Language Model Tokenizers Introduce Unfairness Between Languages
Petrov, Aleksandar, La Malfa, Emanuele, Torr, Philip H. S., Bibi, Adel
–arXiv.org Artificial Intelligence
Recent language models have shown impressive multilingual performance, even when not explicitly trained for it. Despite this, there are concerns about the quality of their outputs across different languages. In this paper, we show how disparity in the treatment of different languages arises at the tokenization stage, well before a model is even invoked. The same text translated into different languages can have drastically different tokenization lengths, with differences up to 15 times in some cases. These disparities persist even for tokenizers that are intentionally trained for multilingual support. Character-level and byte-level models also exhibit over 4 times the difference in the encoding length for some language pairs. This induces unfair treatment for some language communities in regard to the cost of accessing commercial language services, the processing time and latency, as well as the amount of content that can be provided as context to the models. Therefore, we make the case that we should train future language models using multilingually fair subword tokenizers.
arXiv.org Artificial Intelligence
Oct-20-2023
- Country:
- North America
- Haiti (0.14)
- United States (0.04)
- Europe
- United Kingdom > England
- Oxfordshire > Oxford (0.04)
- Switzerland
- Zürich > Zürich (0.04)
- Basel-City > Basel (0.04)
- United Kingdom > England
- Asia
- Sri Lanka (0.04)
- Myanmar (0.04)
- Indonesia > Bali (0.04)
- India (0.04)
- China > Hong Kong (0.04)
- Philippines > Luzon
- Ilocos Region > Province of Pangasinan (0.04)
- Middle East
- Yemen > Amanat Al Asimah
- Sanaa (0.04)
- Syria
- Damascus Governorate > Damascus (0.04)
- Aleppo Governorate > Aleppo (0.04)
- Saudi Arabia
- Mecca Province > Jeddah (0.04)
- Riyadh Province > Riyadh (0.04)
- Qatar > Ad-Dawhah
- Doha (0.04)
- Oman > Muscat Governorate
- Muscat (0.04)
- Lebanon > Beirut Governorate
- Beirut (0.04)
- Israel > Jerusalem District
- Jerusalem (0.04)
- Iraq
- Nineveh Governorate > Mosul (0.04)
- Basra Governorate > Basra (0.04)
- Baghdad Governorate > Baghdad (0.04)
- Yemen > Amanat Al Asimah
- Japan
- Hokkaidō (0.04)
- Kyūshū & Okinawa > Kyūshū
- Fukuoka Prefecture > Fukuoka (0.04)
- Honshū
- Tōhoku (0.04)
- Kantō
- Tokyo Metropolis Prefecture > Tokyo (0.04)
- Saitama Prefecture > Saitama (0.04)
- Chūgoku
- Okayama Prefecture > Okayama (0.04)
- Hiroshima Prefecture > Hiroshima (0.04)
- Chūbu > Ishikawa Prefecture
- Kanazawa (0.04)
- Africa
- Niger (0.04)
- Mauritius (0.04)
- Madagascar (0.04)
- East Africa (0.04)
- Sudan
- Khartoum State > Khartoum (0.04)
- Khartoum (0.04)
- Middle East
- Tunisia > Tunis Governorate
- Tunis (0.04)
- Libya > Benghazi District
- Benghazi (0.04)
- Egypt
- Cairo Governorate > Cairo (0.04)
- Aswan Governorate > Aswan (0.04)
- Tunisia > Tunis Governorate
- North America
- Genre:
- Research Report (0.49)
- Overview (0.46)
- Technology: