Deep and Broad NLP for Big Data and Knowledge Graph Generation

RESTful Microservice API for NLP, JSON-NLP, and Knowledge Graphs

Talk at Indiana University, School of Informatics, Computing and Engineering (SICE), Luddy Hall, room 1106

03/22/2019 10:00 AM

We present our a new research, development, and integration results related to:

• broad and deep NLP components in a scalable big data targeting architecture and API,
• a new NLP annotation standard and Middleware JSON-NLP, and
• HooSIER (HooSIER Semantic Information ExtractoR) our approach to extract core semantic information and Knowledge Graphs (KG) from unstructured text.

We are using a hybrid ensemble of NLP technologies consisting of knowledge based, probabilistic and neural NLP and Information Extraction methods. The underlying technology represents a new kind of NLP API and system architecture that addresses common limitations of open and free NLP technologies by providing a scalable and robust environment. The parallel RESTful Microservice architecture of the NLP API addresses the lack of standardized and uniform output annotations by providing our new JSON Schema for NLP that together with the API abstraction layer also represents a new form of an NLP Middleware. Numerous common and popular NLP pipelines and environments are integrated in this new infrastructure, optimizing their performance and runtime, simplifying the access and management, and making the outputs accessible for advanced NLP-based engineering.

We also present a novel Information Extraction system and text to KG generator (HooSIER), which is domain independent and easily adaptable towards domain specific texts and Information Extraction. It is based on our unified ensemble of Natural Language Processing (NLP) Microservice pipelines. It utilizes lexical, syntactic, semantic, and pragmatic analyses based on their output, including entity and Part-of- Speech (POS) tagging, lexical analysis, syntactic constituency and dependency parsers, anaphora and coreference analysis, and unique approaches to presupposition and implicature computation. Our approach is capable of predicting implied entities and arguments, from implicit subjects, over discourse linked implied arguments, to certain predicates in gapping constructions, and arguments that were subject to ellipsis. The extraction of core semantic relations based on predicate argument structures at the clause level yields abstract graph representation between entities and concepts in the text that represent core semantic content. Entities and relations are typed and linked against existing large KGs using common word and graph embedding disambiguation approaches. Our platform is also capable of generating probabilistic KGs with typed and linked entities.

The core NLP API will soon be available to the general community on the IUB campus. The project is a result of year long research and development activities that included students and colleagues from the different schools at IU Bloomington, and also external collaborators. We should emphasize that significant contribution came from the volunteers and NLP enthusiasts from the NLP Lab.