Cross-Lingual Consistency (CLC) of Factual Knowledge in Multilingual Language Models

Easy-to-use framework for evaluating cross-lingual consistency of factual knowledge (Supported LLaMA, BLOOM, mT5, RoBERTa, etc.).

Our paper was selected for the Outstanding Paper Award in the Multilinguality and Linguistic Diversity track of EMNLP 2023 and the Best Data Award of GenBench Workshop.

Authors: Jirui Qi • Raquel Fernández • Arianna Bisazza

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.

If you find the paper helpful and use the content, we kindly suggest you cite through:

@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.",
}

Environment:

Python: 3.11

Packages: pip install -r requirements.txt

Quick Start

For a quick start, you only need to run the following two lines to get the CLC of two languages in a PLM:

cd 1_easyrun
bash easyrun.sh

You may also like to modify the variables in easyrun.sh

• mname: The model currently supports LLaMA, BLOOM, BLOOMZ, mT5, RoBERTa. Use the full model name on Huggingface! Like bigscience/bloom-3b.
• lang1 & lang2: Abbreviation of languages in ISO 639-1 format. See the tables below for details.
• mini: yes for using BMLAMA-17 and no for using BMLAMA-53.
• weight: Weight metric for RankC, select among softmax, norm1, and norm2.

Supported languages of BMLAMA-17

Language	ISO 639-1	Language	ISO 639-1	Language	ISO 639-1
English	en	French	fr	Dutch	nl
Spanish	es	Russian	ru	Japanese	ja
Chinese	zh	Korean	ko	Vietnamese	vi
Greek	el	Hungarian	hu	Hebrew	he
Turkish	tr	Catalan	ca	Arabic	ar
Ukrainian	uk	Persian	fa

Supported languages of BMLAMA-53

Language	ISO 639-1	Language	ISO 639-1	Language	ISO 639-1
Catalan	ca	Azerbaijani	az	English	en
Arabic	ar	Ukrainian	uk	Persian	fa
Turkish	tr	Italian	it	Greek	el
Russian	ru	Croatian	hr	Hindi	hi
Swedish	sv	Albanian	sq	French	fr
Irish	ga	Basque	eu	German	de
Dutch	nl	Estonian	et	Hebrew	he
Spanish	es	Bengali	bn	Malay	ms
Serbian	sr	Armenian	hy	Urdu	ur
Hungarian	hu	Latin	la	Slovenian	sl
Czech	cs	Afrikaans	af	Galician	gl
Finnish	fi	Romanian	ro	Korean	ko
Welsh	cy	Thai	th	Belarusian	be
Indonesian	id	Portuguese	pt	Vietnamese	vi
Georgian	ka	Japanese	ja	Danish	da
Bulgarian	bg	Chinese	zh	Polish	pl
Latvian	lv	Slovak	sk	Lithuanian	lt
amil	ta	Cebuano	ceb