KIKO: A Tools Framework for Industrial Domain-specific Natural Language Processing Tasks

Domain-specific Natural Language Processing (NLP) tasks need prepared, ground-truth data. However, practical constraints make this task challenging. Domain engineers, who are the subject matter experts, are required to create high-quality training data. However, their availability is limited because they are busy with their day-to-day high-priority project and organizational tasks. Moreover, the data needed to be annotated to train NLP models may not be available at the outset itself. As the prepared data grows, NLP models should take advantage of the new data and train improved machine learning models. Hence, in this paper, we present KIKO, a domain-specific continuous learning NLP framework that allows domain engineers to perform annotations on one document at a time, and retraining when a large enough collection of documents becomes available. KIKO is made of several tools that pipeline together to process documents from the industrial domain. The pipeline enables parsing, pattern-matching, annotation, training, relationship mining, and continuous learning based retraining of the model by using a weaker form of active learning.