README.md

---
language: it
license: cc-by-sa-4.0
multilinguality: monolingual
task_categories:
- token-classification

tags:
- Factuality Detection
- Modality Detection
---
# ModaFact - Dataset

## Dataset Description

### Dataset Summary

ModaFact is a textual dataset annotated with Event Factuality and Modality in Italian. ModaFact’s goal is to model in a joint way factuality and modality values of event-denoting expressions in text.

### Textual data source

Original texts (sentences) have been sampled from [EventNet-ITA](https://huggingface.co/mrovera/eventnet-ita), a dataset for Frame Parsing, consisting of annotated sentences from Wikipedia. 

### Statistics

| Feature | # |
| :---        |    ----:   |
| Sentences     | 3,039|
| Words   | 73,784 |
| Annotations| 10,445 |
| Unique label assignments |33,029|
| Words per sentence (avg.) |24.28|
| Annotations per sentence (avg.) | 3.44|
|Unique label assignments per sentence |10.87|


### Annotation

ModaFact has been originally annotated at token level, adopting the IOB2 style.
Whereas for Modality the schema is unique, for Factuality we provide two representations: a fine-grained representation (FG), which specifies values over three axes (CERTAINTY, POLARITY, TIME), and a coarse-grained representation (CG), which only provides the final factuality value.


Example of **fine-grained representation (FG)**:
```
Per	O
chiarire	B-POSSIBLE-POS-FUTURE-FINAL
la	O
questione	O
la	O
Santa	O
Sede	O
autorizzò	B-CERTAIN-POS-PRESENT/PAST
il	O
prelievo	B-UNDERSPECIFIED-POS-FUTURE-CONCESSIVE
di	O
campioni	O
del	O
legno	O
che	O
vennero	O
datati	B-CERTAIN-POS-PRESENT/PAST
attraverso	O
l'	O
utilizzo	B-CERTAIN-POS-PRESENT/PAST
del	O
metodo	O
del	O
carbonio-14	O
.	O
```

Example of **coarse-grained representation (CG)**:
```
Per	O
chiarire	B-NON_FACTUAL-FINAL
la	O
questione	O
la	O
Santa	O
Sede	O
autorizzò	B-FACTUAL
il	O
prelievo	B-NON_FACTUAL-CONCESSIVE
di	O
campioni	O
del	O
legno	O
che	O
vennero	O
datati	B-FACTUAL
attraverso	O
l'	O
utilizzo	B-FACTUAL
del	O
metodo	O
del	O
carbonio-14	O
.	O
```

#### Labelset

Factuality:

- Fine-grained
  - CERTAINTY: {`CERTAIN`, `PROBABLE`, `POSSIBLE`, `UNDERSPECIFIED`}
  - POLARITY: {`POSITIVE`, `NEGATIVE`, `UNDERSPECIFIED`}
  - TIME: {`PRESENT/PAST`, `FUTURE`, `UNDERSPECIFIED`}

- Coarse-grained
  - {`FACTUAL`, `NON-FACTUAL`, `COUNTERFACTUAL`, `UNDERSPECIFIED`}


Modality:
- {`WILL`, `FINAL`, `CONCESSIVE`, `POSSIBILITY`, `CAPABILITY`, `DUTY`, `COERCION`, `EXHORTATIVE`, `COMMITMENT`, `DECISION`}


### Data format

According to the experimental set presented in the paper (see below, Citation Information) we provide different data formats:
- **token-level BIO sequence labelling**: the dataset is formatted as a two-column `tsv`. The first column contains the token, the second column contains all corresponding labels (factuality and modality), concatenated with `-`. This format makes the dataset ready-to-train with the MaChAmp [seq_bio](https://github.com/machamp-nlp/machamp/blob/master/docs/seq_bio.md) task type.
- **token-level multi-task sequence labelling**: the dataset is formatted as a three-column `tsv`. The first column contains the token, the second column contains all factuality labels, the third column contains the modality label. This format makes the dataset ready-to-train with the Machamp seq_bio **multitask** setting.
- **generative and sequence-to-sequence**: the dataset is formatted as a `jsonl` file, containing a list of dictionaries. Each dictionary has an *Input* field (the sentence) and an *Output* field, a string composed by *token=labels* pairs, separated by `|`. This format makes the dataset ready-to train with sequence-to-sequence and causal/generative models.

### Data Split

For the sake of reproducibility, we provide, for each configuration, the 5 folds used in the paper.
The data split follows a 60/20/20 ratio and has been created in a stratified way. This means each train/dev/test set contains (approx) the same relative distribution of classes.


## Additional Information

An instance of the mT5 model, fine-tuned on ModaFact, is available at [this repo](https://huggingface.co/dhfbk/modafact-ita).

### Licensing Information

ModaFact is released under the CC-BY-SA-4.0 License. 

### Citation Information

If you use ModaFact, please cite the following paper:

```
@inproceedings{rovera-etal-2025-modafact,
    title = "{M}oda{F}act: Multi-paradigm Evaluation for Joint Event Modality and Factuality Detection",
    author = "Rovera, Marco  and
      Cristoforetti, Serena  and
      Tonelli, Sara",
    editor = "Rambow, Owen  and
      Wanner, Leo  and
      Apidianaki, Marianna  and
      Al-Khalifa, Hend  and
      Eugenio, Barbara Di  and
      Schockaert, Steven",
    booktitle = "Proceedings of the 31st International Conference on Computational Linguistics",
    month = jan,
    year = "2025",
    address = "Abu Dhabi, UAE",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.coling-main.425/",
    pages = "6378--6396",
    }
```