PKMC Module II cheat sheet

Two knowledge types: explicit (codified, transferable in documents) and tacit (personal, context-bound, transferred through interaction); Polanyi's dictum: we know more than we can tell.

SECI model (Nonaka & Takeuchi): Socialization (tacit→tacit), Externalization (tacit→explicit), Combination (explicit→explicit), Internalization (explicit→tacit); each conversion requires a 'ba' (shared context).

Codification strategy (Hansen, Nohria & Tierney): knowledge reused via repositories, people-to-documents; fits standardized, repeatable problems. Personalization strategy: knowledge shared person-to-person; fits novel, judgment-heavy problems. Firms should dominate one and support the other.

Resource-based view test for a knowledge asset: valuable, rare, hard to imitate, non-substitutable; a knowledge resource yields sustained advantage only if it is causally ambiguous or socially complex.

Knowledge ladder (Boisot): data → information → knowledge → wisdom; each step adds interpretation, context, and judgment.

Strategic KM maturity staging typically moves from ad hoc capture, to codified repositories, to networked collaboration, to embedded/continuous learning; measure movement, not the label.

Absorptive capacity (Cohen & Levinthal): prior related knowledge determines how much new external knowledge a firm can recognize, assimilate, and exploit; R&D and training build it.

Alavi & Leidner frame KM as process plus outcome: creation, storage/retrieval, transfer, and application; technology supports the process but does not substitute for the social context.

Capture methods by tacit depth: interviews and critical-incident review for reasoning; observation and shadowing for skill; after-action review for team lessons; exit interviews for departing expertise.

Critical knowledge risk ranking = impact of loss × likelihood of loss ÷ ease of replacement; prioritize roles with unique, undocumented, decision-critical knowledge.

Codification quality test: a document is codified only if a competent outsider can act on it without contacting the author; otherwise it is a pointer, not capture.

Lessons-learned failure modes: captured after project closure, no owner, no retrieval path, no validation; fix with in-flight capture, a named owner, and a linked taxonomy term.

Knowledge retention tactics: mentoring/apprenticeship, job rotation, communities, structured handover checklists, and retiring-expert interviews scheduled months before departure, not on the last day.

Cognitive task analysis (e.g., critical decision method) elicits cues, strategies, and anomalies experts use; it captures why, not just what, and is more costly than document review.

Decomposition rule: codify stable, frequently reused knowledge; leave volatile or highly contextual knowledge tacit and route it to expert networks instead.

Retention metrics: percentage of critical roles with a documented successor, average handover lead time, and reuse counts on captured artifacts.

Communities of practice (Wenger): domain (shared topic), community (relationships and trust), practice (shared repertoires, tools, stories); all three must be present.

CoP lifecycle: potential → coalescing → maturing → stewardship → transformation; interventions differ by stage (sponsorship early, leadership succession late).

Legitimate peripheral participation: newcomers learn by doing peripheral tasks and move toward the core; design onboarding as gradual participation, not document reading.

Knowledge-sharing barriers: hoarding incentives, time pressure, lack of trust, status asymmetry, and weak tie strength; structural holes (Burt) are bridged by brokers who span groups.

Social network analysis metrics: degree (direct ties), betweenness (brokerage), density (cohesion), and structural holes; use them to find isolated experts and single points of failure.

Psychological safety predicts team learning behavior more strongly than individual expertise; measure it separately from knowledge-sharing frequency.

Collaboration technology fit: synchronous tools for ambiguous, fast-moving problems; asynchronous repositories for stable reference content; mismatch causes either overload or staleness.

CoP health indicators: active member ratio, response latency to questions, number of new practices adopted outside the community, and leadership succession readiness.

Inverted index supports keyword retrieval; ranking signals include term frequency, inverse document frequency, field boosts, recency, and click feedback.

BM25 scoring: sum over query terms of IDF × (tf × (k1+1)) / (tf + k1 × (1 − b + b × doclen/avglen)); k1 controls term-frequency saturation, b controls length normalization.

Precision = relevant retrieved ÷ retrieved; recall = relevant retrieved ÷ all relevant; F1 = harmonic mean; raising recall usually lowers precision.

Knowledge graph triple: subject–predicate–object; ontologies define classes and properties; OWL adds reasoning (subclass, inverse, cardinality); SPARQL queries RDF graphs.

Semantic search improves recall for synonyms and paraphrase via embeddings or ontology expansion, but can reduce precision without a relevance threshold or reranker.

Faceted search exposes metadata dimensions (author, date, type, topic) so users narrow results; facets must be derived from a controlled vocabulary to stay consistent.

Entity resolution/deduplication: blocking to reduce candidate pairs, then similarity scoring (string, phonetic, embedding) and a merge rule with human review for high-impact entities.

Discovery evaluation: use judged relevance sets, nDCG or precision@k, and query logs; a search change that raises clicks without judged relevance may only be reordering noise.

Controlled vocabulary spectrum by increasing control: term list → synonym ring → taxonomy (hierarchy) → thesaurus (broader/narrower/related) → ontology (classes, properties, axioms).

Taxonomy design rules: one preferred term per concept, synonyms as non-preferred entry terms, mutually exclusive sibling nodes, and a single primary parent to avoid ambiguity.

Facet types (Ranganathan's five, adapted): who, what, when, where, how; faceted classification scales better than a single deep hierarchy for heterogeneous content.

Metadata types: descriptive (title, creator, subject), structural (format, relationships), administrative (rights, retention, provenance); Dublin Core covers 15 core descriptive elements.

Interoperability standards: SKOS for thesauri and taxonomies on the Semantic Web; schema.org for web markup; ISO 25964 for thesaurus construction; DC for cross-domain description.

Governance of vocabulary: a term owner, change-request workflow, versioning, deprecation with redirects, and mapping tables when terms are merged or split.

Quality checks: orphan terms (no content), polyhierarchy without policy, near-duplicate siblings, and vocabulary drift where user queries diverge from preferred terms.

Tagging trade-off: free tagging (folksonomy) captures emergent language and cheaply scales but yields inconsistency; hybrid models seed tags with a controlled vocabulary and allow user extension.

Return on investment = (benefit − cost) ÷ cost; for KM, benefits are often avoided rework hours, faster onboarding, and reduced search time, which must be converted to currency explicitly.

Kirkpatrick levels for KM programs: reaction, learning, behavior change, results; most KM reporting stops at level 1–2 and overstates value.

Leading indicators (contributions, reuse, queries answered, CoP activity) predict outcomes; lagging indicators (cycle time, error rate, time-to-competence) prove them.

Search-time savings estimate = users × searches per period × minutes saved per search ÷ 60 × loaded hourly cost; state assumptions or the number is not auditable.

Balanced scorecard logic applies to KM: financial, customer, internal process, and learning/growth perspectives; no single metric can represent KM value.

Value realization requires a baseline captured before the intervention; without a pre-measure, attribution to KM is not defensible.

Analytics hygiene: distinguish usage telemetry (page views, downloads) from knowledge use (citation in decisions, reuse in deliverables); telemetry alone is a vanity metric.

Benchmarking caution: knowledge-work productivity varies by task novelty; compare like-for-like processes, not cross-industry averages.

Governance roles: executive sponsor (funding, priorities), KM lead (program design), content owners (accuracy, lifecycle), stewards (taxonomy, standards), and users (contribution, flagging).

Records management principle: a record is evidence of a business transaction; retention schedules are set by legal/regulatory obligation plus business need, and disposal must be documented and defensible.

Legal hold suspends normal disposition for identified custodians and systems; destroying held content is spoliation and carries sanctions.

Data classification tiers (e.g., public, internal, confidential, restricted) drive access control, encryption, and sharing rules; classification must be assigned at creation, not retrofitted.

Privacy compliance: collect minimum necessary personal data, define lawful basis, honor access/erasure requests, and apply retention limits; personal data inside knowledge bases is still in scope.

Intellectual property risk: third-party content, contractor-authored material, and scraped data need rights clearance; license terms govern reuse and derivative works.

Access control models: role-based (job function), attribute-based (context such as location and clearance), and need-to-know; ABAC handles fine-grained knowledge sharing better than roles alone.

Governance failure signals: undocumented retention, orphaned content owners, uncontrolled external sharing, and no audit trail for content changes.

Architecture layers: capture/authoring, repository/storage, indexing and search, integration/API, presentation, and analytics; each layer should be replaceable without rewriting the others.

Repository choice: document management systems for controlled, versioned records; wikis for collaborative drafting; graph databases for relationship-heavy knowledge; vector stores for semantic retrieval.

Retrieval-augmented generation (RAG) pipeline: chunk source content, embed into a vector index, retrieve top-k by similarity, rerank, then ground the model's answer in retrieved passages with citations.

RAG quality levers: chunk size and overlap, metadata filters, hybrid keyword+vector retrieval, reranker quality, and freshness of the index; stale indexes produce confident wrong answers.

Integration patterns: APIs and event-driven sync keep repositories consistent; scheduled batch extracts create latency and duplicate records unless a master source is designated.

Single source of truth rule: designate one system of record per entity (e.g., employee, document, product) and reference it elsewhere; multiple masters cause conflicting knowledge.

Security architecture: identity federation (SSO/SAML/OIDC), least-privilege authorization enforced at the API layer, encryption at rest and in transit, and audit logging of access to sensitive content.

Scalability and cost: vector indexes grow with chunk count; embedding refresh, storage, and query latency are recurring costs that must be budgeted alongside the initial build.