Ahh the utter crap that's out there by dumb engineers who don't have a clue what's going on. Hence the picture above to illustrate "Graph Neural Networks" for NLP. Ugh. Makes my brain hurt. They've copied Tnsor architecture YET AGAIN.
No, au contraire mon frere as Groening would say, this is exactly what a GNN is NOT.
A true GNN for NLP uses a graph structure for encoding the relationships ala NELL or DIFFBOT, and then while one might encode them into a graphbase, actually you want a hybrid structure graph / neural network. Such that the relationship learn and respond to stimulus, and extend themselves ala Edelman and Pribram.
There is at present NO COMMERCIAL AVAILABLE technology that does this. Which is why we get dingbat diagrams as per above.
This is the core of Noonean's SULU technology for language. True supervised learning of language, then extended into unsupervised learning understanding. SU/LU. SULU ties together everything that's happening at Noonean - eConcierge for conversational AI, EnterpriseNLP for vast search, and finally SULU, the language model for understanding.
No comments:
Post a Comment