Add disinformation tests

[dcecchini]

Check metrics from HELM.

Are there curated datasets available?

Metrics:

• Self-BLEU
• Entropy (Monte Carlo)

[ArshaanNazir]

https://github.com/stanford-crfm/helm/blob/main/src/helm/benchmark/scenarios/disinformation_scenario.py

[RakshitKhajuria]

@dcecchini,

After going through the paper Truth, Lies, and Automation and the link provided above

These were my findings:

1. They are using two DisinformationScenario -> Narrative Reiteration and Narrative Wedging
2. In Narrative Reiteration test - It tests the ability of models to generate new headlines that promote a given narrative (whether the narrative is true or fase).
3. GPT-3 performed well, easily generating tweet-length messages (as per the paper) that fit the desired theme and did not directly repeat the examples. It can produce similar quality of outputs with similarly low levels of effort for almost any topic or argument whether it is a propaganda or not.

If given GPT-3 inputs and outputs for climate denialist tweets how are we going to evaluate the results if they are misleading or not can you provide us with an example, it would be helpful...

[dcecchini]

Hi @RakshitKhajuria ,

On the Github link, in the docstring of the scenario, it is mentioned that the models were manually evaluated. In our case, we want to automate that and we can think of a few possibilities:

• Use the embedding distance to measure if the prompt is too different than the generated text (if the model goes in line with what is asked) -- need experimentation, as we saw for the clinical tests that it may not always work
• Use a model like T5 for the MNLI task (if the sentences are contradictory or not)
• Use rules to determine if the generated text follows the prompt or not (using keywords for example, which would be created by ourselves for each prompt)

As a first step, please find the prompts used, and generate the texts for at least 3 models, then we can run some experiments.