nala-cub/americas_nli

Dataset Card for AmericasNLI

Dataset Summary

AmericasNLI is an extension of XNLI (Conneau et al., 2018) a natural language inference (NLI) dataset covering 15 high-resource languages to 10 low-resource indigenous languages spoken in the Americas: Ashaninka, Aymara, Bribri, Guarani, Nahuatl, Otomi, Quechua, Raramuri, Shipibo-Konibo, and Wixarika. As with MNLI, the goal is to predict textual entailment (does sentence A imply/contradict/neither sentence B) and is a classification task (given two sentences, predict one of three labels).

Supported Tasks and Leaderboards

[Needs More Information]

Languages

• aym
• bzd
• cni
• gn
• hch
• nah
• oto
• quy
• shp
• tar

Dataset Structure

Data Instances

all_languages

An example of the test split looks as follows:

{'language': 'aym', 'premise': "Ukhamaxa, janiw ukatuqits lup'kayätti, ukhamarus wali phiñasitayätwa, ukatx jupampiw mayamp aruskipañ qallanttha.", 'hypothesis': 'Janiw mayamp jupampix p
arlxapxti.', 'label': 2}

aym

An example of the test split looks as follows:

{'premise': "Ukhamaxa, janiw ukatuqits lup'kayätti, ukhamarus wali phiñasitayätwa, ukatx jupampiw mayamp aruskipañ qallanttha.", 'hypothesis': 'Janiw mayamp jupampix parlxapxti.', 'label
': 2}

bzd

An example of the test split looks as follows:

{'premise': "Bua', kèq ye' kũ e' bikeitsök erë ye' chkénãwã tã ye' ujtémĩne ie' tã páxlĩnẽ.", 'hypothesis': "Kèq ye' ùtẽnẽ ie' tã páxlĩ.", 'label': 2}

cni

An example of the test split looks as follows:

{'premise': 'Kameetsa, tee nokenkeshireajeroji, iro kantaincha tee nomateroji aisati nintajaro noñanatajiri iroakera.', 'hypothesis': 'Tee noñatajeriji.', 'label': 2}

gn

An example of the test split looks as follows:

{'premise': "Néi, ni napensaikurihína upéva rehe, ajepichaiterei ha añepyrûjey añe'ê hendive.", 'hypothesis': "Nañe'êvéi hendive.", 'label': 2}

hch

An example of the test split looks as follows:

{'premise': 'mu hekwa.', 'hypothesis': 'neuka tita xatawe m+k+ mat+a.', 'label': 2}

nah

An example of the test split looks as follows:

{'premise': 'Cualtitoc, na axnimoihliaya ino, nicualaniztoya queh naha nicamohuihqui', 'hypothesis': 'Ayoc nicamohuihtoc', 'label': 2}

oto

An example of the test split looks as follows:

{'premise': 'mi-ga, nin mibⴘy mbô̮nitho ane guenu, guedi mibⴘy nho ⴘnmⴘy xi di mⴘdi o ñana nen nⴘua manaigui', 'hypothesis': 'hin din bi pengui nen nⴘa', 'label': 2}

quy

An example of the test split looks as follows:

.', 'label': 2}

shp

An example of the test split looks as follows:

{'premise': 'Jakon riki, ja shinanamara ea ike, ikaxbi kikin frustradara ea ike jakopira ea jabe yoyo iribake.', 'hypothesis': 'Eara jabe yoyo iribiama iki.', 'label': 2}

tar

An example of the test split looks as follows:

{'premise': 'Ga’lá ju, ke tási newalayé nejé echi kítira, we ne majáli, a’lí ko uchécho ne yua ku ra’íchaki.', 'hypothesis': 'Tási ne uchecho yua ra’ícha échi rejói.', 'label': 2}

Data Fields

all_languages

- language: a multilingual string variable, with languages including ar, bg, de, el, en.
- premise: a multilingual string variable, with languages including ar, bg, de, el, en.
- hypothesis: a multilingual string variable, with possible languages including ar, bg, de, el, en.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

aym

- premise: a string feature.
- hypothesis: a string feature.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

bzd

- premise: a string feature.
- hypothesis: a string feature.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

cni

- premise: a string feature.
- hypothesis: a string feature.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

hch

- premise: a string feature.
- hypothesis: a string feature.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

nah

- premise: a string feature.
- hypothesis: a string feature.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

oto

- premise: a string feature.
- hypothesis: a string feature.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

quy

- premise: a string feature.
- hypothesis: a string feature.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

shp

- premise: a string feature.
- hypothesis: a string feature.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

tar

- premise: a string feature.
- hypothesis: a string feature.
- label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

Data Splits

Language	ISO	Family	Dev	Test
all_languages	--	--	6457	7486
Aymara	aym	Aymaran	743	750
Ashaninka	cni	Arawak	658	750
Bribri	bzd	Chibchan	743	750
Guarani	gn	Tupi-Guarani	743	750
Nahuatl	nah	Uto-Aztecan	376	738
Otomi	oto	Oto-Manguean	222	748
Quechua	quy	Quechuan	743	750
Raramuri	tar	Uto-Aztecan	743	750
Shipibo-Konibo	shp	Panoan	743	750
Wixarika	hch	Uto-Aztecan	743	750

Dataset Creation

Curation Rationale

[Needs More Information]

Source Data

The authors translate from the Spanish subset of XNLI.

AmericasNLI is the translation of a subset of XNLI (Conneau et al., 2018). As translators between Spanish and the target languages are more frequently available than those for English, we translate from the Spanish version.

As per paragraph 3.1 of the original paper.

Initial Data Collection and Normalization

[Needs More Information]

Who are the source language producers?

[Needs More Information]

Annotations

Annotation process

The dataset comprises expert translations from Spanish XNLI.

Additionally, some translators reported that code-switching is often used to describe certain topics, and, while many words without an exact equivalence in the target language are worked in through translation or interpretation, others are kept in Spanish. To minimize the amount of Spanish vocabulary in the translated examples, we choose sentences from genres that we judged to be relatively easy to translate into the target languages: “face-to-face,” “letters,” and “telephone.”

As per paragraph 3.1 of the original paper.

Who are the annotators?

[Needs More Information]

Personal and Sensitive Information

[Needs More Information]

Considerations for Using the Data

Social Impact of Dataset

[Needs More Information]

Discussion of Biases

[Needs More Information]

Other Known Limitations

[Needs More Information]

Additional Information

Dataset Curators

[Needs More Information]

Licensing Information

Creative Commons Attribution Share Alike 4.0 International: https://github.com/abteen/americasnli/blob/main/LICENSE.md

Citation Information

@inproceedings{ebrahimi-etal-2022-americasnli,
    title = "{A}mericas{NLI}: Evaluating Zero-shot Natural Language Understanding of Pretrained Multilingual Models in Truly Low-resource Languages",
    author = "Ebrahimi, Abteen  and
      Mager, Manuel  and
      Oncevay, Arturo  and
      Chaudhary, Vishrav  and
      Chiruzzo, Luis  and
      Fan, Angela  and
      Ortega, John  and
      Ramos, Ricardo  and
      Rios, Annette  and
      Meza Ruiz, Ivan Vladimir  and
      Gim{\'e}nez-Lugo, Gustavo  and
      Mager, Elisabeth  and
      Neubig, Graham  and
      Palmer, Alexis  and
      Coto-Solano, Rolando  and
      Vu, Thang  and
      Kann, Katharina",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.acl-long.435",
    pages = "6279--6299",
    abstract = "Pretrained multilingual models are able to perform cross-lingual transfer in a zero-shot setting, even for languages unseen during pretraining. However, prior work evaluating performance on unseen languages has largely been limited to low-level, syntactic tasks, and it remains unclear if zero-shot learning of high-level, semantic tasks is possible for unseen languages. To explore this question, we present AmericasNLI, an extension of XNLI (Conneau et al., 2018) to 10 Indigenous languages of the Americas. We conduct experiments with XLM-R, testing multiple zero-shot and translation-based approaches. Additionally, we explore model adaptation via continued pretraining and provide an analysis of the dataset by considering hypothesis-only models. We find that XLM-R{'}s zero-shot performance is poor for all 10 languages, with an average performance of 38.48{\%}. Continued pretraining offers improvements, with an average accuracy of 43.85{\%}. Surprisingly, training on poorly translated data by far outperforms all other methods with an accuracy of 49.12{\%}.",
}

Contributions

Thanks to @fdschmidt93 for adding this dataset.