Source-linked AI summary
A systematic Approach to constructing a Chance-and-Risk Matrix for Semiconductor Supply Chains
Ema Salkić, Alexander Fichtl, Philipp Ulrich, Hans Ehm, Marta Bonik, Georg Groh
TL;DR
Semiconductor supply chains lack a scalable way to continuously structure and prioritize risk intelligence from public corporate disclosures. The paper builds an OWL-backed knowledge-graph pipeline with LLM extraction, deduplication, scoring, and expert validation; across five companies, it produces 76,207 items with 92.6% validity and identifies trade restrictions as the dominant systemic risk.
Problem
No scalable automated system continuously feeds structured, scored risk and opportunity intelligence from publicly available corporate documents into the ALLPROS matrix.
Method
The pipeline retrieves corporate documents, extracts risks and opportunities with LLMs, consolidates them in an ontology-backed knowledge graph, and ranks them using algorithmic, LLM, and expert-validation layers.
Results
76,207 extracted items achieved 92.6% validity, while LLM-adjusted rankings correlated with expert judgment at ρ=0.55 for risks and ρ=0.72 for opportunities.
Takeaways & Limitations
The resulting matrices identify trade restrictions as the dominant systemic risk across all value-chain positions, while selected opportunities are complementary across companies.
Takeaways & Limitations
The system uses only English-language documents, covers five of twelve ontology-defined supply-chain clusters, relies on one expert per company, and lacks formal recall measurement.
Abstract
from arXiv · showhide
Semiconductor supply chains face escalating risks from geopolitical tensions, geographic concentration, and rapid technological shifts, yet no scalable system continuously extracts, structures, and prioritizes risk intelligence from public corporate disclosures. We present an end-to-end pipeline that retrieves corporate documents for semiconductor companies and uses large language models (LLMs) to extract the risks and opportunities they describe. It organizes these into a knowledge graph linking each item to its category, sources, and related events, then merges duplicates and ranks them with a three-layer mechanism combining an algorithmic formula, an LLM relevance adjustment, and expert validation. Applied to five companies across the value chain, the pipeline produces 76,207 scored items, of which an independent check finds 92.6% valid. The automated rankings match expert judgment at an average Spearman correlation of 0.55 for risks and 0.72 for opportunities, and the resulting matrices identify trade restrictions as the dominant cross-company risk.
1 INTRODUCTION
Semiconductor supply chains are fragile and exposed to concentrated capacity, geopolitical tensions, and costly disruptions. The paper proposes a scalable pipeline that turns corporate disclosures into structured, scored intelligence for the ALLPROS chance-and-risk matrix.
- Motivation: Global semiconductor supply chains are vulnerable because fabrication is capital-intensive, lengthy, and geographically concentrated.A single advanced fab requires over USD 10 billion, while fabrication cycles span months and critical capacity is concentrated in a few regions.
- Motivation: The 2020–2023 chip shortage halted automotive production worldwide and caused GDP losses estimated in the hundreds of billions of dollars.
- ALLPROS: ALLPROS coordinates semiconductor companies, EU Member State governments, and the European Commission across the European value chain.Its Supply Chain working group maps twelve clusters, from raw materials and equipment to electronics manufacturing services and end-user applications.
- ALLPROS: The chance-and-risk matrix plots estimated impact against resources required, with upper-right items representing the most critical issues for coordinated European action.
- Contribution: The pipeline retrieves corporate documents, extracts risks and opportunities with LLMs, consolidates them in an OWL knowledge graph, and ranks them through algorithmic, LLM, and expert layers.
- Contribution: The modular workflow converts unstructured public documents into machine-queryable, scored intelligence and can support targeted due diligence or policy-impact assessments.
2 BACKGROUND
The paper combines supply-chain risk management, ontology-based knowledge representation, and LLM-based extraction to address shortcomings in scalable semiconductor risk intelligence. Prior approaches automate extraction or structure information, but generally do not combine formal ontologies, automation, and ranking for risks and opportunities.
- Supply-chain risk management: Semiconductor supply-chain risk management is difficult because fabrication is geographically concentrated, capital-intensive, and spans hundreds of steps over several months.
- Knowledge representation: An ontology supplies formal classes, properties, and constraints, while a knowledge graph instantiates that schema with concrete entities and typed relationships.
- LLM-based extraction: LLMs enable zero-shot classification and taxonomy-guided extraction, while Sentence-BERT embeddings support semantic deduplication and LLM-as-a-judge evaluation.
- Supply-chain risk management: Existing assessment techniques such as AHP, FMEA, and Bayesian Networks remain dependent on manual expert input and periodic reassessment.
- Related approaches: An LLM-based supply-chain framework automates extraction but lacks structured knowledge representation and scoring or ranking mechanisms.
- Related approaches: Existing knowledge-graph systems may lack formal ontologies or scoring, while semiconductor applications often use internal data and address only one risk dimension.
3 METHODOLOGY
The methodology builds an OWL-backed knowledge graph from public corporate disclosures, consolidates extracted items through semantic clustering, and ranks them with event-aware scoring before expert validation. Its design combines taxonomy, provenance, typed relationships, recency, historical events, and relevance adjustment.
- Pipeline: The six-stage pipeline defines taxonomies, formalizes them in OWL, retrieves documents, extracts items, clusters duplicates, and scores risks and opportunities.
- Ontology Design: The ontology contains 9 risk categories, 90 risk subclasses, 8 opportunity categories, and 90 opportunity subclasses, enriched beyond a simple hierarchy.
- Ontology Design: The OWL ontology uses disjoint Risk and Opportunity classes, typed relationships, provenance, and twelve supply-chain clusters covering 128 companies.
- Source Retrieval and Data Extraction: Source retrieval uses company-disclosed annual reports, filings, ESG and sustainability reports, presentations, and governance documents, searched through 44 queries per company.
- Source Retrieval and Data Extraction: A two-pass LLM process filters irrelevant chunks, extracts distinct risks and opportunities with taxonomy labels and justifications, and generates OWL instances while preserving provenance.
- Semantic Clustering and Consolidation: Embedding-based clustering uses 384-dimensional Sentence-BERT vectors and cosine similarity threshold τ = 0.85, with Union-Find transitive closure consolidating near-duplicates.
- Impact Analysis and Expert Validation: The deterministic item score is S = w_c · w_t · f_e, combining category importance, publication recency, and event linkage.
- Impact Analysis and Expert Validation: Event-linked items receive higher scores through β = 0.3, while the logarithmic event factor limits the influence of any single historical event.
4 RESULTS
Across five semiconductor companies, the pipeline generated and consolidated large volumes of scored risk and opportunity data, then ranked items through algorithmic scoring, LLM relevance adjustment, and expert comparison. Trade restrictions emerged across all companies, while opportunities varied by value-chain position.
- 76,207 scored risk and opportunity individuals were produced across the five target companies.
- Semantic clustering reduced company-level extractions by 21–70%, with the largest decreases for Infineon and Siltronic.Infineon decreased by 70.3% and Siltronic by 64.7%, while Texas Instruments decreased by 21.1%.
- Financial and strategic risks accounted for roughly half of extracted risks, while financial and operational opportunities dominated the opportunity side.
- Recent geopolitical and trade-policy actions received the highest event scores, with US export restrictions on China’s chip sector ranking first at 1.97.Events from the 2020s averaged 1.34 versus 0.59 for events from the 2000s.
- Geopolitical risk ranked first and strategic opportunity ranked first in algorithmic rankings across all companies.The category weight for geopolitical and strategic items was wc=1.4.
- LLM relevance re-scoring moved items by 20–33 rank positions on average, with maximum displacements of up to 93 positions.Infineon’s top algorithmic risk fell from rank 1 to rank 8, while a manufacturing-concentration risk rose from rank 77 to first place.
- Average LLM-to-expert correlation was ρ=0.55 for risks and ρ=0.72 for opportunities.The strongest agreements were Infineon risks at ρ=0.97 and Air Liquide opportunities at ρ=1.00.
- Trade restrictions appeared among the selected risks for all five companies, while selected opportunities differed by value-chain position.Examples included market expansion for Texas Instruments, foundry diversification for Intel, and automotive electrification for Infineon.
5 DISCUSSION
The pipeline achieves strong extraction and knowledge-graph utility, while LLM re-scoring generally improves alignment with expert judgments, especially for opportunities and peripheral companies.
- Extraction quality: 92.6% precision is achieved when Partial items are included, compared with 76.0% strict precision in the 500-item extraction sample.Risk extractions are more precise than opportunity extractions, at 85.6% versus 66.4% strict precision.
- Subclass assignment accuracy: 70.8% subclass-level accuracy rises to 89.4% at the parent-category level, with errors concentrated among semantically close categories.Recurring ambiguities involve Operational versus Financial, Technological versus Strategic, and Financial versus Compliance classifications.
- Clustering effectiveness: 0.89 average intra-cluster cosine similarity is reported for clusters of size ≤5, which comprise 85.6% of all clusters.The weak correlation between subclass agreement and embedding similarity is r=0.23, reflecting different inputs for clustering and classification.
- Knowledge graph utility: 80.2% of risk subclasses are industry-wide, while three-hop provenance tracing links historical events to individual risk instances.The graph also propagates 30 historical events across all five companies and supports year-over-year and value-chain comparisons.
- Scoring validation: LLM re-scoring raises expert agreement from ρ=0.44 to 0.55 for risks and from 0.36 to 0.72 for opportunities.The largest gains occur for companies peripheral to the semiconductor core, while Infineon’s algorithmic ranking already reaches ρ=0.97.
6 LIMITATIONS AND FUTURE WORK
The pipeline is constrained by language, retrieval, clustering, ingestion, coverage, validation, and recall limitations that motivate several future extensions.
- Data coverage: English-only processing under-covers companies such as Air Liquide, whose filings are largely in French.Multilingual models are identified as a direct way to address this gap.
- Data coverage: Retrieval uses 44 query templates per company, excluding document types such as earnings-call transcripts and supplier audit reports.Vision-language models could broaden evidence extraction from charts and infographics in annual reports.
- Data quality: Cross-company source contamination accounts for most invalid extractions, motivating metadata pre-filtering and post-extraction attribution checks.The proposed mitigation combines document-header filtering with an attribution check after extraction.
- Clustering: Single-linkage clustering with fixed threshold τ=0.85 can create overly large clusters in highly repetitive corpora.Average-linkage clustering and sensitivity analysis across thresholds are proposed to bound worst-case cluster sizes.
- Scope and validation: The system processes documents in a single batch, covers only five of twelve ontology-defined supply-chain clusters, and relies on one domain expert per company.Future work includes incremental ingestion, broader company coverage, and multiple-expert validation.
- Evaluation: No formal recall measurement has been performed because it requires human-annotated ground truth.A targeted recall study on a small annotated document subset is proposed.
7 CONCLUSION
The paper presents a six-stage pipeline that converts corporate disclosures into scored, ontology-backed knowledge-graph intelligence for semiconductor supply-chain risks and opportunities. Applied to five companies, it produces 76,207 items with 92.6% precision, expert correlations of ρ=0.55 for risks and ρ=0.72 for opportunities, and trade restrictions as the dominant systemic risk.
- Conclusion: 76,207 risk and opportunity individuals are extracted from five companies spanning the ALLPROS value chain with 92.6% precision.The items are consolidated through embedding-based clustering.
- Conclusion: ρ=0.55 for risks and ρ=0.72 for opportunities are the correlations between LLM-adjusted rankings and domain-expert judgment.The rankings use a three-layer scoring mechanism.
- Conclusion: Trade restrictions are identified as the dominant systemic risk across all value-chain positions, while selected opportunities are complementary across companies.The conclusion connects these results with a collaborative alliance strategy.