nlp-brin-id/triplets-all

We create this dataset from nlp-brin-id/id-hoax-report-merge-v2.

In this subset, triplets candidates are described as Positive sentence A; Positive Sentence B; Negative sentence C. Sampling space are defined as follows:

• [HOAX label] Title Hoax; Title Hoax; Title Non-Hoax
• [HOAX label] Title Hoax; Title Hoax; Content Non-Hoax (if not empty)
• [HOAX label] Title Hoax; Content Hoax (if not empty); Title Non-Hoax -
• [HOAX label] Title Hoax; Content Hoax (if not empty); Content Non-Hoax (if not empty)
• [HOAX label] Title Hoax; Title Hoax; Fact Hoax (if Fact != null),
• [HOAX label] Title Hoax; Content Hoax (if not empty); Fact Hoax (if Fact != null),
• [NON-HOAX label] Title Non-Hoax; Title Non-Hoax; Title Hoax
• [NON-HOAX label] Title Non-Hoax; Title Non-Hoax; Content Hoax
• [NON-HOAX label] Title Non-Hoax; Content Non-Hoax (if not empty); Title Hoax
• [NON-HOAX label] Title Non-Hoax; Content Non-Hoax (if not empty); Content Hoax (if not empty)
• [NON-HOAX label] Title Non-Hoax; Fact Non-Hoax (if not empty); Title Hoax
• [NON-HOAX label] Title Non-Hoax; Fact Non-Hoax (if not empty); Content Hoax (if not empty)
• [NON-HOAX label] Content Non-Hoax (if not empty); Fact Non-Hoax (if not empty); Title Hoax
• [NON-HOAX label] Content Non-Hoax (if not empty); Fact Non-Hoax (if not empty); Content Hoax (if not empty)

For creating the subset, we permute hard negative samples for 10 epochs dependent to the class category.
For each epoch, we flip coins to decide whether the triplet uses 'Title', 'Content' (a long description of claim in Title), or 'Fact'.
Note that: 'Fact' represents hard negative or contradicting sentence in Hoax class samples, while in Non-Hoax subset it represents supports (Positive sentence).