Bug fix CNN/DM and XSum initialization

[alejorba]

This PR builds on the changes made in PR #16, where a bug related to the CNNDM variable in summac/benchmark.py was identified.

Specifically, when the CNNDM variable is undefined, lines 44 and 54 can raise a NameError: name 'CNNDM' is not defined:

summac/summac/benchmark.py

Lines 41 to 61 in 9e4f357

	def get_cnndm_document(self, aid):
	global CNNDM
	if self.cnndm is None:
	if CNNDM is None:
	CNNDM = load_dataset("cnn_dailymail", "3.0.0")
	self.cnndm = CNNDM
	self.cnndm_id2article = {}
	for cut in ["test", "validation"]:
	self.cnndm_id2article.update({d["id"]: d["article"] for d in self.cnndm[cut]})
	return self.cnndm_id2article[aid]
	def get_cnndm_reference(self, aid):
	global CNNDM
	if CNNDM is None:
	CNNDM = load_dataset("cnn_dailymail", "3.0.0")
	self.cnndm = CNNDM
	if self.cnndm_id2reference is None:
	self.cnndm_id2reference = {}
	for cut in ["test", "validation"]:
	self.cnndm_id2reference.update({d["id"]: d["highlights"] for d in self.cnndm[cut]})
	return self.cnndm_id2reference[aid]

The fix proposed by @forrestbao effectively resolve this issu, but during testing I noticed some performance concerns.>

To address this, I propose an alternative fix that leverages class variables. Through this new approach:

• The CNN/DM and XSum datasets are only loaded when an instance of the SummaCBenchmark class is created.
• The datasets are loaded only once and reused across all instances of the class.

In addition, the GDrive link for the SummEval dataset provided in lines 271-276 (apparently from the 4/19/2020 update on the README.md file of the original repo https://github.com/Yale-LILY/SummEval) is broken.

summac/summac/benchmark.py

Lines 271 to 276 in 9e4f357

	if not os.path.exists(dataset_folder):
	print("==== SummEval dataset not found, downloading from scratch")
	os.makedirs(dataset_folder)
	# From the 4/19/2020 update on the README: https://github.com/Yale-LILY/SummEval
	download_file_from_google_drive("1d2Iaz3jNraURP1i7CfTqPIj8REZMJ3tS", fn)

I replaced it with a valid GCS bucket link that can be found on the README.md file of the same repo under the "Human annotations" header (https://storage.googleapis.com/sfr-summarization-repo-research/model_annotations.aligned.jsonl)

Changes

• Fixed NameError by modified dataset loading in the SummaCBenchmark class.
• Replaced broken GDrive link with a working GCS bucket link, for the SummEval dataset.