slhenty's picture
add dataset format to cf-stsb section
12cc1c1
|
raw
history blame
7.36 kB
---
license: unknown
viewer: false
---
**==========================================**
**_IN PROGRESS - NOT READY FOR LOADING OR USE_**
**==========================================**
---
# Dataset Card for climate-fever-nli-stsb
## Dataset Description
- **Homepage:**
- **Repository:**
- **Paper:**
- **Leaderboard:**
- **Point of Contact:**
### Dataset Summary
The CLIMATE-FEVER dataset modified to supply NLI-style (**cf-nli**) features or STSb-style (**cf-stsb**) features that SentenceBERT training scripts can use as drop-in replacements for AllNLI and/or STSb datasets.
There are two **cf-nli** datasets: one derived from only SUPPORTS and REFUTES evidence (**cf-nli**), and one that also derived data from NOT_ENOUGH_INFO evidence based on the annotator votes (**cf-nli-nei**).
The feature style is specified as a named configuration when loading the dataset: cf-nli, cf-nli-nei, or cf-stsb. See usage notes below for `load_dataset` examples.
### Usage
Load the **cf-nli** dataset
```python
# if datasets not already in your environment
!pip install datasets
from datasets import load_dataset
# all splits...
dd = load_dataset('climate-fever-nli-stsb', 'cf-nli')
# ... or specific split (only 'train' is available)
ds_train = load_dataset('climate-fever-nli-stsb', 'cf-nli', split='train')
## ds_train can now be injected into SentenceBERT training scripts at the point
## where individual sentence pairs are aggregated into
## {'claim': {'entailment': set(), 'contradiction': set(), 'neutral': set()}} dicts
## for further processing into training samples
```
Load the **cf-nli-nei** dataset
```python
# if datasets not already in your environment
!pip install datasets
from datasets import load_dataset
# all splits...
dd = load_dataset('climate-fever-nli-stsb', 'cf-nli-nei')
# ... or specific split (only 'train' is available)
ds_train = load_dataset('climate-fever-nli-stsb', 'cf-nli-nei', split='train')
## ds_train can now be injected into SentenceBERT training scripts at the point
## where individual sentence pairs are aggregated into
## {'claim': {'entailment': set(), 'contradiction': set(), 'neutral': set()}} dicts
## for further processing into training samples
```
Load the **cf-stsb** dataset
```python
# if datasets not already in your environment
!pip install datasets
from datasets import load_dataset
# all splits...
dd = load_dataset('climate-fever-nli-stsb', 'cf-stsb')
# ... or specific split ('train', 'dev', 'test' available)
ds_dev = load_dataset('climate-fever-nli-stsb', 'cf-stsb', split='dev')
## ds_dev (or test) can now be injected into SentenceBERT training scripts at the point
## where individual sentence pairs are aggregated into
## a list of dev (or test) samples
```
<!--
### Supported Tasks and Leaderboards
[More Information Needed]
### Languages
[More Information Needed]
-->
## Dataset Structure
### Data Instances
[More Information Needed]
### Data Fields
[More Information Needed]
### Data Splits
[More Information Needed]
## Dataset Creation
### Curation Rationale
SentenceBERT models are designed for 'Domain Adaptation' and/or 'Fine-tuning' using labeled data in the downstream task domain. As a bi-encoder, the primary objective function is real-valued similarity scoring. Typical training datasets use NLI-style features as input, and STSb-style features as model evaluation during training, and to measure post-hoc, _intrinsic_ STSb performance. Classification tasks typically use a classifier network that accepts SentenceBERT encodings as input, and is trained on class-labeled datasets.
So, to fine-tune a SentenceBERT model in a climate-change domain, a labeled climate change dataset would be ideal. Much like the authors of the CLIMATE-FEVER dataset, we know of no other _labeled_ datasets specific to climate change. And while CLIMATE-FEVER is suitably labeled for classification tasks, it is not ready for similarity tuning in the style of SentenceBERT.
This modified CLIMATE-FEVER dataset attempts to fill that gap by deriving NLI-style features typically used in pre-training and fine-tuning a SentenceBERT model. SentenceBERT also uses STSb-style features to evaluate model performance both during training and after training to gauge _intrinsic_ model performance on STSb.
### Source Data
#### Initial Data Collection and Normalization
see CLIMATE-FEVER
#### Who are the source language producers?
see CLIMATE-FEVER
<!--
### Annotations
-->
### Annotation process
#### **cf-nli**
For each Claim that has both SUPPORTS evidence and REFUTES evidence, create labeled pairs in the style of NLI dataset:
| split | dataset | sentence1 | sentence2 | label |
|---|---|---|---|---|
| {'train', 'test'} | 'climate-fever' | claim | evidence | evidence_label SUPPORTS -> 'entailment', REFUTES -> 'contradiction' |
> Note that by defintion, only claims classified as DISPUTED include both SUPPORTS and REFUTES evidence, so this dataset is limited to a small subset of CLIMATE-FEVER.
### **cf-nli-nei**
This dataset uses the list of annotator 'votes' to cast a NOT_ENOUGH_INFO (NEI) evidence to a SUPPORTS or REFUTES evidence. By doing so, Claims in the SUPPORTS, REFUTES, and NEI classes can be used to generate additional sentence pairs.
| votes | effective evidence_label |
|---|---|
| SUPPORTS > REFUTES | _SUPPORTS_ |
| SUPPORTS < REFUTES | _REFUTES_ |
In addition to all the claims in **cf-nli**, any Claims that have,
* **_at least one_** SUPPORTS or REFUTES evidence, AND
* NEI evidences that can be cast to effective _SUPPORTS_ or _REFUTES_
are included in the datasset.
### **cf-stsb**
For each Claim <-> Evidence pair, create labeled pairs in the style of STSb dataset:
| split | dataset | score | sentence1 | sentence2 |
|---|---|---|---|---|
| {'train', 'dev', 'test'} | 'climate-fever' | cos_sim score | claim | evidence |
This dataset uses 'evidence_label', vote 'entropy', and the list of annotator 'votes' to derive a similarity score for each claim <-> evidence pairing. Similarity score conversion:
> `mean(entropy)` refers to the average entropy within the defined group of evidence
| evidence_label | votes | similarity score |
|---|---|---|
| SUPPORTS | SUPPORTS > 0, REFUTES == 0, NOT_ENOUGH_INFO (NEI) == 0 | 1 |
| | SUPPORTS > 0, REFUTES == 0 | mean(entropy) |
| | SUPPORTS > 0, REFUTES > 0 | 1 - mean(entropy) |
| NEI | SUPPORTS > REFUTES | (1 - mean(entropy)) / 2|
| | SUPPORTS == REFUTES | 0 |
| | SUPPORTS < REFUTES | -(1 - mean(entropy)) / 2 |
| REFUTES | SUPPORTS > 0, REFUTES > 0 | -(1 - mean(entropy)) |
| | SUPPORTS == 0, REFUTES > 0 | -mean(entropy) |
| | SUPPORTS == 0, REFUTES > 0, NEI == 0 | -1 |
The above derivation roughly maps the strength of evidence annotation (REFUTES..NEI..SUPPORTS) to cosine similarity (-1..0..1).
<!--
#### Who are the annotators?
[More Information Needed]
### Personal and Sensitive Information
[More Information Needed]
## Considerations for Using the Data
### Social Impact of Dataset
[More Information Needed]
### Discussion of Biases
[More Information Needed]
### Other Known Limitations
[More Information Needed]
## Additional Information
### Dataset Curators
[More Information Needed]
### Licensing Information
[More Information Needed]
### Citation Information
[More Information Needed]
### Contributions
[More Information Needed]
-->