GlobalDNA performed the first analysis of economic activity of a country through the prism of financial transaction network and developed novel approach for extraction of aspect-based sentiment extraction from news and social media texts.
Analysis of economic activity was based on anonymized dataset of transaction logs from payment processing system covering a period of several years. Transaction logs, together with Statistical classification of economic activities in the European Community (NACE) enabled derivation of two classes of networks: (a) one node is a company and an edge correspond to some type of activity between the two companies, and (b) one node is one NACE category and an edge correspond to some level of activity between companies from the NACE categories.
The analysis of these networks provided insights into the structure of the economy on the country level such as identification and characterization of the core industries and companies based on the economic activity. We developed algorithms for automatic extraction of supply and value chains in the economy on both company or category networks. Finally, we proved that we can improve the prediction of the probability of default based on the position of the company in the financial network.
GlobalDNA also developed an approach for extraction of sentiment-based aspect extraction from news and social media texts based on deep learning techniques. We focused on open-world (i.e. there is no predefined schema of aspects, which are identified directly from data) aspect-based approach (i.e. identify sentiment towards specific target along different aspects. The approach uses two main inputs: list of targets which we want to analyze (e.g. car brands or models, companies, politicians) and social media texts (e.g. reviews, Twitter, Reddit) or news articles. The approach we developed is based on topic modeling with neural networks using the paradigm of attention. Algorithm for aspect-based sentiment extraction provided a novel neural network architecture which enabled unsupervised identification of aspects and topic modeling social media posts and news articles according to these aspects and sentiment.
