More settings for retrofitting and fine-tuning

[jxmorris12]

This is Table 2 from Retrofitting Contextualized Word Embeddings with Paraphrases.

We're missing a couple of the row dimensions (new settings for retrofitting):

• MRPC dataset
• PAN dataset
• Ability to combine multiple things together (Quora + MRPC + PAN)

And most of the column dimensions, various fine-tuning tasks:

• MPQA
• CR (what is this?)
• SST-2
• SICK-R
• {STS-15, STS-16, STS-Benchmark}
• MRPC
• SICK-E

We certainly don't need to support every scenario, but more would be nice. They claim retrofitting with Sampled Quora gives a ~4% boost in accuracy on MPQA so that could be a good place to start.

[jscuds]

First Steps

1. I'll start w/ MPQA. I'm going to use Version 3.0 (came out in 2015 so I think it's a reasonable assumption that this is the one Shi used). The site for the dataset is here [link], and I confirmed it's not on HF datasets via this GH Issue (which I saw you had some comments on from 18 months ago, haha).
2. I think MRPC in datasets is structured similarly to quora/('glue','qqp'). The ParaphraseDatasetElmo framework should just require me to add a few additional lines to allow for MRPC in retrofitting.
3. CR is "Customer Review Datasets." I found the authors' website and under the heading "Data Sets" they have a 5 product review and a 9 product review -- without specification by Shi I'll use the 5 product review.
4. As I come across additional datasets that aren't included in HF's library, I'll put the raw zip files on my Google Drive in a datasets directory

[jxmorris12]

I think the things we added are sufficient at this point