Structured Products
Written by Patrick Reum
Jun 18, 2024
The Capmatix TranslAIte (CTAI) project was launched in the first quarter of 2024. Originally it was intended to support the translation of regulatory documents, but other internal use cases for the tool have emerged such as the optimization of emails. The tool enables LPA employees to use Chat GPT 4 Turbo customized to the LPA context. Users can choose from various use cases, for example translating text modules for PRIIPs KID cases, composing professional emails, or summarizing texts and generating the corresponding output.
Over the course of approximately 60 days and across three sprints, a Scrum-organized team consisting of three developers with specialties in areas such as architecture, backend, GUI and DevOps, a product manager and a product owner developed an internal tool for LPA. The Capmatix TranslAIte tool has been set up in a dedicated internal Azure environment. Additionally, various authentication options for users were explored, and ultimately EntraID was implemented. There are two categories of users: the regular LPA end user and the administrator, who performs additional CRUD operations and takes care of user management. During that time, our development and product team has gained valuable insights and experience in dealing with artificial intelligence and its integration into projects.
A significant part of the project was the development of prompts. In particular, the use of AI for translations posed challenges: prompts require a precise and comprehensive context for the translation, but there is also the question of how to evaluate an output. The common prompt techniques were considered as a starting point for our investigations (see details on page 10). To address the question of translation quality in the context of PRIIPs KIDs we investigated multiple measuring methods. Usually, the quality of translations is assessed by comparing the machine-generated text with a human “correct” translation. There are various methods with specific characteristics, such as the ROUGE method (Recall-Oriented Understudy for Gisting Evaluation), which compares matches between machine and human translation using synonyms and overlaps of whole units or phrases. The BLEU score (Bilingual Evaluation Understudy), on the other hand, evaluates exact matches. Since many text phrases and sections in KID texts are strictly defined by the European Banking Authority, the focus was placed on the BLEU score to assess the quality of the prompts (see below). To summarize, it was found that the balancing act between necessary precise translations (e.g., in the section “Purpose”) and text with free wording, context-sensitive translations (e.g., section “What is the product”) cannot be represented using a single prompt. Several prompts per section lead to better results but limit the usability and value of the application. The phenom enon of hallucinations, known from GPT, posed an additional difficulty, particularly for precise translations.
The additional features of the tool include user- friendly functions, such as the ability to copy and paste both the input and the output, as well as the option to rate the output. The entire feedback is regularly evaluated by the Structured Products product team to derive further action strategies and improvements on our prompts.
LPA has acquired considerable experience in the field of artificial intelligence (AI) within a remarkably short period of time and offers all our employees an easy way to access GPT-4. We are now utilizing this knowledge to advance the development and deployment of custom-tailored AI models to further improve the quality of the translation results. We also expect that the new models will help in improving our Support capabilities as they are trained in existing user manuals for the various LPA Software products.
This report explores the world of structured products, discussing their future in the era of automation and disruptive technologies. It gathers insights from the forefront of RegTech implementation and analyses distinctions between structured products markets in Europe and Asia.
Senior Product Manager, Germany
Product Owner, Germany
Written by Simon Blechinger
17 Feb, 2025
Written by Simon Blechinger
17 Feb, 2025
Written by Simon Blechinger
5 Feb, 2025
Written by Sebastian Höft
24 Jan, 2025
Written by Sebastian Höft
24 Jan, 2025
Written by Julie Bradini
24 Jan, 2025