使用知識圖譜改進 GenAI.pdf

1、Using Knowledge Graphs to Improve GenAIKMWorld 2024Keynote Tuesday November 19th,2024Dave Clarke,EVP Semantic Graph TechnologyCompany news July 2024-Synaptica acquired by and merges with Squirro Combines enterprise Taxonomy and Ontology with GenAI End-to-end integration of GenAI and Semantic Layer t

2、ooling Four months on already delivering GraphRAG solutions in multiple industriesEnterprise Taxonomy and Ontology SoftwareFor over two decades,Synaptica has been helping organizations around the world to organize,classify,and discover enterprise knowledgeEnterprise Knowledge GraphsSynaptica uses W3

3、C semantic technology,industry-standards,and graph databases to build enterprise knowledge graphsOrganize:develop ontology schema in a no-code UI and manage single-source-of-truth taxonomiesClassify:classify content to taxonomies and perform entity and fact extractionDiscover:guide search,navigation

4、,insights,recommendations,GraphRAG conversationsUn-silo:uniform metadata and unified search across any CMSRecall:mitigate the risk of missing informationPrecision:reduce time wasted on irrelevant resultsInferencing:generate new knowledge from the graphProcess automation:improve efficiency and consis

5、tency Business Value of TOMS and KGs Graphite taxonomy and ontology managementSSOT taxonomies as a semantic RDF graphKnown entity extraction(inline tagging)Document-level aboutness classificationContent-to-concept annotationsContent-aware knowledge graphTaxonomy powered semantic searchTaxonomy guide

6、s RAG&process automationML novel entity and information extraction Identify new concepts for inclusion in taxonomiesRDF knowledge graph supports machine-inferencingBusiness insights and analyticsTaxonomy&OntologyKnowledgeGraph(KG)GenAISemanticsGraphRAGClassification&ExtractionEnd-to-End Integrated S

7、olutionshttps:/commons.wikimedia.org/wiki/File:Albert_Einstein_sticks_his_tongue_1951.jpg“The definition of insanity isdoing the same thing over and over again and expecting a different result.”Albert Einstein(probably never said this)https:/commons.wikimedia.org/wiki/File:Albert_Einstein_sticks_his

8、_tongue_1951.jpg“The definition of insanity isdoing the same thing over and over again and expecting a different result.”Albert Einstein(probably never said this)Wouldnt we be insane if we built machines that give a different answer over and over again in response to the same question?Because“probab

9、ly the right answer”isnt good enough.Why does GenAI need graphs?Large Language Models(LLMs)and Retrieval-Augmented Generation(RAG)are GenAI tools commanding huge interest.They use probabilistic methods to understand and generate text.Knowledge graphs are machine-intelligible representations of conte

10、xtualized connected knowledge.They enable a deterministic counter-balance for GenAI process.Why Knowledge Graphsclassification graphs precision&recallcontent graphsinsights&recommendations agentic graphsbusiness process automationuser graphsaccess control&personalization Using Knowledge Graphs to Im

11、prove GenAImany types of KGmany benefits of KGUsing Knowledge Graphs to Improve GenAItaxonomy&ontology enrichment:concepts,categories,synonyms,relationships,definitions,ground-truth exemplarscontent graph enrichment:content structure,content inter-connectednessGenAI query analysiscontent summarizati

12、on entity&fact extractionWhere is the Semantic Layer in all this?https:/enterprise- KnowledgeIts at the heart of what Squirro doesWorked Example of an Agentic OntologyAn agentic ontology is a graph structured with process-oriented semantic concepts,relationships and properties.It provides the framew

13、ork for how GenAI agents perceive,reason,make decisions,and act.It provides GenAI agents with the enterprise domain knowledge to manage complex interactions with humans and other systems.The transparency of the graph enables human agency over the outcomes of GenAI agents.Relationship declares depend

14、enciesConcept property embeds function call lookupRelationship triggers action by function callBusiness Process Automation1.User enters tech support request:2.RAG classifies the request to Troubleshooting:3.Graph processRequiresInfo dependencies instruct the LLM to generate a follow up question for

15、the user:5.User answers question in natural language:7.A Jira ticket is created using function calling8.The LLM then generates natural language to inform the user of the follow up action and ticket reference.6.The graph has a processHasTeam action predicate directs gathered information to the Suppor

16、t team:4.Graph processCallsDataLookup dependency instructs the RAG to query server log dataDiscoverSearchChatSummarizeCompareAnalyzePredictAutomate“Squirro is the only noteworthy company for Conversational Generative AI in Europe.”Privacy LayerSquirros Enterprise-Grade GenAI PlatformSome Current Sem

17、antic GenAI Projects1.A classification graph(standard taxonomy)for public-facing conversational search in the B2C sector ROI matching consumers to business services with over 95%accuracy2.An agentic ontology graph for managing tech support ticketing in the telecoms hardware sector ROI automation of

18、business process with consistent quality of outcome3.A classification and process graph for regulatory compliance in the financial services sector ROI risk mitigationTakeaway 1 Knowledge graphs improve GenAI-accuracy,insights,automation,personalization Reciprocally,GenAI can enrich and build-out kno

19、wledge graphsTakeaway 2 If you have already invested in enterprise taxonomies,you have a gold-mine of structured knowledge to bootstrap your GenAI initiativesTakeaway 3 Squirro has built out a robust enterprise GenAI platform incorporating the semantic layer,data privacy,guardrails,and agentic toolsTakeawaysThank youBooths 301 to 305 in the Expo hallhttps:/