Grid Computing Cluster – the Development and ... - Lim Lian Tze

Grid@USMwhole#2entity#1273 nodes 7 nodes 5 nodes 6 nodes148699 7553medical building#1medical institution#1medical man#1hypernymhypernymhypernymhospital#1 hospital#2doctor#18 64hypernym hypernymsurgeon#1 1 allergist#1 2has instance3Hippocrates#1Figure 1: Discrete relations and non-discrete neighbourhoods of lexical items and their meanings. Hypernymy (the ‘is-a’relation) links surgeon to doctor etc, but does not link hospital to surgeon nor doctor directly. Instead, all lexical items related tothe medical profession are within nearby neighbourhood of each other by the conceptual vector model.clusters consisting of 340 CPUs. Closerto home, we had previously deployedour machine translation system on agrid cluster for load balancing.An NLP application would needto understand – or seen to be capableof understanding – ‘meaning’ asconveyed by natural language discourse,if it was to perform anythingdeemed ‘intelligent’. While the exactdefinition of ‘meaning’ is still debatedamong linguists and philosophers, weattempted to create a simple computationalrepresentation of word (lexicalitem) meanings. This model is used byBlexisma2, our distributed multi-agentsystem, to construct a semantic lexicaldatabase and to perform semanticprocessing of natural language texts.REPRESENTING LEXICALMEANING WITH COMPUTERSWe designed a Semantic Lexical Base(SLB) to store computer representationsof word meanings. The meaningof a word or lexical item can bemodelled by its relations to other lexicalitems. For example, the Englishword heavy is used to convey the intensityof rain – we do not say bigrain or even strong rain in English.In addition, rain intuitively calls tomind other words like snow, sunny,wet, water. We model such explicitlinguistic relations with discrete linksbetween lexical items (and betweentheir meanings), and implicit ‘conceptualneighbourhood’ with mathematicalvectors. Essentially, the conceptualvector model projects words and theirmeanings onto a spherical space, inwhich meanings with similar ‘conceptualthemes’ (e.g. water, weather) arelocated close to each other. Figure 1illustrates how this hybrid representationscheme captures multiple facetsof lexical item meanings.The SLB is designed to handle polysemy,i.e. that a surface lexical form(e.g. mouse) can have multiple meanings(a small, furry animal; a computeraccessory; a timid person; etc).In other words, the SLB will store arecord for each acknowledged meaning.One of the key aims of Blexisma2is to automatically acquire as manymeaning records (called acceptions)as possible from multiple sources,including dictionaries, terminologylists and Web articles. We chose touse multiple learning sources becauseno single source can claim exhaustivecoverage of a language’s lexicon.Moreover, if the information or definitiongiven by a particular source isbadly written to the point of introducingnoise, the negative effect can beautomatically reduced when data ofbetter quality from another source isprocessed. Therefore, we also store arecord (called a lexie) for each meaningentry acquired from a differentsource. The overall SLB organisationis shown in Figure 2.Blexisma2 is expected to run continuouslyand iteratively, so that newlycoinedterms or new usages of existingwords will be captured in theSLB as they appear. The quality ofacception and lexie objects will alsobe refined with each iterative cycle,as each refined record would ‘feed’positive inputs into the learning processof other lexical meanings. Weimplemented the SLB as a PostgreSQLdatabase, hosted on utmkblex in theUSM Campus Grid.BLEXISMA2 AGENTSBlexisma2 is a distributed multi-agentsystem, implemented with the Mad-Kit platform (http://www.madkit.net). Each Blexisma2 agent has a specificrole or function to perform. Theymay communicate with each other bysending messages, managed by theMadKit kernel. An agent may requestthe help of other agents to completeits task. Conversely, an agent may24 GRID-ENABLED BLEXISMA2