The Importance of Legal Professionals from Minority Backgrounds: A Discussion on their Significance
In the rapidly evolving world of artificial intelligence (AI), a growing number of researchers are focusing on breaking new ground in the realm of minoritised languages. These communities, often operating with limited resources, face significant challenges in a highly competitive international market. However, innovative approaches are being developed to build AI systems that effectively support low-resource languages and minimise language modeling biases.
One such approach is the use of Small Language Models (SLMs), compact AI models that are faster, cheaper, and feasible to run on devices with limited computational resources. This efficiency enables deployment in resource-constrained environments, broadening access for speakers of low-resource languages without the need for expensive infrastructure.
Another promising initiative is the UnifiedCrawl Dataset, a framework aimed at improving language model performance on low-resource languages by collecting massive multilingual data via extensive web crawling. This dataset, designed to be effective even when linguistic data is scarce, has shown improved results for multiple low-resource languages compared to prior techniques.
Recent research has also shown that selectively translating small quantities of data from a low-resource target language and blending it with original high-resource language data can substantially enhance model alignment and performance. This approach improves multilingual fluency, instruction-following, and reasoning capabilities of large models while maintaining preference for outputs judged by other language models.
Voice AI and speech technologies are another area of focus, with Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Neural Machine Translation (NMT) technologies increasingly supporting low-resource languages and dialects. This convergence helps make spoken content globally accessible and inclusive, addressing the oral language barrier common in low-resource contexts.
AI is also being used to assist human annotation and prompt generation, helping to mitigate bias and improve data quality for low-resource languages. By using language models as evaluators or judges, the annotation and prompt creation processes can be scaled across multiple languages, maintaining quality and consistency while reducing human workload.
These initiatives collectively address both the scarcity of data for low-resource languages and the biases that large language models may inadvertently encode, moving toward more equitable and inclusive AI systems. However, concerns remain about the potential for a unified global AI system to reproduce existing discriminations and serve only the more widely spoken languages.
In South Africa, the lack of scientific vocabulary in indigenous languages like isiZulu highlights the need to avoid replicating historic racial biases in AI language models. Organisations like Masakhane are working on translating scientific papers into various low-resource languages to expand datasets for further training of AI models.
Lelapa AI, a company based in Africa, is developing AI technologies specifically tailored to African languages and cultural contexts. The term 'low-resource' language is debated, often applied to any language other than English without considering the unique challenges specific language groups face.
Research in minoritised languages can serve as a model for innovation within the broader technological ecosystem. Collaboration and community involvement, such as through initiatives like the Universal Knowledge Core and PanLex, can help tackle language modelling biases and support the development of minoritised language models.
However, it is important to acknowledge that when language models fail to support languages other than English, they erase cultural histories and ways of thinking. Inaccurate translation in AI apps has resulted in asylum seekers being detained due to translation inaccuracies.
As the focus shifts to multilingual language technologies, it is a long way from solving the problems faced by individuals like Tizita, a computer science student in Addis Ababa, who faces challenges working with mainstream AI platforms due to lack of support for her language (Amharic) and cultural customs.
In conclusion, the goal is to build a vocabulary and platform for people to talk about their research in their own language and thereby take cultural ownership over science. By addressing the challenges faced by minoritised languages, we can create a more inclusive and equitable technological landscape that respects and preserves cultural diversity.
The author(s) of this piece do not necessarily reflect the views of the website.
Technology is being used to develop Small Language Models (SLMs), compact AI models that are efficient to run on devices with limited computational resources, thereby broadening access for speakers of low-resource languages without the need for expensive infrastructure. Another promising initiative is the UnifiedCrawl Dataset, which aims to improve language model performance on low-resource languages by collecting massive multilingual data via web crawling, despite the scarcity of linguistic data.
