An Expedient Mind: PREFACE

AN EXPEDIENT MIND

Tartan Hen Publications : Home | more books | Contact : feedback@tartanhen.co.uk

LANGUAGE

Preliminary comments
And so we come to language. When I planned the structure of this book I thought that the introduction of this topic would require a major shift of emphasis. But I find instead, that I have already covered most of the issues involved. That in itself, is significant. For my robot, the phenomenon of language is not a radical and peculiarly departure from what went before. It requires only a few relatively simple amendments to existing structures and procedures.

The spoken word
A word is a sound pattern. A word is a set of squiggles on a page. More recently, a word has also come to be a set of binary digits stored in a computer storage device. In all these forms, to humans and robot alike, a word is a sensory experience. And as such, it will be incorporated into the perceptions, as these are processed to form the #concepts of our robot’s understanding.
The spoken word must have preceded all other forms in the ascent of Man, so we will concern ourselves with that form in the first instance.
#Concepts are formed from repeating chunks of experience. If the robot is exposed to an environment in which these word-experiences occur in close association with the patterns of experience which form the #concepts, then they too, will be gathered up and become part of that repeating chunk of experience. Words, the sound pattern of each word, will become just another property or attribute of the #c-object and hence of the #concept formed from that #c-object.
We have noted that the #concept store is organised and indexed for content-related retrieval. It follows that the sound pattern of a word can be used to retrieve its associated #concept. It is particularly useful in that respect, for the sound pattern property of its word-label is (almost) unique to each #concept. There is a degree of ambiguity because some word-sounds have more than one meaning - that is, they refer to more than one #concept. But by and large word-sounds are a more specific indicator than most other properties or attributes which are often shared by a great many #concepts and can be used for retrieval only in groups. A similar phenomenon can be observed by any GOOGLE-user. The use of several keywords is usually desirable to narrow the search for a wanted website. But each site has a unique address, and if that is known it will provide direct access to a specific site. Word-sounds provide a similar access-retrieval facility which is almost as effective.

The written word
The written word cannot be incorporated into a repeating pattern of experience in the same way. It cannot occur simultaneously with the normal range of experiences associated with a given concept. It can however, be associated in that way with the word-sound. So word-sound patterns will be learned first and then the written form of a word will be linked, through that, to the appropriate #concept.

Phonetics
There is a snag. The automatic recognition of the spoken word - particularly of connected speech in all its forms, is a very difficult problem. The pattern of sounds which corresponds to a particular word can be analysed into units called phonemes. These are the fundamental units of word-sound pattern.
That much is known, and there have been many clever schemes devised to analyse and recognise the spoken word. Difficulties are created, however, by the ambiguity of sounds, the way they are slurred together, by the eccentricities of individual pronunciation and by the vagaries of national and regional accents. I have a suspicion that recognition has more to do with the pattern of muscular control necessary to create the sounds, rather than with the sounds themselves. But that is irrelevant. I don’t want to get drawn into a discussion of these things.
I acknowledge the importance of the topic from the point of view of the practical implementation of a robotic system, but it is, as far as I am concerned, a digression. I shall take it, that the problem will be solved somehow, someday, by the people who know about such things. Until that day comes, however, I will allow my robot to conduct his conversations in the written form. We will by-pass the phonetic problem and make an artificial, but straight-forward association, between the written word and its associated #concept. An Outline of the language interpretation system
Each word will be associated with a #concept. (We will deal with ambiguity later.) As each word is encountered, its associated #concept will be accessed and copied from the #concept-store. The copy-version will then be linked into a growing #interpretation structure.
We have seen this before. The system is not simply analogous to the normal interpretation process. It is that process. The only new element, is that some words are associated, not with #concepts which existed before the introduction of language skills, but to a new kind of #concept (called “meta-concepts”) which refer to other #concepts and to the way these other #concepts are normally linked in the #interpretation. For the most part, however, most words are associated with the #concepts we have described earlier in the book.

The Construction Kit Theory
I call this approach “The Construction Kit Theory of Language” (or just the “Kit Theory”). It has been around a while. What has also been around for a while, are some vigorous objections to the idea. I want to show that these objections are invalid and the kit theory is sound.
The kit theory holds that the communication of a sentence from one individual (or robot) to another, is rather like one handing over a modelling construction kit. The kit is designed to enable the receiver to build a interpretation structure similar to the one in the speaker’s head. The one who receives this kit, does not need to be handed the necessary construction parts. They are all already available in the receiver’s head. What is need (and what is communicated) are the names which identify the parts. That’s what words are - labels - labels for #concepts, labels for the necessary building bricks.

Notation
When I refer to a word I will put the expression in quotation marks. The word John is therefore denoted “John”. When I refer to the #concept which is associated with the word “John” I will use uppercase letters and put the expression inside curly brackets, like this - {JOHN}. The curly bracket notation refers to a complex structure and the terms which appear inside the brackets indicate the content of that structure.

Word classifications.
To explain the system more fully I will need to classify the words and their associated #concepts.
The correspondence with traditional word classifications is close but not perfect. Roughly speaking, each noun is associated with an #entity and each verb is associated with a #scenario.

#Meta-concepts
That leaves words like “of”, “where” and “how”, without any kind of classification in my system. Explaining how such words fit into the kit theory is, or was, a major problem for its proponents. In my version of the kit theory, these words are associated with #meta-concepts.
A #meta-concept is a #concept which refers to other #concepts. It defines a pattern of #concepts and the structural links which usually connect them within an #interpretation. The word “on” for example, will be associated with a #concept {ON}. Within this structure, two other #concepts are related by physical location. (eg: “The book is ON the table”)

{ON} =
#entity=E1,
#perception
dimension=height
location= ext(L1,L2)
#entity=E2
#perception
dimension=height
location=ext(L3,L4)
relationship: equal(L1,L4)

I have missed out a good deal of the structure of {ON}. The list of components shown above is restricted to those parts which are relevant. In words, this structure indicates -

---------------------------------------------------------------------------------
There are two #entities (E1 and E2)

A #perception, which is part of E1, refers to its height (channel unspecified). This occupies a location which extends over the interval from location L1 to location L2.

A second #perception, which is part of E2, has a similar height value which extends from location L3 to location L4.

L1 (the base of E1) is identical with L4 (the top of E2)
----------------------------------------------------------------------------------

The interpretation of the sentence, is a process which will identify the objects which correspond with E1 and E2 and in doing so, give them the appropriate relationship with respect to their locations. In this context we could consider the terms "interpretation" and "understanding" to be synonymous.

#signal = [ ID, channel, dimension, location, magnitude ]

#perception = [ ID, inputs, relationships, properties, count, response ]

But the story does not end there. The #meta-concept {ON} will offer alternative ways of making the necessary connection. These connections will not necessarily be concerned with spatial locations. Consider these sentences -

“He did it ON Monday”

To understand this sentence, and build an appropriate representation, the interpretation procedure needs to examine the representations of the #concepts involved. {DID IT} is not a physical object. It is a #scenario and a #scenario does have a physical location. All the #entities participating in the action will have a common location.
The #concept {MONDAY} does not have a physical location and so the initial or default interpretation which is based on physical location would be inappropriate. {MONDAY} is a period of time. The #scenario {DID IT} also has a time-span. The connection is made. The action took place during the period of time specified as “Monday”.

“The patient is ON antibiotics”.

{ANTIBIOTICS} is a substance and the incidence data within that #concept, should indicate its typical use. Inside {ANTIBIOTICS} there will be a short #scenario containing several #entities. There will be one representing a substance, and another representing a person. The #scenario will represent the person ingesting the substance and feeling better as a result. The salient will point at the substance as the true meaning of “antibiotics”. Matching up {ANTIBIOTICS} with {THE PATIENT} is simply a question of linking the patient with the person taking the substance. The structure {THE PATIENT} would also indicate a person who was not feeling well and was receiving treatment designed to make him feel better. These are quite complex #scenarios. I shall leave it at that for the time being. More examples will follow.

“She spoke ON penal reform”

This is a pretty hard one to explain. {SHE SPOKE} is a complex #scenario which involves the utterance of words. It goes something like this ...

{SHE} caused sounds to be made. When applied to these words, the interpretation process created a new anonymous structure which I will denote {....}. This structure has an identity but it has no content. It is a structure representing the #interpretation made by someone who hears what the speaker is saying. It also represents the structure which is in the speaker’s head.
The structure {PENAL REFORM} is also complex, and appears to be totally unrelated. In trying to match these for the purposes of linking to form an #interpretation, the procedure will find that the only dangling reference in {SHE SPOKE} which seems to be in need of amplification, is the anonymous structure {....}. When this is linked to the structure {PENAL REFORM} we see that {PENAL REFORM} has become the subject matter of her address.
This example raises a number of interesting side-issues.
We can use almost any preposition instead of “on” and get a similar result. “She spoke to penal reform”, “She spoke of penal reform”, “She spoke in penal reform”, “She spoke about penal reform”, “She spoke for penal reform”, “she spoke above penal reform”.
The choice of preposition, in some of these examples, may appear somewhat eccentric. Anyone listening to such a sentence, would be troubled by that. But although there would be a lingering doubt, that some import had been missed, the general drift - that the topic of conversation was penal reform, is still fairly easy to grasp.
In each case the words involved in a sentence (and their associated #concepts) bring along enough contextual information to make the correct linkage possible.
The second point of interest is that the interpretation of the sentence requires a direct reference to the interpretation process itself. Quite a lot of utterances are concerned with what is happening inside people’s heads. “I said that ...”, “He thinks that ...” Even the word “but” suggests that something has happened which contradicts an expectation. An expectation is something we find inside a person’s head. Without that ability to represent a {MIND} and to the #interpretation which that {MIND} might contain, we would be unable to tackle the interpretation of utterances of this kind.
Before getting embroiled in that, however, I want to describe the representation structure for a number of less complicated #concepts.

Tartan Hen Publications : Home | more books | Contact : feedback@tartanhen.co.uk