Expanding global news reach using machine translation to increase the volume and speed of Spanish and Portuguese news publications. Challenges included ensuring high-quality, accurate translations in AP’s house style, and handling how news stories evolve incrementally without re-translating whole stories.

​

HOW THE PLURALIT TEAM WORKED WITH AP

• Collaborative Development: We worked closely with AP’s editorial, technical, and product teams to align solutions with their goals.

• Agile Methodology: We adopted weekly sprints to test, refine, and deploy AI capabilities rapidly.

• PoC/ Pilot/ Pre-Production: We successfully delivered all 3 use cases through to pre-production, using AP’s development framework and meeting gate criteria for progression from one stage to the next.

• Feedback-Driven Improvements: We leveraged real-time user insights to enhance functionality and user satisfaction.

• Transparent Integration: We adopted AP’s development and communication practices to act as an integrated partner to their editorial, engineering and data science teams.​​

SOLUTION

• PoC -GenAI model exploration and optimisation: Competitive evaluations of leading LLMs (Claude Sonet 3.5, GPT-4o-mini, GPT-4o, DeepL and AWS Bedrock), were conducted, with prompts improved iteratively over weekly sprints based on qualitative and quantitative editorial feedback, to achieve the AP editorial style.

​​

• Custom Translation Workflows: Delivered GenAI capabilities that generated high-quality draft translations for Spanish and Portuguese news articles, enabling faster editorial review and publication.

​​

• ‘Write-through’ translation: re-translation of news updates for published articles (considering previous translations to maintain coherence and a strong AP style whilst reducing editors’ effort) required a complex prompting approach. 

​​

• Advanced Evaluation Metrics: We developed a measurement and evaluation solution built on two main pillars:

  • User Experience Metrics (UX) captured important qualitative aspects such as user satisfaction, perceived translation accuracy, and the frequency of content that needs to be regenerated. These metrics provided a clear understanding of service quality and usability from the editor’s perspective.

  • Operational Metrics monitored the system's technical performance comprehensively. They tracked critical parameters like response time, token consumption during OpenAI calls, API usage frequency, and usage patterns. This information enabled predictive scalability analysis, resource optimisation, and the identification of specific opportunities for improving system performance.

​​

• Pilot - User interface development: A simple, editor-friendly user interface was developed to allow co-working of the GenAI tools alongside  AP's existing editorial systems. The UI also incorporated analytics and feedback feeding into Microsoft PowerBI to inform ongoing performance optimisation.