The FFG funded GENTIO project has ended with MODUL Technology contributing advances in AI-based NLP and NER/NEL approaches, development of its Semantic Knowledge Base, and new classification and summarization algorithms. State of the art AI has improved our Natural Language Processing, Named Entity Recognition and Named Entity Linking components. Good NER combined with relation extraction can feed our Semantic Knowledge Base with new entities and relations automatically identified in news articles. To explore new possibilities to tackle this, we started looking at fine tuning Large Language Models (LLMs) with promising early results, the subject of a conference presentation. Transformers models were compared for the task of text summarization.
A significant impact was made in the task of news article classification, following the IPTC News categories. Here, we decided on fine tuning on top of BERTopic, a neural model for topic classification. The initial topic classification is fed through a Transformer model to align topics to the IPTC News categories. We could show higher accuracy scores on annotated news corpora than just using a language model like RoBERTa to classify articles directly, also published this year at a conference. GENTIO has enabled MODUL Technology to advance in its deployment of state of the art AI (e.g. Transformers models, language models) for core text understanding tasks than underlie ‘Web intelligence’, i.e. extracting trends and insights from Web data.

The FFG funded project EPOCH, coordinated by MODUL Technology, demonstrated the groundbreaking use of machine learning/AI approaches to time series forecasting combined with Web intelligence – the analysis of topics and trends in online news and social media over time. Supported by webLyzard technology’s Web intelligence platform and wood price data from KPMG, MODUL Technology’s AI experts combined statistical forecasting models with features extracted from the online news and social media based on mentions of terms relevant to the wood market and combined sentiment of documents mentioning the terms (e.g. ‘Borkenkäfer’, the bark beetle which damages trees).  Four categories of terms were created, and were tested for their predictive power (the extent to which they improve the accuracy of price prediction compared to pure statistical forecasting). We compared a number of models (Random Forest, XGBoost, LSTM, GRU) with four different datasets (hardwood, softwood, nadelholz (coniferous), rohpapier (raw paper)). The rolling mean sentiment feature was found to be the most predictive for future wood prices, especially for the terms related to weather and pests. We provide a slideset explaining the wood price prediction experiments and a Web interface which allows you to visualise the accuracy and feature importance in prediction using different models and datasets. The experiment code is also available to test the models with your own price data. 

• Slides
• Web interface
• Experiment code

This groundbreaking work was extended to look at predictions of prices in other markets and make use of recent Transformers models such as FinBERT to generate the sentiment features from a news corpus. The developed models outperformed classical models, showcasing the fact that combining basic economic logic (e.g., supply and demand) with affective and background knowledge can lead to significant improvements. MODUL Technology will continue to bring this advance in price prediction to other domains through future projects. 

Journal Articles

• Kaplan, H., Weichselbraun, A., Braşoveanu, A.M.P. (2023) Integrating Economic Theory, Domain Knowledge and Social Knowledge into Hybrid Sentiment Models for Predicting the Crude Oil Markets. Cognitive Computation. Springer Nature. DOI: 10.1007/s12559-023-10129-4.
• Weichselbraun, A., Steixner, J., Brasoveanu, A.M.P., Scharl, A., Gobel, M & Nixon, L.B. (2022). Automatic Expansion of Domain-Specific Affective Models for Web Intelligence Applications. Cognitive Computation 14(1), pp. 228-245. DOI: 10.1007/s12559-021-09839-4.

Book Chapters

• Braşoveanu, A M.P., & Andonie, R. (2022). Visualizing and Explaining Language Models. In Integrating Artificial Intelligence and Visualization for Visual Knowledge Discovery, pp. 213-237. Cham: Springer International Publishing. DOI: 10.1007/978-3-030-93119-3_8.

Modul Technology is excited to start working on the new project AI-CENTIVE, funded by the FFG via the ICT of the Future program. The goal of the 3 year project is to build and manage a complex mobility data ecosystem which will support changing citizens’ mobility behaviour towards choosing more sustainable options. It is an interdisciplinary project that combines data science, deep learning, weather prediction, information visualisation and tourism management. The project is implemented in collaboration with webLyzard technology, Zentralanstalt für Meteorologie und Geodynamik (ZAMG), ummadum, Data Intelligence Initiative (DIO) and BOKU. 

Nowadays the data on mobility choices and the context in which they are taken is split. The aim of the project is to turn it into actionable information thanks to efficient knowledge extraction. A common mobility data ecosystem will merge this data, allowing new AI models that learn to anticipate mobility choices based on varying contexts such as the availability of more environmentally friendly options. Customised incentives leveraging the AI predictions will motivate citizens to make more sustainable mobility choices. As such, AI-CENTIVE will contribute to Austria’s achievement of the UN 2030 Agenda for Sustainable Development. 

Stay tuned … we will very soon go online with the project’s website!