The Anglocentric nature of scholarly communication has many implications, such as limiting both publication and access from other language communities (even for major languages); putting minoritized languages at risk in the academic domain; and excluding many from peer review. This scholarly digital divide is worsened by the limited availability of linguistic data and language technologies for machine translation (MT) in most languages. The OSCAIL project follows UNESCO's Recommendation on Open Science in which “openness” entails embracing multiple forms of diversity–including linguistic diversity–to reach a wider audience. OSCAIL will gather domain-specific data in project languages (with good, medium, and weak support following the Digital Language Equality metric, building and testing cutting-edge MT processes. OSCAIL gathers experts in multiple disciplines (MT, NLP, translation evaluation, scholarly publishing) who will work with linguistic communities. Using the data gathered, project partners will explore new techniques to boost performance of large language model (LLM)-based MT systems for scholarly translation. Outputs will be integrated into the Open Journal System (OJS), the world’s most widely used end-to-end scholarly publishing platform for article submission, peer review, and production. OSCAIL will focus on three key use cases: (1)peer review: using MT to enable reviewers to work in their preferred language, and allowing authors to write and respond in the language of their choice; (2)e-discovery: allowing researchers, journal managers, and library and information science professionals to access and cite multilingual publications via translated metadata and content; (3) plain language summarisation and MT of content into project languages for lay readers.
Evaluation aligned with all 3 use-cases will follow best practice, combining automatic and human evaluation and engaging with end users, following reproducible steps. Ethics is at the heart of the project aims, processes, and evaluations, to produce recommendations for the scientific community. We aim to improve the linguistic openness of science, to reduce the digital divide in the scholarly domain, to promote recognition of scholars working in languages other than English, and to increase the value of research and research driven products and services by facilitating regional or national targeted impacts, among others.
OSCAIL will achieve the aims of the SOL call, with experienced experts at each step in the process of MT of scientific knowledge. Several partners are involved with work as part of the European data strategy, such as ALT-EDIC, and the European Language Data Space, and will use these and other resources for data gathering. Partners also have a history in research for MT of scientific publications. We propose a modular framework, that may be updated and enhanced (e.g., the best available LLMs to be deployed through new architectures for final task of MT, summarization, including peer-review support, etc. The LLMs can be boosted/adapted for MT language-pairs with limited effort, or replaced with more recent versions, new entities/words being added to “dictionaries”) for translation and transposition of figures and tables, aiming to produce highly accurate output appropriate for peer review, summarised output for e-discovery, and simplified output for lay readers, catering to a mix of producers and users. MT and evaluation across high-resource and low-resource language scenarios will employ automatic and human metrics, maximising reproducibility and highlighting correlations between metrics. Evaluation with end users is also necessary to address the cultural barriers. The project will produce protocols, guidelines, and interconnectors for open science publication platforms, cognisant of the ethical issues involved and maximising the appropriate use of current technologies in overcoming language and cultural barriers for knowledge sharing in Europe.