(demo) FactCC & Frank & QAGS results on pairwise MNLI model metrics (sample results using df[:10])

[lihebi]

For frank, we have multiple system summaries for a doc, so we report regular correlation results in terms of:

• system-level
• summary-level

For qags-cnndm and factCC datasets, we don't have docID in the input data. Instead, we only have (doc, sum, human_score). Thus, we will report:

• system-level
• pool-level: where all rows are assigned the same ID; thus all rows are in one large batch to compute correlation. The file name is still summary-level. We won't have summary-level results as there's only one system.

The data is obtained with g2/env.py, i.e., it contains both bertscore and mnli models as the metrics.

The sample results are using the first 10 rows of the dataframe (df[:10]), just for demo purposes. The full experiments are running.