The Challenge

Artificial neural networks have many applications in Natural Language Processing (NLP). Applications include text classification, language creation, answering questions, image captioning, language translation, named entity recognition, speech recognition, and many more. The project goal was to use text mining and machine learning to extract economic sentiment indicators from millions of disparate documents.

The Solution

Since the release of BERT (Bidirectional Encoder Representations from Transformers) by Google in November 2018 there has been a flurry of activity around pre-trained transformer models. Since then, new models from Google (XLNet), Facebook (RoBERTa), Microsoft (MtDNN), and OpenAI(GPT2) have been released each achieving state of the art performance for some set of NLP tasks. Elder Research built a weakly supervised text sentiment classifier using the latest NLP tools such as transformer architecture and transfer learning. The team created aggregated metrics , plots, etc. for extracting net sentiment/attention on different economic areas from a vast array of sources. The solution incorporated a search feature to allow users to  quickly find the sentences that reference a certain topic/sentiment across the set of documents. Using these models as a starting point, we can quickly develop high quality models that can be used to process large amounts of text, making possible new signals to be used for decision making. 

Results

By combining semi-supervised methods with the latest in pre-trained models, we were able to quickly construct high quality models to make valuable new sources of data available for the client to inform decision-making such as rapid portfolio rebalancing based on continuous market signaling at speeds previously unattainable by human research analysts.