UIMA-HPC: High-Performance Knowledge Mining

UIMA-HPC will enable data mining applications to make efficient use of High Performance Computing resources. The focus will be on the bio-pharmacological area for which e.g. the PubMed database holds more than 20 million entries. Researchers in this field need to find answers to questions such as the following: For a given base structure, are there any structure variants already mentioned in literature, and if so, are there any indications of their effects? Are structure variants protected by third-party rights or are they freely available? These questions cannot be answered by sheer keyword searches.

Therefore the project aims to realize an HPC-based solution for the automated analysis of multi-modal pharmaco-chemical document databases, taking the patent-search use-case as an initial solution design driver. The combination of text and structure analysis is an innovative approach, but will be based on an existing and welltested data analysis architecture: the Unstructured Information Management Architecture (UIMA). UIMA is a software architecture which specifies component interfaces, design patterns and development roles for creating, describing, discovering, composing and deploying multi-modal analysis capabilities. The UIMA specification is being developed by a technical committee at OASIS (Organization for the Advancment of Structured Information Standards).

The UIMA-HPC approach centres on the workflows for the automated annotation of a document corpus, the work- flow comprising analysis components within the UIMA architecture. The individual “annotation engines”, such as text-mining of a document or analysis of diagrams within a document based on Optical Character Recognition (OCR), are of a computational complexity such that parallelization at the level of the heterogeneous “node” of a modern HPC system is highly appropriate, meaning parallelization for deployment on multi-core and/or GPU-accelerated processors. Handling the large quantity of documents - and the related load-balancing issues created by the diversity of computational complexity relating to individual documents - to be analyzed by independent instantiations of the annotation engines for the workflow is handled at the level of the nodes of the HPC compute system as a whole and will be realized within an adaptation of the UNICORE software system. An example workflow is shown in Figure 1 where red-coloured framed boxes denote UIMA-analysis pipelines and orange triangles split or collect data to achieve load-balancing and parallel execution of pipelines.

UIMA-HPC is a collaborative project funded in part by the German Federal Ministry for Education and Research (BMBF – Bundesministerium für Bildung und Forschung, Förderkennzeichen 01IH11012) and running for three years; the Consortium is led by FHG-SCAI and includes Forschungszentrum Jülich GmbH, scapos AG, and Taros Chemicals GmbH.

The Jülich Supercomputing Centre puts its R&D effort in the development of algorithms and tools for the distribution and collection of data as well as the calculation of the appropriate number of parallel analysis streams and monitoring.

• Mathilde Romberg
Jülich Supercomputing Centre

  top