Learning to Rationalize for Nonmonotonic Reasoning with Distant Supervision

The black-box nature of neural models has motivated a line research that aims to generate natural language rationales explain why model made certain predictions. Such rationale generation models, date, have been trained on dataset-specific crowdsourced rationales, but this approach is costly and not generalizable new tasks domains. In paper, we investigate the extent which can reason about predictions, relying only distant supervision with no additional annotation cost for human-written rationales. We multiple ways automatically using pre-trained knowledge from related tasks, train generative capable composing explanatory unseen instances. demonstrate our defeasible inference task, nonmonotonic reasoning task in an may be strengthened or weakened when information (an update) introduced. Our shows promises at generating post-hoc explaining more less likely given information, however, it mostly generates trivial reflecting fundamental limitations models. Conversely, realistic setup jointly predicting update its type challenging, suggesting important future direction.