Extracting information from technical media

Vehicle manufacturers and suppliers produce to the highest quality and safety standards. At the same time, vehicles are under special observation by the authorities. The entire supply chain is well regulated and vehicles are regularly inspected once they are in use. The relevant quality standard for the automotive industry, IATF 16949, has stipulated since 2020 that every manufacturer involved in the vehicle is responsible for monitoring its products, including parts installed in vehicles, in the field. This is often very difficult for component and parts manufacturers further down the supply chain. PUMACY, a company specializing in data analysis, supports vehicle suppliers worldwide in collecting and structuring data from a wide variety of information sources and evaluating it using innovative data analysis and AI methods.

As part of the EU project EUHubs4Dataof which DAISEC is a member, an open platform for vehicle-related communication was developed in the experiment OVeC (Open Platform for Vehicle-related Communication). The aim is to enrich official recall data with additional information from publicly accessible sources. While the processing of purely technical information sources, such as those offered by the vehicle and traffic authorities, is not a major technical challenge, the evaluation of specialist publications aimed at the wider public is a much more complex task. For this reason, the DAISEC developed a prototype news extractor for the OVeC experiment, with the help of which information relating to vehicle recalls in news articles can be recognized and assigned to the respective vehicle recalls in a structured manner. A parallel technical goal of the project was the evaluation of knowledge graphs for use within the Carwatch platform.

Graphic: Overview of the automated categorization of vehicle recalls.

A knowledge graph is generally understood to be a knowledge base that uses a model or topology structured by graphs to represent and process data. Using nodes, edges and labels, any number of complex structures can be created and linked together. The technically relevant text passages for the project were extracted from automotive-related websites using a web scraper. Wikidata provided the necessary structural information for building the knowledge graph, in particular information about car manufacturers and vehicle models. In the in-depth analysis, additional information regarding the context, e.g. the year of manufacture of the vehicle and the relevant components or parts, was recognized from the texts. Two different artificial intelligence methods were used for testing:

LLM-based text classifier: Large language models (LLMs) have already delivered promising results in information extraction in many application areas. This was also confirmed by the exemplary application in the OVeC project.

Transformer-based text classifier: In a second approach, a transformer-based multi-label text classification model was trained using an existing Carwatch dataset. This approach also led to good and practically usable results.

In the coming months, PUMACY will continue to test the results of the project using the extensive Carwatch database and test the strengths and weaknesses of the two methods with regard to the different types of information. If the results are confirmed with larger and more complex data volumes, the project results will be implemented in the operational Carwatch platform.

The Pumacy Technologies AG group of companies enables its customers to make optimum use of product and process knowledge as well as the ideas and experience of their employees within the company. To this end, the company has developed a unique system and a tried-and-tested portfolio of methods and software during a large number of consulting and software projects in the industry. As a result, customers receive a customized solution to create stable value from their knowledge.