Search

CBR applications have been deployed in a wide range of sectors, from pharmaceuticals; to defence and aerospace to IoT and transportation, to poetry and music generation; for example. However, a majority of these have been built using monolithic architectures which impose size and complexity constraints. As such these applications have a barrier to adopting new technologies and remain prohibitively expensive in both time and cost because changes in frameworks or languages affect the application directly. To address this challenge, we introduce a distributed and highly scalable generic CBR system, Clood, which is based on a microservices architecture. This splits the application into a set of smaller, interconnected services that scale to meet varying demands. Experimental results show that our Clood implementation retrieves cases at a fairly consistent rate as the casebase grows by several orders of magnitude and was over 3,700 times faster than a comparable monolithic CBR system when retrieving from half a million cases. Microservices are cloud-native architectures and with the rapid increase in cloud-computing adoption, it is timely for the CBR community to have access to such a framework.

Semantic annotation is an enabling technology which links documents to concepts that unambiguously describe their content. Annotation improves access to document contents for both humans and software agents. However, the annotation process is a challenging task as annotators often have to select from thousands of potentially relevant concepts from controlled vocabularies. The best approaches to assist in this task rely on reusing the annotations of an annotated corpus. In the absence of a pre-annotated corpus, alternative approaches suffer due to insufficient descriptive texts for concepts in most vocabularies. In this paper, we propose an unsupervised method for recommending document annotations based on generating node descriptors from an external corpus. We exploit knowledge of the taxonomic structure of a thesaurus to ensure that effective descriptors (concept summaries) are generated for concepts. Our evaluation on recommending annotations show that the content that we generate effectively represents the concepts. Also, our approach outperforms those which rely on information from a thesaurus alone and is comparable with supervised approaches.