A UIMA Database Interface for Managing NLP-related Text Annotations

A UIMA Database Interface for Managing NLP-related Text Annotations Giuseppe Abrami and Alexander Mehler Text-technology Lab, Goethe University Frankfurt Robert-Mayer-Straße 10, 60325 Frankfurt am Main {abrami, mehler}@em.uni-frankfurt.de Abstract NLP and automatic text analysis necessarily involve the annotation of natural language texts. The Apache Unstructured Information Management applications (UIMA) framework is used in several projects, tools and resources, and has become a de facto standard in this area. Despite the multiple use of UIMA as a document-based schema, it does not provide native database support. In order to facilitate distributed storage and enable UIMA-based projects to perform targeted queries, we have developed the UIMA Database Interface (UIMA DI). UIMA DI sets up an environment for a generic use of UIMA documents in database systems. In addition, the integration of UIMA DI into rights and resource management tools enables user and group-specific access to UIMA documents and provides data protection. Finally, UIMA documents can be made accessible for third party programs. UIMA DI, which we evaluate in relation to file system-based storage, is available under the GPLv3 license via GitHub. Keywords: UIMA, database interface, text annotation, Neo4J, MongoDB, Web Services, REST 1. Introduction NLP and automatic text analysis necessarily involve the annotation of natural language texts. In various projects (e.g. (Da Silva et al., 2006; Kano et al., 2009; Hemati et al., 2016)), especially in an NLP context (e.g. (Savova et al., 2010; Eckart de Castilho and Gurevych, 2014; Patter- son et al., 2017; Kreimeyer et al., 2017; Rudzewitz et al., 2017; Kassner et al., 2017)), the Apache Unstructured In- formation Management applications (UIMA) (Ferrucci et al., 2009) is used as a standard architecture for text anno- tation. Accordingly, there are various tools for processing and producing UIMA compliant documents (e.g. (Ogren et al., 2008; Zeldes et al., 2009; Ferrucci et al., 2010; Hemati et al., 2016; Niekler et al., 2017)). UIMA documents are specified by means of UIMA Type System Descriptor (UIMA TSD) files in order to enable interchangeability. The usage of UIMA in numerous projects and tools shows that this framework can be seen as a de facto standard in the context of NLP (Wilcock, 2017). However, despite its widespread usage, it is surprising that native database sup- port is rarely found for UIMA documents. An exception is (Fette et al., 2013) who introduced an approach to stor- ing UIMA documents by means of relational databases. In this article we want to close the gap between UIMA and databases by looking at several database systems beyond relational ones. In order to make UIMA-based projects (as, for example, TREEANNOTATOR (Helfrich et al., 2018)) conform to the requirements of modern text technology, database support is essential. This becomes clear when comparing database usage with file system-based storage and retrieval of UIMA documents. Generally speaking, approaches using file system-based storage lead to several bottlenecks: (a) No redundancy: The document-based schema of the UIMA framework, which is usually stored via XMI1 files, means that the results of UIMA processes can- not be stored decentrally …

Content truncated to 500 words