Last month, GraphAware data scientists attended Nodes 2023 – a 24-hour online graph community gathering dedicated to learning how to integrate graph technologies into ML and dev projects. The hot topics this year were:

  • Building Intelligent Applications: APIs, Libraries, and Frameworks
  • Machine Learning and AI
  • Visualization: Tools, Techniques, and Best Practices

To contribute to all the inspiring content on the latest innovations in graph technology for applications and ML models, our scientists discussed Entity Resolution (ER) and Relation Extraction, comparing Dependency Graphs vs Large Language Models.

Entity Resolution (ER)

ER, in a nutshell, is reconstructing and identifying a unique entity whose information is parsed among multiple databases. This means consolidating various heterogeneous pieces of information into distinct personas or organisations. For intelligence agencies, this could mean detecting the same, unique suspect from different reports. Likewise, it’s useful in detecting financial crime – a criminal can register under one name in the tax registry, a different name in the National Registry, and yet another for a bank account in an attempt to cheat the system.

Using Knowledge Graph software, like Hume, allows holistic and complete investigations that allow you to see the same person from different perspectives, different databases, connected by links that may be hard to miss. Sometimes, one can discover new information by combining these single pieces and investigating the links.

In our section presented by Giuseppe Futia (Senior Data Scientist), Pantelis Krasadakis (Technical Sales Engineer), and Federica Ventruto (Junior Data Scientist), we will show how Hume facilitates ER and our approach to a real scalable end-to-end solution :

Relation Extraction: Dependency Graphs vs Large Language Models

Relation Extraction is the identification of semantic relationships between entities, and classification of them into predefined categories. Knowledge Graph-building from unstructured data is a necessary but also challenging task, which is where Relation Extraction comes in. For example, take an intelligence agency investigating a complex web of corruption. Should they want to analyse thousands of pages of text to extract insights, they have to apply different pipelines performing different tasks related to the natural language processing domain – and one of them is relation extraction, essentially the task to extract the relationships between entities, in this case people, organisations and money. Main techniques to accomplish this task include Dependency Parsing or Large Language Models.

We explored how to leverage dependency graphs and GPT language models, comparing them using the real-world Gupta Leaks scandal and highlighting some interesting insights from graph data science.

Discover more with Federica Ventruto, Junior Data Scientist:

Watch GraphAware data scientists at Nodes 2023 unpack the power of Knowledge Graphs, with Entity Resolution, Relation Extraction and LLMs.