Second
Assumption
Human sentence processing is effected by memory
associations stored in human memory--i.e. by the content and organization of human memory
and all its associations--as memory reacts to the input signal.
The process has nothing to do with fixed sequences
of logical operations.
Thus, human sentence processing will vary from
sentence to sentence as the memory associations vary. For one linguistic
situation, the brain may rely more upon semantic associations, for another, upon syntactic
associations, for still another, upon stored chunk of hybrid linguistic data.
In this latter case, associations involving the
chunks have become so familiar that short cuts to them have sprung up (to satisfy the law
of least effort).
 |
In the Logos System, sentence processing is also effected
solely by the content and organization of associative memory. |
|
In the Logos system, by associative memory is meant the
words and structures of a given language (and their associations) that have been learned
through exposure to that language and assimilated into memory. This assimilation, of
course, is effected by human developers, who work inductively with representative text
samples of the given language. |
|
Rules in the Logos System used to process sentences are
comprised entirely of semantico-syntactic patterns found in actual text. These patterns
capture the semantico-syntactic associations learned regarding the given language, based
on the sampling. |
|
Here too, the parts and types of memory involved in
processing a sentence vary from from one linguistic situation to another, in true
opportunistic fashion, now veering more towards syntax, now towards semantics.
|
|
In sum, in both the brain and the Logos System, associative
memory itself is the processor, as this memory reacts opportunistically to the input
signals. See Mental Model. |
|
Third
Assumption
The brain abhors complexity and seeks wherever
possible to simplify its processes, driven in this by the law of least effort.
The brain achieves simplification largely through
the use of abstraction.
The brain cell (neuron) by its very architecture
would appear to be an abstraction device, with many dendrites for input and a single axon
for output.
The brain clearly employs abstraction in dealing
with the complexity of natural language. No one knows how the brain represents
language internally, but there are hints from language acquisition (click here for
discussion).
|
The Logos System relies upon abstraction as its chief device
for coping with complexity. |
|
Abstraction is the principle upon which the SAL Representation Language of
the Logos System is based. |
|
The semantico-syntactic patterns of the Logos System
knowledge base are patterns at several levels of abstraction and thus enjoy the power of
generality. |
|
In both the brain and the Logos System, the problem of
complexity of natural language is handled by dealing with natural language at the level of
semantico-syntactic abstractions. |
|
Fourth
Assumption
Humans acquire language by means of the general
mechanism for learning, viz. the assimilation of what is unknown to what is already
known.
The language learner assimilates new, unknown
patterns by associating them with already known patterns. This assimilation
occurs when some resemblance (or distinction) between the unknown and the already
known is recognized.
Language learning advances dramatically when the
brain discerns commonality in differing, already known linguistic patterns and
thus, through the power of abstraction, begins to generalize these patterns.
When this recognition/assimilation process is
effected at a sufficiently abstract abstract level, the learning process becomes very
efficient.
This abstraction mechanism, in part at least, might
well explain how virtually infinite language capacity may result from merely finite
language exposure.
|
The assimilation of the unknown to the already
known perfectly describes the Logos approach to sentence analysis. |
|
In effect, an unknown input sentence is
"learned" (decoded) by associating the various parts of the unknown
sentence with already known (stored) semantico-syntactic patterns which,
at some level of abstraction, they resemble. |
|
These already known, stored patterns are the
product of prior, finite exposure to language by Logos System developers who have
generalized what they have seen in language (simulating presumed brain processes). |
|
The capacity to discern analogy among diverse phenomena is
fundamental to this generalization process. |
|
Seeing that x (the unknown) is like y
(the already known) entails seeing them under a common (more general or abstract)
classification. |
|
The SAL Representation Language, being an abstraction
language, supplies the classifications by which an unknown x can
be linked to an already known y in an efficient manner. |
|
The stored, semantico-syntactic (SAL) patterns of the
system's memory, being abstract, encompass virtually all actual and potential sentences of
a given source language. |
|
Thus, after generalizing its language exposure as a finite
set of abstract, semantico-syntactic patterns, the system's memory is prepared to deal
with (decode) an infinite variety of new, previously unseen sentences. |
|
Fifth
Assumption
In the brain, syntactic processing and semantic
processing are integrated rather than separate, sequential processes.
Evidence for this is seen in the fact that we often
understand half-uttered sentences, and can often complete the sentences begun by others.
This further suggests that human sentence
processing must be deterministic, producing a single parse rather than a syntactic parse
forest that must then be semantically pruned.
|
In the Logos System, the SAL Representation Language
integrates semantics and syntax, seeing them as the two extremes of a continuum. |
|
The integration of syntax and semantics in SAL allows
syntactic processing and semantic processing to be conducted simultaneously. |
|
By applying semantics to the syntactic parse, the Logos
System is able to produce a single, deterministic parse rather than a parse forest which
must then be pruned by means of semantics. |
|
Regarding
Translation Proper
So far we have not discussed how translation per
se occurs. We hypothesize, however, that the same faculty of
associative memory applies. In effect, in learning a foreign language, the brain
associates target language patterns found to be equivalent to semantico-syntactic
patterns of the source language. The Logos Model follows this methodology, using
contrastive linguistics (at the semantico-syntactic level) as the basis of its transfer
modules.
(As target language fluency increases, source
language mediation no doubt begins to drop out. A very distant approximation of this
may be said to also occurs in the Logos Model, as the source language sentence is
re-expressed in the SAL metalanguage. Target generation takes place, then, not from
source strings but from SAL strings which loosely approximate a conceptual metalanguage.
This realization is quite imperfect, to be sure, but is indicative of the
orientation of the Logos Model.)
Mental Model as a Basis for the Logos Model
The graphic of the Mental Model,
shown below, illustrates conjectures regarding the role of associative memory in human
sentence processing.
Key Characteristics of the
Mental/Logos Model
1. The brain reacts
opportunistically to language input, using whatever resources it needs to derive
intelligence from the input signals. Resources are drawn into this process by memory
associations. Because associations are different, the same sentence will be
processed differently by different people.
2. To derive intelligence
from the input signal, the brain must first deal with signal ambiguity and
complexity. The prevalence of fan-out and fan-in circuitry in the anatomy of the
brain suggests that analysis (fan-out) and synthesis (fan-in) are fundamental to its
operations. (See graphic below for illustration).
3. The processor for
accomplishing this work is associative memory itself and its organization
(interconnections). There is no circuitry devoted specifically to managing sentence
processing. There is, however, architecture designed for this purpose.
4. Sentence processing is
done in incremental stages along a pipeline architecture, vaguely akin to the incremental
processes of the visual cortex.
5. Syntax and semantics are
not handled in separate operations but are rather different aspects of an integrated
process. Syntax and semantics form a semantico-syntactic continuum.
6. Memory is
content-addressable. This explains why mental processes do not slow down as the knowledge
store increases. It is clear that of the billions of cells in the brain, only those which
should become involved with a given input signal do in fact become involved.
The above are the properties of
the Mental Model which we sought to emulate in the Logos Model.
