A global AI community requires language-diverse publishing (2408.14772v2)

Abstract: In this provocation, we discuss the English dominance of the AI research community, arguing that the requirement for English language publishing upholds and reinforces broader regimes of extraction in AI. While LLMs and machine translation have been celebrated as a way to break down barriers, we regard their use as a symptom of linguistic exclusion of scientists and potential readers. We propose alternative futures for a healthier publishing culture, organized around three themes: administering conferences in the languages of the country in which they are held, instructing peer reviewers not to adjudicate the language appropriateness of papers, and offering opportunities to publish and present in multiple languages. We welcome new translations of this piece. Please contact the authors if you would like to contribute one.

• The paper demonstrates that English dominance in AI publishing hinders non-native scholars by biasing peer review on language quality.
• It uses a six-year analysis of ICLR reviews to reveal how linguistic barriers impose financial and academic burdens on NNES researchers.
• The authors propose multilingual conferences, revised reviewer guidelines, and educational reforms to foster a more inclusive global AI community.

The Necessity for Language-Diverse Publishing in the Global AI Community

The paper by Haley Lepp and Parth Sarin from Stanford University argues for the critical need to address linguistic inclusivity within the global AI research community. The paper posits that the current hegemony of English in academic publishing and conference proceedings significantly undermines diversity and perpetuates various inequities.

Linguistic Exclusion in AI Research

Lepp and Sarin meticulously detail how the dominance of English in AI publication venues contributes to the exclusion of non-native English speakers (NNES). The top 100 ranked computer science journals and conference proceedings are predominantly in English, which poses a substantial barrier to NNES researchers. The authors highlight that even scholars who have invested considerable time and resources to master academic English face rejection based on language proficiency rather than the substance of their research. This phenomenon is substantiated by an analysis of ICLR peer reviews over the past six years, where a significant number of reviews explicitly or implicitly criticize the language quality of submissions from NNES authors.

Implications of Monolingual Research Practices

The ramifications of a monolingual research landscape are manifold. Firstly, linguistic exclusion leads to heavily uneven linguistic output, marginalizing research in so-called "low-resource" languages. Secondly, it imposes limitations on global education and hiring, as English proficiency becomes a de facto prerequisite for academic and professional advancement in AI. Thirdly, it acts as a financial burden on NNES researchers who often must resort to expensive translation and editing services to meet publication standards.

The Role of Translation Technology

While the advent of automatic writing assistance and translation tools like ChatGPT has been celebrated for promoting inclusivity, the authors argue that this technological solution is symptomatic of a broader issue of linguistic exclusion. They assert that these tools should not obscure the need for systemic changes that allow researchers to publish in their native languages. Moreover, preliminary interviews with multilingual ICLR scholars reveal that expressing oneself in a non-native language can result in a different representation of one's personality and thoughts, further impacting the diversity of knowledge in AI research.

Proposed Interventions

Lepp and Sarin propose several interventions to create a more inclusive AI research community:

1. Multilingual Conferences: Conferences should accommodate the languages of their host countries. This could involve hiring translation services and encouraging scholars to present in languages other than English to unveil the linguistic diversity inherent in the computing community.
2. Reviewer Guidelines: Peer reviewers should be explicitly instructed not to evaluate the "appropriateness" of the language used in submissions. Instead, editors should explore publishing multilingual scholarships, allowing researchers to disseminate their work in multiple languages. The cost burden of translation should not fall solely on NNES researchers but should be distributed across the research community.
3. Educational Reforms: Graduate students should be encouraged, if not required, to take language courses to foster a more linguistically tolerant and diverse environment. Embracing multiple languages within educational paradigms would support the goal of a truly global AI community.

Future Implications and Speculation

The implementation of these recommendations could lead to significant shifts in the AI research landscape. Practically, it may democratize the field of AI research by removing linguistic barriers, thus fostering a more inclusive environment where a broader spectrum of ideas and perspectives can thrive. Theoretically, this could accelerate advancements in AI by leveraging diverse cognitive frameworks and problem-solving approaches from across the globe.

Speculatively, as AI continues to evolve, the integration of linguistic diversity could pave the way for more sophisticated multilingual AI models and applications, thereby bridging gaps between different linguistic communities and enhancing global collaborations.

In conclusion, Lepp and Sarin's work compellingly argues that for the AI community to be truly global and inclusive, there must be a concerted effort to address linguistic disparities. This involves systemic changes to current academic practices, encouraging a shift away from English-centric norms towards a more equitable and diverse research environment.