Data viz biz Tableau forks out for natural language startup Data visualisation firm Tableau has made its third-ever acquisition in a bid to speed up use of natural language query technology and bring in more users. The start-up, ClearGraph, was founded in 2014 and chugged down $1.53m in seed funding in 2015. According to a canned statement, its platform offers non-experts the ability …
No doubt this time round it will involve machine learning because.....
Snarky comments aside there are 3 big problems with asking questions in a natural language.
1) It's vocabulary is never complete. Humans make new nouns and verbs at the drop of a hat. So IRL the classic step 1 of an NL algorithm (look up all words in the sentence) is actually BS.
2) There is probably no complete grammar for any real natural language anywhere. Some do exist but they are f**king huge, compared to those for even big languages like Ada, C++ (claimed to be impossible without parsing the whole source 3 times) and COBOL.
3) Questions like "How many really expensive houses are for sale in London right now" have huge implicit context associated with them, not to mention they are only appropriate in the context of a data base of UK (or global?) house prices. If you're DB does not have that information the question is basically meaningless.
What I could see is a system that has to have items explicitly explained to it (what does "very expensive" mean in this context?) and retains that data for reuse, so that over time its answers become more intelligent sounding (like a real PFY learning on the job, hopefully not turning into a BOFH)
What has changes is a lot more on line resources of things like parts of speech dictionaries to help with the brute force task, and of course processor speeds rising by a 1000 and memory sizes by a 1000 also (which sounds like any increase in processor speed has been cancelled by increase in the search space ;-( )
I cannot believe that this is the first thing I see this AM. I was just thinking about creating a class I can teach children with the above title. Just 15 minutes ago I started putting together words that should exist but make no obvious sense.
I even came up with a 3 letter word with no definition in English.
FIK
PLAP no definition
TARNK no definition
STAMBLE no definition
But better are some that do have definitions:
────────────────────────────────────
STARP
Noun:
in geometry, the pentagrammic antiprism is one in an infinite set of nonconvex antiprisms formed by triangle sides and two regular star polygon caps, in this case two pentagrams. This polyhedron is identified with the indexed name U79 as a uniform polyhedron.
────────────────────────────────────
GLAUD
Noun. ward, seal, type of barrier : A juffetless glaud is no glaud at all. (somebody playing games??????)
────────────────────────────────────
FLUK Fluctuates
FIK (Urban DIctionary = Fuck I Know) but nothing in English. I speak a number of languages and FIK does exist in quite a few but not English.
han fik et blåt øje - he got a black eye in Danish
────────────────────────────────────
Albanian
fik (first-person singular past tense fika, participle fikur)
to extinguish (fire) {Your'e fired?}
to switch off, to turn off
to bring misfortune, to ruin, to destroy
────────────────────────────────────
So let's just fik a glaud.
or
I want a house with a starp and a juffetless glaud!
That's sort of the point. I'd never heard of "STARP" till you mentioned it.
There is work to deduce what kind of word a word is by its place in a sentence and assuming the sentence is grammatical.
Which, if you're looking at a system that learns to speak a human language the way a human does may be more appropriate.
Despite the fact that ultimately the human brain is a huge multi layer neural net I don't think, and all human thought maps to that architecture, I think the brain uses intermediate abstractions and massive training sets won't help identify what they are.
I am doing brain mapping right now at the local university. Our neural structure can be amazing. For vision the entire retina is one to one mapped on the front of the occipital lobes with inversion taking place there. For language there are two main areas. Nouns including names are mapped into Brocca's area close behind the frontal lobes on the sinister (left) side of the brain. It is incredibly precise. It is like very thin layers of neural tissue and those layers are mapped according to age when language(s) are learned.
One of the mistakes most of us make is to use too much "baby language". Instead we should use the actual language with simple and very clearly spoken words. We all have the most neurons in the brain at about 3 months after birth, if not premature. From that point forward it is a matter of use it or lose it. That is called Synaptic pruning. The more new and interesting things we are exposed to from just after birth the more of those neurons we retain. It is why we will often put up a spinning toy with bright colours and similar items. Three dimensional toys are by far the best (no stupidity phones with flat interfaces).
Language is stored in the layers as time passes. The other areas are Wernicke's speech area and the posterior sinister fusiform gyrus. Wernicke's is used primarily for verbs, adverbs and similar as well as conjunctions etc. The posterior fusiform gyrus is for reading. Even tiny amounts of damage there can make it nearly impossible to read or at the best just a single character at a time.
In Brocca's area the words one learns later in life are more easily affected by even very tiny amounts of damage. The damage is often not direct but is called a disconnection syndrome. That is where the white matter (axons, the "wiring") are cut somehow. The memory is intact but access is limited.
The main thing is that brain structure is a huge number of very small entities that each have very specific functions. Also, it is analogue, not digital in nature.
When it comes to language what you learn very early is what you will know the rest of your life. Brain pruning is finished at late puberty. Learning any other language is then more difficult. I learned English and Danish at the same time. English has about 34 or so phonemes but Danish has 52. That is the largest number for Western European languages. It makes it much easier for me to learn other languages anywhere. I am multi lingual as is my mother.
There is far more I could say about all of this but it would be much better if you could somehow attend one of my courses I am planning, even one for children.... ☺
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
