Communicating With The Machine

 

If we want to build an AGI machine, communicating with the machine will be essential – not the trivial approach of trying to change the statistics on the words, but using words with their precise meanings. Words can have many meanings, and they change their meanings in the presence of other words – “on” can be an adverb or a preposition, some words have dozens of meanings. But there is much more than the meanings of words – we mention some other aspects.

Parent Information

The MULTIPARENT operator details precisely what a word is doing, but sometimes you can’t even tell whether a word is a noun or a verb. What happens then?

We may encounter a point of uncertainty. We can be unsure about the part of speech (a noun or a verb?), the type of noun (Simple, Infinitive, Clausal) or verb (transitive, ditransitive, or a lot of others), and the specific definition. These points need to be marked and monitored, and if necessary decided on without sufficient evidence.

Sometimes you can’t even tell whether a word is a noun or a verb. What happens then?

With definitions, sometimes the best we can do is reduce the pool of possibilities (worthwhile when for some words we have an initial set of twenty different definitions).

A point of uncertainty can be tracked – if it is not relevant to the problem at hand it can be ignored.

Speed

This all sounds very slow? Important decisions take time, particularly if a complex scenario is being played out over time. Some of the uncertainty can be stripped out beforehand, while other uncertainties must remain to define the problem.

Collocated Verbs

There is a heap of collocated verbs – fed up with, tire out, tiring out, tired out, shut off, shut down, shut out – about 2,500 in total. They are expensive, as each one requires one or more wordgroups - but often the meaning is not easily deducible from the components.  “fed up” might mean you have been fed too much.

There are gerunds (tiring out the opposition is a strategy) and adjectives (the tired out old man) mixed in as well.

Tire out as a verb, with its present participle also as a gerund, and its past tense as an adjective.

If the system is going to work well, it needs a mountain of detail - detail which only your Unconscious Mind can think about in any detail.

With existing analytic tools, we are used to simplifying the problem. When problems get complex, that is a bad strategy, as the simplification may prevent any solution being found.

Passive Verb Phrase

A passive verb phrase is modelled by keeping the subject/object form of active verb phrases, allowing a subject to be given by a “by” preposition.

The phrase “something derived” looks like

A “that” and an “is” are inserted, then a SPLITTER to the subject. The “that” is connected to the object end of the verb phrase.

If no candidate for the subject of the verb phrase can be found, an UnknownObject is connected.

Turning a description in English into an active structure is worthwhile, but sometimes the words are clumsy. A fragment of the definition of a mortgage:

“the transfer of title is voided upon the payment of the debt”

If you are paying down the debt every month, “the payment of the debt” doesn’t sound very specific. What is meant is “full and final payment of debt”, but this isn’t very specific either. When “the loan balance falls to or below zero” is much more precise.

What about the furphy that people rarwly consult a dictionary, but this system uses a dictionary for every word? Yes, but people take 20-25 years to build an abstract picture of the world, we can't afford that time, and people can't reason deeply - the severe limit on their input won't allow it.

Conclusion

We haven’t shown any formulae in action – this is trivial in comparison with English. We built an analytic system to handle formulae a long time ago, but the system without the connectives of a natural language interface to tie it all together was too hard to use.

Some numbers:

Words                                  45,000
Wordgroups                       10,000 (fed up, run amok, shut the stable door, etc.)
Word definitions               87,000
(not all words have definitions - only the base form of a verb is defined)

We hope we have given you some insight into the workings of an active semantics model. It is intended for large and complex problems, where a human has severe limitations, or as an explanatory tool for complex text (pieces of legislation or large specifications), where the text serves as an abstract working model, and can be poked and prodded to see what it does.

An AGI machine will need to be communicated with – the easiest way is to use a natural language like English. Solving that problem solves a lot of other problems too. You will be able to tell it when it goes wrong and how to fix it, or it may be able to point out why you are mistaken.