What does Chat2Trade do?
Chat2Trade automatically translates client requests submitted by chat or email into a structured output format that can be used by internal pricing tools to provide a quote. Prior to Chat2Trade, RFQs, that can vary in complexity, were manage manually by salespeople who interpreted the request and filled in a response form with the characteristics of each product on a dedicated platform. Answering these requests manually is time-consuming: it could require several minutes per quote, depending on the complexity of a request. This step is now reduced to a simple copy-and-paste in less than a second.
This project has become a benchmark within the Bank… Chat2Trade has turned the NLP automation embedded in these systems into a standard of delivery for sales and trading workflow tools, to the extent that a new system is now almost inconceivable without that AI component.
Manuel PEREIRA COIMBRA, Data scientist, CIB Global Markets Data & AI lab
Technically, to translate a client's request from plain text extracted from a chat or email to a structured output format, the request must first be understood. To do this, the Global Markets Data & AI lab uses a pre-trained open-source language model, BERT - developed by Google - that is then fine tunned internally on historical requests from BNP Paribas clients, or synthetic data built from a set of templates. In NLP (Natural Language Processing) lingo, the models learn by performing ‘text classification’ (matching a sequence of text with a category) and ‘named entity recognition’ (assigning concepts to entities within the text). Once a model has been trained, the outputs of the classification and tagging tasks are used to create a final structured output format.
