【Session Theme: AGI】
With Articut NLP AI system, we build a knowledge graph frame that has both “core knowledge” and “contextual experience.” Its name is AItlas. The features of AItlas include:
- The model is generated via automatic text-reading, it doesn’t require any supervised tagging or knowledge classification.
- The data format and the content of the model is “human-readable.” Besides, the programming logic is transparent and that makes it possible to verify the operation of any AI system that adopts AItlas.
In practical scenarios, the “core knowledge” will be released by the manufacture (Droidtown Linguistic Tech), and the “contextual experience” will be cumulated as the AI system interacts with humans. With this arrangement, the AI system can not only know the basic principles of how the world works, but also know that under some context, some humans require unique interaction. One thing worth to mention is that these two layers of knowledge (the core one and the contextual ones) are not required to be uploaded to the cloud. That means, all human rights of privacy is safe with the system since it works perfectly within a local machine.
Knowledge Graph works with statistics, graph theory and computer to convert texts into visible network to illustrate the inner structure of a scientific system. For example, “Mt. Lala” is classified as a “Natural Landscape”, and a “Human Activity Area” when it’s also “Part of Taoyuan City”. These classification tasks require domain experts manually designate what classes an entity or event belongs to. Knowledge Graph is considered as a difficulty that needs to be solved so AI can go into next generation.
Droidtown Linguistic Tech. proposes a Hybrid AI approach to build AItlas knowledge graph system. On one hand, it parses and understands the contents of the texts via a rule-driven NLP engine (Articut), on the other hand, it self-learns the relationships among entities in the texts unsupervised. The process doesn’t require involvement of domain experts, but the human-readable output also preserve the possibility that human experts CAN put their hands into the knowledge graph.
An early experiment shows that after an AI system learns that “a dog” is a kind of “animal” with AItlas, it automatically inferences that “a cat” is also a kind of “animal” since their syntactic structure are the same and the context are extremely similar. That is to say, AItlas helps an AI system learned what “a cat” is even when it is the first time it meets “a cat” in the context.
If there’s no further instruction or inference in the contextual experience, AItlas would suggest that “a cat” should be treated as if it is “a dog.”
This seemly problem can easily be solved due to the human-readable data structure with clear semantic logic represented. Human can perform a backward lookup and see the root cause that triggers a mistake then fix it. We believe that only when an AI is thinking in humans’ ways, the AI system is a trustworthy partner.
About PeterWolf (Wen-jet Wang)
Founder of Droidtown Linguistic Tech. Co., Ltd.