AriGraph (Part 2) - Knowledge Graph Construction and Retrieval Details
12Excited to have Petr run through the inner details about AriGraph! We go through how the Knowledge Graph is constructed using LLMs to generate triplets from free text and how the old triplets are updated.
We also go through how the retrieval is done by firstly coming up with a list of important queries (along with importance score) to search for based on the goal and the observation, then iteratively finding out semantically similar triplets to these queries. The number of iterations depends on the importance score. Finally, the episodic memory is added back as the observation text for the episode with the most number of triplets selected.
Overall, it is a very exciting approach how the knowledge graph is updated dynamically, and how the knowledge graph is retrieved in an iterative manner as well.
See Part 1 here: • AriGraph: Learning Knowledge Graph Wo...
~~~
0:00 Introduction
1:05 Constructing Knowledge Graph
5:53 Updating Knowledge Graph
9:50 How the Subset of Triplets are Selected for Updating
13:46 How disambiguation of terms are done
17:41 Why disambiguation may be better at retrieval than at construction (My Insights)
20:14 Retrieval of Knowledge Graph with Iterative Expansion
43:05 Iterative Search with ReAct (future work)
46:04 Incorporating Episodic Memory
55:13 Q&A Benchmark with AriGraph
1:01:30 Conclusion
~~~
AI and ML enthusiast. Likes to think about the essences behind breakthroughs of AI and explain it in a simple and relatable way. Also, I am an avid game creator.
Discord: / discord
LinkedIn: / chong-min-tan-94652288
Online AI blog: https://delvingintotech.wordpress.com/
Twitter: / johntanchongmin
Try out my games here: https://simmer.io/@chongmin