Thursday, March 28, 2024

Using a GPT to get updates in topics of interest

About a month ago, I created a GPT based on ChatGPT 4.0.  It's easy to create but requires some fine-tuning.  I used Ross Dawson's approach detailed here:  Creating custom GPTs for news and Information Scanning; and I adjusted it to suit my own purpose.  The results have been mixed but I'm reasonably happy with what I got today. 

Prompt:  What is the latest news about knowledge graphs?

GPT answer

The latest discussions and advancements in knowledge graphs (KGs) emphasize their growing importance and integration with other technologies like large language models (LLMs) and digital twins, highlighting their pivotal role in AI, information retrieval, and supply chain management.

1. Knowledge Graph Conference 2024: Scheduled for May 6-10, 2024, at Cornell Tech NYC, this conference aims to unite industry leaders and researchers to shape the future of semantic technologies and AI. The event features over 80 presentations, 20+ workshops, and masterclasses, facilitating significant networking opportunities. (Source: Knowledge Graph Conference May 2024.)

2. Unifying Large Language Models and Knowledge Graphs: This area explores the synergy between KGs and LLMs, such as ChatGPT and GPT-4, focusing on overcoming the limitations of LLMs in capturing and accessing factual knowledge by leveraging the structured, factual framework provided by KGs (Source: Unifying Large Language models..., 2023).

3. Knowledge Graphs & LLMs in Multi-Hop Question Answering: The integration of KGs with LLMs facilitates more effective multi-hop question answering by leveraging both structured and unstructured data. This combination allows for improved information retrieval, especially in complex queries that span multiple steps or require understanding connections between disparate pieces of information (Source: Knowledge Graphs and LLMs: Multi-Hop Question Answering), June 2023.

4. Learning to Update KGs by Reading News: This research explores automatic updating of KGs using information from news streams, highlighting the challenges in current text-based KG updating methods that rely on domain-specific, hard-to-maintain rules and systems (Source: ACL Anthology, 2019) .

5. Digital Twins and Knowledge Graphs for Supply Chain Resilience: Knowledge graphs underpin modern digital twin technology, offering scalable, detailed insights into complex, interconnected systems. This application is particularly notable in enhancing supply chain resilience, exemplified by companies like Caterpillar incorporating knowledge graphs for better asset management (Source: Digital Twins, knowledge graphs, and boosting supply chain resilience, 2022).

6. Applications of LLMs and KGs: Highlighting multi-hop question answering and the integration of textual data with KGs, the use of LLMs alongside KGs addresses challenges in information retrieval and enables handling complex questions that span across multiple documents or require connection between various data points (Source: LLMs & Knowledge Graphs, September 2023).

Importance: The evolving landscape of knowledge graphs signifies a substantial shift towards more interconnected, intelligent systems that can manage complex information networks. This is especially relevant in enhancing AI's understanding and processing capabilities, augmenting supply chain resilience, and enabling more sophisticated question-answering and data retrieval methodologies.

These developments represent critical steps forward in harnessing the power of knowledge graphs across various domains, from AI research and semantic web technologies to practical applications in supply chain management and digital twin technologies.

How is this useful (or not)?

  • It is a discrete number of items that I can relatively quickly review to determine whether useful for me to read more thoroughly or not.
  • The short descriptions tell me something before I even track down the source.
  • This particular set of answers were mostly high value.  It's not always the case but tuning the GPT could help fix that by more judiciously selected sources. It's also been difficult to get it to focus on the most recent resources.  This answer included mostly 2023-2024 sources, but it included a 2019 document.  It was an interesting one but advances in graphs and LLMs suggest the challenges of 2019 may have been addressed by now. 
  • It takes me to resources I was not aware of and would not necessarily come across doing a Google search. 
  • I also find the way ChatGPT provides sources to be very unfriendly.  I had to manually recreate the links for the above.  Perhaps it's something I can adjust in the GPT's instructions.  

No comments: