate vocabulary for each column. Create the word$-$to$-$index and index$-$to$-$word dictionaries for article column and convert the text in article column to the index vector.

$4.$ You should use an encoder$-$decoder architecture to generate the text. The encoder should be a RNN $($e$.$g$.,$ LSTM$,$ GRU, etc.$)$ and the decoder should be a fully connected $($dense$)$ layer. Note that the initial hidden states for the decoder will be the output hidden states of the encoder model. Also, the input sequence to the decoder will have the same length that of the target sequence starting with $.$ In the inference stage, the process will run till token is encountered.

$5.$ Train the model with cross$-$entropy loss. On the test set, generate the summary using beam search. Your model should be able to generate text of at least $10$ words.

$6.$ Use the rouge module in PyTorch $\backslash (\{\}^\{2\}\backslash ),$ calculate and report the average ROUGE$-1,$ ROUGE$-2,$ and ROUGE$-$L scores for the test set.