metadata
configs:
- config_name: af
data_files:
- split: test
path: af.tsv
- config_name: ar
data_files:
- split: test
path: ar.tsv
- config_name: az
data_files:
- split: test
path: az.tsv
- config_name: be
data_files:
- split: test
path: be.tsv
- config_name: bg
data_files:
- split: test
path: bg.tsv
- config_name: bn
data_files:
- split: test
path: bn.tsv
- config_name: ca
data_files:
- split: test
path: ca.tsv
- config_name: ceb
data_files:
- split: test
path: ceb.tsv
- config_name: cs
data_files:
- split: test
path: cs.tsv
- config_name: cy
data_files:
- split: test
path: cy.tsv
- config_name: da
data_files:
- split: test
path: da.tsv
- config_name: de
data_files:
- split: test
path: de.tsv
- config_name: el
data_files:
- split: test
path: el.tsv
- config_name: en
data_files:
- split: test
path: en.tsv
- config_name: es
data_files:
- split: test
path: es.tsv
- config_name: et
data_files:
- split: test
path: et.tsv
- config_name: eu
data_files:
- split: test
path: eu.tsv
- config_name: fa
data_files:
- split: test
path: fa.tsv
- config_name: fi
data_files:
- split: test
path: fi.tsv
- config_name: fr
data_files:
- split: test
path: fr.tsv
- config_name: ga
data_files:
- split: test
path: ga.tsv
- config_name: gl
data_files:
- split: test
path: gl.tsv
- config_name: he
data_files:
- split: test
path: he.tsv
- config_name: hi
data_files:
- split: test
path: hi.tsv
- config_name: hr
data_files:
- split: test
path: hr.tsv
- config_name: hu
data_files:
- split: test
path: hu.tsv
- config_name: hy
data_files:
- split: test
path: hy.tsv
- config_name: id
data_files:
- split: test
path: id.tsv
- config_name: it
data_files:
- split: test
path: it.tsv
- config_name: ja
data_files:
- split: test
path: ja.tsv
- config_name: ka
data_files:
- split: test
path: ka.tsv
- config_name: ko
data_files:
- split: test
path: ko.tsv
- config_name: la
data_files:
- split: test
path: la.tsv
- config_name: lt
data_files:
- split: test
path: lt.tsv
- config_name: lv
data_files:
- split: test
path: lv.tsv
- config_name: ms
data_files:
- split: test
path: ms.tsv
- config_name: nl
data_files:
- split: test
path: nl.tsv
- config_name: pl
data_files:
- split: test
path: pl.tsv
- config_name: pt
data_files:
- split: test
path: pt.tsv
- config_name: ro
data_files:
- split: test
path: ro.tsv
- config_name: ru
data_files:
- split: test
path: ru.tsv
- config_name: sk
data_files:
- split: test
path: sk.tsv
- config_name: sl
data_files:
- split: test
path: sl.tsv
- config_name: sr
data_files:
- split: test
path: sr.tsv
- config_name: sv
data_files:
- split: test
path: sv.tsv
- config_name: ta
data_files:
- split: test
path: ta.tsv
- config_name: th
data_files:
- split: test
path: th.tsv
- config_name: tr
data_files:
- split: test
path: tr.tsv
- config_name: uk
data_files:
- split: test
path: uk.tsv
- config_name: ur
data_files:
- split: test
path: ur.tsv
- config_name: vi
data_files:
- split: test
path: vi.tsv
- config_name: zh
data_files:
- split: test
path: zh.tsv
Dataset Description
This is the BMLAMA53 dataset of our EMNLP 2023 paper Cross-Lingual Consistency of Factual Knowledge in Multilingual Language Models. If you are interested in more datapoints, see our BMLAMA17.
If you find the dataset useful, please cite it as follows:
@inproceedings{qi-etal-2023-cross,
title = "Cross-Lingual Consistency of Factual Knowledge in Multilingual Language Models",
author = "Qi, Jirui and
Fern{\'a}ndez, Raquel and
Bisazza, Arianna",
editor = "Bouamor, Houda and
Pino, Juan and
Bali, Kalika",
booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.emnlp-main.658/",
doi = "10.18653/v1/2023.emnlp-main.658",
pages = "10650--10666",
abstract = "Multilingual large-scale Pretrained Language Models (PLMs) have been shown to store considerable amounts of factual knowledge, but large variations are observed across languages. With the ultimate goal of ensuring that users with different language backgrounds obtain consistent feedback from the same model, we study the cross-lingual consistency (CLC) of factual knowledge in various multilingual PLMs. To this end, we propose a Ranking-based Consistency (RankC) metric to evaluate knowledge consistency across languages independently from accuracy. Using this metric, we conduct an in-depth analysis of the determining factors for CLC, both at model level and at language-pair level. Among other results, we find that increasing model size leads to higher factual probing accuracy in most languages, but does not improve cross-lingual consistency. Finally, we conduct a case study on CLC when new factual associations are inserted in the PLMs via model editing. Results on a small sample of facts inserted in English reveal a clear pattern whereby the new piece of knowledge transfers only to languages with which English has a high RankC score. All code and data are released at https://github.com/Betswish/Cross-Lingual-Consistency."
}