Conclusions

The human langauge model is probably 10^{9 ± 1} bits, in agreement with Turing

• Demonstrated the feasibility of a neural network framework.
• Need to develop a full model (character + lexical + semantic + syntactic).
• Need to evaluate on a standard test set.
• Need to compare with human performance.

Human language is probably machine learnable from a text corpus

• Demonstrated for lexical acquisition.
• Need to demonstrate for syntax.

AI is unsolved - but we are close

• Simulation refines Shannon's 1950 results (the goal) to about 1 bpc.
• The best models (lexical + semantic) have achieved 1.2 bpc.
• The goal needs to be refined by human testing.