Investigation of the economic impact of the European Union public procurement notice descriptions by using natural language processing

Global public procurement amounts to $11 trillion annually. More than 250,000 European Union (EU) public authorities spend around C2.3 trillion every year which constitutes 14 percent of the European Union GDP. Even small improvements in the procurement processes have a large effect on the effectiveness of public funds usage. Therefore, competitive and cost-effective public procurement processes are essential for the effective use of public resources. Public authorities in EU are responsible for applying EU commission directives such as ensuring market competitiveness, transparency and reliable data. To control public procurement activities, predicting future award metrics before bidding is crucial while there exist thousands of documents about procurement calls. The data used in this work consists of European Union's multilingual public procurement notices. The data covers raw contract description texts from 2011 to 2018 spanning 22 EU languages and extracted from publicly available contract documents in Tenders Electronic Daily (TED) which is the online version of 'Supplement to the Official Journal' of the EU. In the experiments, fine-tuned multilingual transformer models, sentence embeddings approach, and a feature based approach are proposed to predict four economic metrics including the single bid awards, the number of offers, foreign contract awards, and contract price effectiveness. In all prediction tasks, proposed methods outperform all baselines including models that use word ngrams and economic indicators. This suggests that notice descriptions play an important role in the outcome of public procurement calls. We also showed that multilingual training improves prediction accuracy in all tasks suggesting that EU dataset can be also used for public procurement calls of other countries. Developed models can be used by public procurement officials to automatize the examining process of procurement notices and detect the ones causing low competition. Participating firms can also use the models to predict potential competition they will face to make better decisions and reduce the future risks. Note that there is still gap to improve model performances by adding economic indicators into models. But the best performing models in this work focus only procurement notice descriptions.