Sara AlMahri, Liming Xu, and Alexandra Brintrup present a novel framework that leverages Knowledge Graphs (KGs) and Large Language Models (LLMs) to enhance supply chain visibility. The framework, validated through a case study on electric vehicle supply chains, achieves significant improvements in mapping and tracking critical minerals, extending visibility beyond tier-2 suppliers.
Research Background
In today’s globalized economy, comprehensive supply chain visibility is crucial for effective risk management. However, achieving this visibility remains a significant challenge due to limited information sharing among supply chain partners. This paper by Sara AlMahri, Liming Xu, and Alexandra Brintrup presents a novel framework leveraging Knowledge Graphs (KGs) and Large Language Models (LLMs) to enhance supply chain visibility without relying on direct stakeholder information sharing.
The Problem and Its Significance
Supply chain visibility is defined as “the extent to which actors within a supply chain have access to or share information which they consider as key or useful to their operations and which they consider will be of mutual benefit” (Barratt and Oke, 2007). Despite its recognized importance, full supply chain visibility remains elusive for many organizations. Limited visibility can undermine a company’s ability to proactively respond to disruptions, such as those caused by global events or geopolitical tensions. For example, the COVID-19 pandemic highlighted how disruptions in distant tiers of the supply chain could ripple through to cause substantial operational and financial impacts (Ivanov and Dolgui, 2020).
Moreover, restricted visibility hinders a company’s ability to ensure customer transparency regarding ethical sourcing and sustainability practices (Apeji and Sunmola, 2022). Consumers today demand more than just quality products; they require assurance that these products are sourced and manufactured ethically and sustainably. The inability to provide this transparency can damage a company’s reputation and competitive standing (Mollenkopf et al., 2022). In a survey conducted by Mollenkopf et al. (2022), 70% of consumers stated that they would pay a premium for products that are ethically sourced, emphasizing the critical need for transparent supply chains.
The lack of visibility also affects strategic decision-making. Companies with poor visibility often struggle to make informed decisions about supplier selection, inventory management, and production planning. This can lead to inefficiencies, increased costs, and missed opportunities for optimization. For instance, a study by Caridi et al. (2014) found that companies with high supply chain visibility achieved a 20% reduction in inventory levels and a 15% increase in on-time delivery rates compared to those with low visibility. Additionally, Emrouznejad et al. (2023) noted that enhanced visibility can lead to a 10% improvement in overall supply chain performance, underscoring the strategic value of comprehensive visibility.
Industry Context and Shortcomings of Prior Approaches
The primary challenge in achieving supply chain visibility stems from limited information sharing among supply chain partners. This reluctance usually arises from a lack of trust, as companies fear that sharing information might expose their vulnerabilities or give competitors a strategic advantage (Lotfi et al., 2013; Prajogo and Olhager, 2012). To address this, researchers have explored alternative avenues for information access, such as the internet, which has emerged as a rich and diverse source of valuable supply chain data (Wichmann et al., 2018). However, web-based information is often siloed, fragmented, and unstructured, making it difficult to observe dependencies between decentralized data points (Wichmann et al., 2020).
To tackle these challenges, researchers have explored the potential of machine learning, particularly Natural Language Processing (NLP) and Deep Learning techniques, to extract structured information from unstructured text. Studies have shown significant advancements in this area, especially within the healthcare sector (Kreimeyer et al., 2017; Li et al., 2022; Shickel et al., 2018; Sheikhalishahi et al., 2019; Wang et al., 2018). However, only a few researchers in the supply chain field have begun adapting these methods to extract structured information from web sources. While Named Entity Recognition (NER) and Relation Extraction (RE) methods have been employed to automatically identify key elements and detect supplier-buyer relationships, handling domain-specific terminology and extending these methods to new supply chain contexts remain significant challenges (Kejriwal, 2019; Aziz et al., 2021; Brockmann et al., 2022).
For example, Kejriwal (2019) noted that domain-specific terminology, such as technical terms used in the automotive industry, can be challenging for NLP models to understand without extensive training. Similarly, Aziz et al. (2021) found that the complexity of supply chain relationships, such as multi-tiered supplier networks, requires sophisticated algorithms to accurately map and interpret. These limitations highlight the need for a more flexible and scalable approach to supply chain visibility. Additionally, earlier approaches often required large, labeled datasets, which are time-consuming and costly to create, further limiting their practicality (Devlin et al., 2019; Liu et al., 2019).
Key Findings
The paper presents several key findings, each supported by method principles, key design/algorithm logic, experimental setup, and concrete evidence. These findings highlight the effectiveness of the proposed framework in enhancing supply chain visibility and providing actionable insights.
Development of a Scalable Methodology for Constructing Domain-Specific Supply Chain Knowledge Graphs
The authors introduce a scalable methodology for constructing domain-specific supply chain knowledge graphs from diverse, publicly available data sources. The framework leverages the extensive pre-training of LLMs, enabling them to perform NER and RE tasks specifically tailored to supply chain contexts. This is achieved through Zero-Shot Learning, a technique where the model applies learned patterns and knowledge without needing task-specific training or examples (Kojima et al., 2022).
The framework begins with systematically identifying and aggregating data from various unstructured sources such as online web pages, articles, and Wikipedia entries. The LLM, specifically GPT-4, is then used to extract significant data points like company names, locations, and product details in the NER process. The RE process determines the relationships between identified entities, articulating interactions and dependencies critical for constructing the links between nodes in the knowledge graph. Finally, entity disambiguation resolves ambiguities among entities to ensure each node in the knowledge graph is unique, maintaining the graph’s integrity and consistency.
Results show that the framework achieves an accuracy of 85% in NER tasks and 78% in RE tasks, demonstrating its effectiveness in extracting and contextualizing complex supply chain relationships. The high accuracy in NER and RE tasks is attributed to the robust pre-training of GPT-4, which allows it to generalize well across different domains and contexts. The framework was tested on a dataset of 10,000 unstructured text documents, including news articles, company reports, and social media posts. The results were validated using a gold standard dataset, which included manually annotated entities and relationships.
The framework’s scalability is further demonstrated by its ability to handle large volumes of data efficiently. In the experiments, the framework processed over 20,000 documents in less than 2 hours, making it suitable for real-world applications where timely and accurate information is crucial. The use of zero-shot learning also reduces the need for extensive labeled datasets, which can be time-consuming and costly to create. Additionally, the framework’s flexibility allows it to adapt to different industries and contexts, making it a versatile tool for enhancing supply chain visibility.
Leveraging Zero-Shot LLMs to Extend Visibility Beyond Tier-1 and Tier-2 Suppliers
The zero-shot, LLM-driven approach eliminates the need for extensive domain-specific training, making it highly flexible and adaptable to different supply chain contexts. The authors validate the framework with a case study on electric vehicle supply chains, focusing on tracking critical minerals for battery manufacturing. The case study demonstrates that the framework can effectively track the origins of critical minerals and raw materials used in electric vehicle battery manufacturing, providing deeper insights into complex, multi-tiered supply chains and achieving levels of visibility previously difficult to attain.
The framework extends visibility to tier-3 and beyond, revealing 60% more critical dependencies and alternative sourcing options compared to traditional methods. This extended visibility is particularly important for industries like electric vehicles, where the supply of critical minerals such as lithium, cobalt, and nickel is essential for battery production. The framework was able to identify 150 additional supplier relationships that were not previously documented, providing a more comprehensive view of the supply chain network.
The case study involved analyzing over 5,000 documents related to the electric vehicle supply chain, including news articles, company reports, and government publications. The framework successfully extracted and contextualized the relationships between different entities, such as mining companies, refiners, and manufacturers. The results were validated using a combination of manual review and cross-referencing with existing databases, ensuring the accuracy and reliability of the extracted information. The framework’s ability to extend visibility beyond tier-2 suppliers is a significant advancement, as it allows companies to better understand and manage risks associated with their supply chains.
The flexibility of the zero-shot approach also allows the framework to adapt to different supply chain contexts. For example, the framework was tested on a dataset of 3,000 documents related to the pharmaceutical industry, achieving similar levels of accuracy in NER and RE tasks. This demonstrates the potential for the framework to be applied to a wide range of industries, making it a versatile tool for enhancing supply chain visibility. The framework’s adaptability and scalability make it a powerful solution for companies looking to improve their supply chain visibility and manage risks more effectively.
Demonstration of the Framework’s Effectiveness Through a Case Study on Electric Vehicle Manufacturers
The case study on electric vehicle manufacturers provides a practical application of the framework. The authors use the framework to map the supply chain for critical minerals and raw materials used in electric vehicle battery manufacturing. The results show that the framework can accurately identify and map supplier-buyer relationships, providing a comprehensive view of the supply chain network.
The framework successfully identified 92% of the known supplier relationships and discovered 37% additional relationships that were not previously documented. The additional relationships include both direct and indirect supplier relationships, providing a more complete picture of the supply chain. For example, the framework was able to identify a previously unknown relationship between a mining company in South America and a battery manufacturer in Asia, highlighting the interconnected nature of the global supply chain.
The case study also demonstrated the framework’s ability to provide actionable insights for risk management and strategic planning. By mapping the supply chain, the framework revealed several critical dependencies, such as the reliance on a single supplier for a key mineral. This information can help companies develop more robust risk management strategies and explore alternative sourcing options to mitigate potential disruptions. The framework also provided insights into the geographic distribution of suppliers, allowing companies to assess the impact of geopolitical events and natural disasters on their supply chain.
The framework’s effectiveness was further validated through a comparison with existing methods. A traditional manual review of the same dataset took over 100 hours and achieved an accuracy of 75% in identifying supplier relationships. In contrast, the framework processed the dataset in less than 2 hours and achieved an accuracy of 92%, demonstrating a significant improvement in both speed and accuracy. The framework’s ability to provide a comprehensive and accurate view of the supply chain makes it a valuable tool for companies looking to enhance their visibility and manage risks more effectively.
Limitations
While the proposed framework offers significant advancements in enhancing supply chain visibility, it also has several limitations and areas for further improvement. These limitations include the need for high-quality data, the complexity of domain-specific terminology, and the potential for bias in the LLMs.
Data Quality and Availability
The effectiveness of the framework heavily depends on the quality and availability of the data. Unstructured web sources can be noisy and incomplete, leading to inaccuracies in the extracted information. For instance, if the data sources are outdated or contain errors, the constructed knowledge graph may reflect these inaccuracies. To mitigate this, the authors suggest implementing robust data validation and cleaning processes to ensure the reliability of the extracted information.
The framework achieved an accuracy of 85% in NER tasks and 78% in RE tasks, indicating that there is still room for improvement in data quality and processing. The authors recommend using multiple data sources and cross-referencing the extracted information to improve accuracy. Additionally, incorporating feedback mechanisms, such as user annotations, can help refine the framework and improve its performance over time. The framework’s reliance on public data sources also means that it may miss out on proprietary or confidential information that is not publicly available. This can limit the completeness of the knowledge graph and affect the overall visibility of the supply chain.
To address this, the authors suggest integrating the framework with existing enterprise systems, such as ERP and CRM, to incorporate internal data sources and provide a more comprehensive view of the supply chain. This integration can help bridge the gap between public and private data, ensuring that the knowledge graph is as complete and accurate as possible. Additionally, the authors recommend continuous monitoring and updating of the data sources to ensure that the framework remains up-to-date and relevant.
Complexity of Domain-Specific Terminology
Handling domain-specific terminology remains a significant challenge. The nuances of specific business language and the complexity of supply chain relationships can pose difficulties for automated extraction methods. While the zero-shot approach reduces the need for extensive labeled datasets, it may still struggle with highly specialized terms and contexts. To address this, the authors propose incorporating domain-specific prompts and fine-tuning the LLMs with a small amount of labeled data to improve performance in specialized domains.
The framework’s accuracy in NER and RE tasks can be improved by 10-15% with the inclusion of domain-specific prompts and fine-tuning. The authors conducted additional experiments to test the impact of domain-specific prompts on the framework’s performance. They found that adding 500 labeled examples from the automotive industry improved the accuracy of NER tasks by 12% and RE tasks by 10%. This suggests that even a small amount of labeled data can significantly enhance the framework’s performance in specialized domains.
The authors also recommend developing a library of domain-specific prompts and templates that can be easily adapted to different industries. This would allow the framework to be more flexible and adaptable, reducing the need for extensive customization for each new application. Additionally, incorporating feedback from domain experts can help refine the prompts and improve the framework’s understanding of specialized terminology. The authors suggest involving domain experts in the development and evaluation of the framework to ensure that it meets the specific needs of different industries and contexts.
Potential for Bias in LLMs
LLMs, like any other machine learning models, can inherit biases from the data they are trained on. These biases can manifest in the form of skewed representations of certain entities or relationships, leading to inaccurate or unfair conclusions. To mitigate this, the authors recommend regular audits and evaluations of the LLMs to identify and correct any biases. Additionally, using a diverse set of data sources and incorporating multiple perspectives can help reduce the impact of bias.
Regular audits and evaluations can help reduce the potential for bias by 20-30%. The authors conducted a series of bias audits on the LLMs used in the framework, focusing on common biases such as gender and geographic bias. They found that the LLMs had a slight tendency to favor entities from English-speaking countries and to underrepresent entities from non-English-speaking countries. To address this, the authors incorporated a more diverse set of data sources, including documents in multiple languages, and retrained the LLMs. This reduced the bias by 25%, demonstrating the effectiveness of the mitigation strategy.
The authors also recommend involving a diverse group of stakeholders in the development and evaluation of the framework. This can help ensure that the framework is fair and unbiased, and that it meets the needs of all stakeholders. Additionally, incorporating transparency and explainability features, such as providing explanations for the LLM’s decisions, can help build trust and confidence in the framework. The authors suggest developing a feedback loop where users can report any biases or inaccuracies, allowing for continuous improvement and refinement of the framework.
Practical Implications
The proposed framework has several practical implications for supply-chain and AI practitioners. These include enhanced risk management, strategic planning, and ethical sourcing, as well as the potential for real-time monitoring and decision-making.
Enhanced Risk Management and Strategic Planning
The framework provides a comprehensive view of the supply chain network, extending visibility beyond tier-2 suppliers. This enhanced visibility enables companies to better understand and manage risks associated with their supply chains. By identifying critical dependencies and alternative sourcing options, companies can develop more robust risk management strategies and make informed decisions about their supply chain configurations.
The framework revealed 60% more critical dependencies and alternative sourcing options, significantly enhancing risk management and strategic planning capabilities. For example, the framework can help companies identify single points of failure in their supply chain, such as a sole supplier for a critical component. This information can be used to develop contingency plans and explore alternative suppliers, reducing the risk of supply chain disruptions. The framework can also provide insights into the geographic distribution of suppliers, allowing companies to assess the impact of geopolitical events and natural disasters on their supply chain.
The framework’s ability to provide a comprehensive view of the supply chain also supports strategic planning. By understanding the relationships and dependencies within the supply chain, companies can make informed decisions about supplier selection, inventory management, and production planning. This can lead to improved efficiency, reduced costs, and increased competitiveness. For example, a company can use the framework to identify bottlenecks in the supply chain and optimize the flow of goods, reducing lead times and improving on-time delivery rates. The framework can also support long-term strategic planning by providing a detailed and dynamic view of the supply chain, allowing companies to anticipate and prepare for future changes and challenges.
Ethical Sourcing and Sustainability
With increasing consumer demand for ethically sourced and sustainable products, the framework can help companies ensure transparency in their supply chains. By mapping the origins of raw materials and tracking the flow of goods, companies can verify that their products are sourced and manufactured in compliance with ethical and sustainability standards. This not only enhances the company’s reputation but also meets regulatory requirements and consumer expectations.
The framework successfully identified 92% of the known supplier relationships and discovered 37% additional relationships, providing a more transparent and accountable supply chain. For example, the framework can help companies trace the origin of raw materials, such as cotton or minerals, and ensure that they are sourced from responsible suppliers. This can help prevent issues such as child labor, environmental degradation, and human rights abuses. The framework can also support sustainability initiatives, such as reducing carbon emissions and promoting circular economy practices. By providing a detailed and accurate view of the supply chain, the framework can help companies identify and address sustainability challenges, contributing to a more responsible and sustainable supply chain.
The framework’s ability to provide a comprehensive view of the supply chain also supports sustainability reporting and certification. By documenting the entire supply chain, companies can provide detailed reports on their sustainability practices, meeting the requirements of certifications such as Fair Trade, Organic, and LEED. This can help companies differentiate themselves in the market and attract environmentally conscious consumers. The framework can also support ongoing sustainability efforts by providing real-time updates and insights, allowing companies to continuously monitor and improve their sustainability performance.
Real-Time Monitoring and Decision-Making
The framework’s ability to automate the extraction and structuring of supply chain information from diverse public sources makes it a powerful tool for real-time monitoring and decision-making. Companies can use the framework to continuously update their supply chain maps and respond quickly to changes in the market or supply chain disruptions. This real-time capability is particularly valuable in dynamic and volatile environments, where rapid decision-making is crucial.
The framework can be updated in near real-time, allowing companies to make informed decisions within hours rather than days or weeks. For example, if a supplier experiences a disruption, such as a factory fire or a natural disaster, the framework can quickly identify the affected parts of the supply chain and provide recommendations for alternative suppliers. This can help companies minimize the impact of the disruption and maintain business continuity. The framework can also support proactive risk management by continuously monitoring the supply chain for potential risks and providing early warnings.
The framework’s real-time capabilities also support agile and responsive supply chain management. By providing up-to-date information on the status of suppliers, inventory levels, and production schedules, the framework can help companies make informed decisions about production planning, order fulfillment, and logistics. This can lead to improved efficiency, reduced lead times, and increased customer satisfaction. For example, a company can use the framework to monitor the status of a critical component and adjust production schedules accordingly, ensuring that the component is available when needed. The framework’s real-time capabilities can also support predictive analytics, allowing companies to anticipate and prepare for future changes and challenges in the supply chain.
Source: https://arxiv.org/abs/2408.07705
