Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -32,7 +32,7 @@ DRUID contains real-world (query, context) pairs to facilitate studies of contex
|
|
32 |
|
33 |
## Dataset Structure
|
34 |
|
35 |
-
We release two versions of the dataset: DRUID and DRUID+.
|
36 |
|
37 |
- id: A unique identifier for each dataset sample. It also indicates claim source.
|
38 |
- claim_id: A unique identifier for each claim in the dataset. One claim may correspond to multiple samples, corresponding to different evidence pieces retrieved from different webpages.
|
@@ -48,13 +48,13 @@ We release two versions of the dataset: DRUID and DRUID+. \texttt{DRUID} is a hi
|
|
48 |
- relevant: Whether the evidence is relevant with respect to the given claim. Has been manually annotated for the DRUID samples.
|
49 |
- evidence_stance: The stance of the evidence, i.e. whether it supports the claim or not, or is insufficient. Has been manually annotated for the DRUID samples.
|
50 |
|
51 |
-
Samples based on claims from 'borderlines' follow a sligthly different structure, as the "claims" in this case are based on
|
52 |
|
53 |
## Dataset Creation
|
54 |
|
55 |
### Claim Collection
|
56 |
|
57 |
-
We sample claims verified by fact-checkers using [Google's Factcheck API](https://developers.google.com/fact-check/tools/api/reference/rest). We only sample claims in English. The claims are collected from 7 diverse fact-checking sources, representing science, politics, Northern Ireland, Sri Lanka, the US, India, France, etc. All claims have been assessed by human fact-checkers.
|
58 |
|
59 |
### Evidence Collection
|
60 |
|
|
|
32 |
|
33 |
## Dataset Structure
|
34 |
|
35 |
+
We release two versions of the dataset: DRUID and DRUID+. DRUID is a high-quality (corresponding to top Cohere reranker scores) subset of DRUID+ manually annotated for evidence relevance and stance. The dataset contains the following columns:
|
36 |
|
37 |
- id: A unique identifier for each dataset sample. It also indicates claim source.
|
38 |
- claim_id: A unique identifier for each claim in the dataset. One claim may correspond to multiple samples, corresponding to different evidence pieces retrieved from different webpages.
|
|
|
48 |
- relevant: Whether the evidence is relevant with respect to the given claim. Has been manually annotated for the DRUID samples.
|
49 |
- evidence_stance: The stance of the evidence, i.e. whether it supports the claim or not, or is insufficient. Has been manually annotated for the DRUID samples.
|
50 |
|
51 |
+
Samples based on claims from 'borderlines' follow a sligthly different structure, as the "claims" in this case are based on the [borderlines](https://github.com/manestay/borderlines) dataset. Therefore, they have no corresponding claimant, claim source or fact-check verdict, etc.
|
52 |
|
53 |
## Dataset Creation
|
54 |
|
55 |
### Claim Collection
|
56 |
|
57 |
+
We sample claims verified by fact-checkers using [Google's Factcheck API](https://developers.google.com/fact-check/tools/api/reference/rest). We only sample claims in English. The claims are collected from 7 diverse fact-checking sources, representing science, politics, Northern Ireland, Sri Lanka, the US, India, France, etc. All claims have been assessed by human fact-checkers. We also collect claims based on the [borderlines](https://github.com/manestay/borderlines) dataset, for which we can expect there to be more frequent conflicts between contexts.
|
58 |
|
59 |
### Evidence Collection
|
60 |
|