debatelab/deepa2

# e-SNLI to DeepA2

## Original Dataset

NLI dataset with explanations for why the inferential relation indicated by the `label` holds between `premise` and `hypothesis`.

### Features

```yaml
premise: "This church choir sings to the masses as
    they sing joyous songs from the book at a church."
hypothesis: "The church has cracks in the ceiling."
label: "neutral"
explanation_1: "Not all churches have cracks in the
    ceiling"
explanation_2: "There is no indication that there are
    cracks in the ceiling of the church."
explanation_3: "Not all churches have cracks in the
    ceiling."
```



### Source

* [esnli at github](https://github.com/OanaMariaCamburu/e-SNLI)
* [esnli at hugging face](https://huggingface.co/datasets/esnli)

### License

MIT License

### Citation

```bibtex
@incollection{NIPS2018_8163,
  title = {e-SNLI: Natural Language Inference with Natural Language Explanations},
  author = {Camburu, Oana-Maria and Rockt\"{a}schel, Tim and Lukasiewicz, Thomas and Blunsom, Phil},
  booktitle = {Advances in Neural Information Processing Systems 31},
  editor = {S. Bengio and H. Wallach and H. Larochelle and K. Grauman and N. Cesa-Bianchi and R. Garnett},
  pages = {9539--9549},
  year = {2018},
  publisher = {Curran Associates, Inc.},
  url = {http://papers.nips.cc/paper/8163-e-snli-natural-language-inference-with-natural-language-explanations.pdf}
}
```

## Preprocessing e-SNLI for DeepA2

e-SNLI examples are grouped by `premise` and splitted into chunks of three items with different labels (but identical premise). These three items are then merged into a single preprocessed example, e.g.:

```yaml
premise: "This church choir sings to the masses as 
   they sing joyous songs from the book at a church."
hyp_neutral: "The church has cracks in the ceiling."
hyp_entailm: "The church is filled with song."
hyp_contrad: "A choir singing at a baseball game."
expl_neutral:
-  "Not all churches have cracks in the ceiling"
-  "There is no indication that there are cracks 
   in the ceiling of the church."
-  "Not all churches have cracks in the ceiling."
expl_entailm:
-  "'Filled with song' is a rephrasing of the 'choir 
   sings to the masses.'"
-  "hearing song brings joyous in the church."
-  "If the church choir sings then the church is 
   filled with song."
expl_contrad:
-  "A choir sing some other songs other than book 
   at church during the base play. they cannot see 
   book and play base ball same time."
-  "The choir is at a chruch not a baseball game."
-  "A baseball game isn’t played at a church."
```

## DeepA2-ESNLI

### Construction

From each preprocessed eSNLI example, we build multiple DeepA2 items.

First, we can build simple propositional arguments (including formalizations) of the form

```
(1) premise
(2) if premise, then hyp_entailm
-- with modus ponens from 1,2 --
(3) hyp_entailment
```

or, respectively:

```
(1) premise
(2) if hyp_contrad, then not premise
-- with modus tollens from 1,2 --
(3) not hyp_contrad
```

Second, we construct argument source texts by shuffling premises, hypotheses and explanations. In doing so, some of the sentences serve as distractors, others may be left out to the effect that a premise and/or the conclusion are implicit. 

Thirdly, we construct lists of reasons and conjectures, which identify explicitly stated parts of the argument and map them to their counterparts in the argdown snippet.


### Features

- [x] `source_text`
- [ ] `title`
- [x] `gist`
- [x] `source_paraphrase`
- [ ] `context`

<!-- -->

- [x] `reasons`
- [x] `conjectures`

<!-- -->

- [x] `argdown_reconstruction`
- [x] `erroneous_argdown`
- [x] `premises`
- [ ] `intermediary_conclusion`
- [x] `conclusion`

<!-- -->

- [x] `premises_formalized`
- [ ] `intermediary_conclusion_formalized`
- [x] `conclusion_formalized`
- [ ] `predicate_placeholders`
- [ ] `entity_placeholders`
- [x] `misc_placeholders`
- [x] `plchd_substitutions`