Continuous improvement of a document treatment chain using reinforcement learning

We tackle the problem of continuous improvement of a treatment chain which extracts events from open-source documents. We use the human operators' corrections to allow the treatment chain to learn from its errors, and self-improve generally. We apply reinforcement learning (specifically Q-learning) to this problem, where the actions are the services of a treatment chain for the extraction of information. The objective is to use the user feedback to allow the system to learn the ideal configuration of the services (order, gazetteers, and extraction rules) based on the characteristics of the documents treated (language, type, etc.). We carry out the first experiments with automatically generated feedback data, and the results are encouraging.