Towards a Domain-Informed Search Engine for NASA Earth Science Data

Abstract:

NASA Earth Science Data Systems (ESDS) are well-suited for machine learning-based improvements to their operational components, such as domain-informed search engines. Current ESDS, such as the Goddard Earth Sciences Data and Information Services Center (GES DISC), organize datasets categorically by satellite missions, platforms, or instruments, making datasets of interest to a user Findable, Accessible, Interoperable, and Reusable, i.e., FAIR--if the user knows the relevant datasets. For non-dataset-specific queries (e.g., “wildfire”), however, current search engines may not return any result. Also, other information relevant to the search term, such as related datasets, metadata and documentation, and events and phenomena, remain essentially opaque, thus limiting the potential value of archived data to users. Our overall goal is to make archived data not only FAIR but also FAIRER, i.e., FAIR with Enhanced Relationships. To that end, a knowledge base is needed. To automate and scale the development of such a knowledge base, natural language processing (NLP) methods are needed to efficiently analyze the vast amounts of human-readable structured and unstructured data and documents across various scientific disciplines. We have conducted NLP- related work on short documents (e.g., classifying tweets), using algorithms such as Naïve Bayes, Convolutional Neural Network (CNN), Hierarchical Attention Network (HAN), and semi-supervised learning (with tri-training). Our current work focuses on improving the discoverability of GES DISC datasets and related information (i.e., making the data FAIRER). Key early results include (1) extraction of named entities and relations from published documents, using the Span-based Entity and Relation Transformer (SpERT); (2) creation of concept maps of events as domain models with which to extract, via decision trees, related geophysical variables; and (3) extraction of Earth science-related keywords from published documents, using the Bidirectional Encoder Representations from Transformers (BERT). Thus far, we have applied NLP methods to various external documents. We will next apply similar NLP and learning algorithms to GES DISC-internal resources, such as trouble tickets and NASA mission data product technical reports.