Distantly supervised labels for BW dataset

[leonhardhennig]

Think of a way to distantly / heuristically label sentences in the Businesswire dataset (after NER validation by Crowdee, not yet done) with the following relation types:

• org:facility_or_location (org, fac/loc) - see maybe Docred P706? P276, FewRel P276 relation types
• loc:event_or_disaster (loc, disaster-type/event)
• org:insolvency (org, cause_of_insolvency) (insolvency 'trigger' is not a NER yet?)
• org:layoffs (org, loc)
• org:strike (org, loc)
• org:turnover (org, number/money) (Umsatz)
• org:revenue (org, number/money) (Gewinn)
• org:industry (org, industry) (i.e. industry branch of the company, e.g. 'IT')
• org:fin_event (org, fin_event) (financial event, i.e. sentences with ORG and FINANCIAL_EVENT)

Maybe train DISTRE on GIDS and then do predictions on BW dataset to add GIDS relation types?

@harbecke suggested to try to use prompts to pre-label BW docs?

[harbecke]

One possible approach is to write one or several prompts that capture the relation and pre-label the dataset with those. Example: How much revenue did $entity_company$ make? [input sentence] Answer span: $entity_revenue$ in input sentence. If the dataset is in TacRED format, I can do this without much effort. For examples where one entity is missing we need a strategy. (E.g questions without entity or one entity with yes or no answers)