There's so much excitement about the Neural Network implementations for understanding and finding answers in short paragraphs of text. Unfortunately this approach never scales very well.
New attempts to add 10^19 parameters to encode it all still won't work. The reason why is that you really aren't understanding what is being said nor what is being said in the content.
That's why we place our efforts into advanced grammar and sentence morphological analysis. This is different from Google Universal Sentence Encoder which is a N dimensional hyperspace. When I tested these kinds of solutions they didn't do very well at all.
We had a run in with someone with a statistics background who claimed that NLP didn't work as well. Yah, well what did they do? You always have to ask exactly what was tried. Simple part of speech? yah that doesn't work.
Graph encoding morphological grammar is our state of the are technique but we don't stop there. We then apply a grammar transformation and world-map modeling as well as a few other things to map a conversation. I'm sure behind the scenes they are starting some of that at Google and Amazon but not qute all of it. They will always be handicapped by their core neural network approach. And I'm someone who's been a pioneer and believer in neural networks - for what they do well - for well over 30 years. I believe 1986 was my first one.
The reason why most people don't use grammar is they don't know the theory and they aren't long time researchers in the NLP field. It's a lot easier to get excited by the new stuff and believe thats the cool stuff. Well in some ways they are right, but not in accuracy, precision, or ability. Not yet. Maybe in ten years. They have to go through similar efforts in research.
It can be frustrating running a language tech startup when others have a hard time understanding why what you are doing is better, until they see the results. Then it's just bang obvious. Fast. Precise. Great stuff.
And that continues to get even better as we move into voice and conversational systems.
Like Tesla, we build our company layer by layer, technology advance ontop of technology advance.
Eventually we will arrive at very natural human seeming systems interacting with massive amounts of information and able to encode facts and world information in a fully autonomous form. Maybe that's 2030. But what other companies are striving for that? Not Amazon or Google, their approaches are intellectual dead ends, a party trick that does the basics well but won't scale.
No comments:
Post a Comment