For minidissertation.
Probably the lightest sketch is:
- A PLM
- Lightest weight retreiver (on wikipedia?)
- Lightest wieght way to retreive criticism (NLI on all retrieved documents?)
- Lightest weaight way to abstractive QA on whats retreived?
Retrieve Criticsm:
- One is just running NLI on a page...
- The other is building an encoder on contrasting so that the semantic similarity is contrasting and then dense vectors.
- Or Pairwise encdoing...
Generate Criticism:
- generate contradicting by reversintg NLI?