Feeds:
Posts
Comments

Archive for September, 2008

The notions of meaning, semantic mapping and relevance are nebulous not because they are fanciful, though that could be argued in many cases. The conceptions or mental representations of these notions–formed in people’s minds– are not entirely clear nor plainly understood.

Many would argue that one cannot know the concepts in people’s mind; particularly those involved in computer programming and computational linguistics.  They consider thoughts to be ephemeral and what people may have in mind relevant but murky at best. Because meaning is the effect of interpreting a sign— something that refers to an object—on the interpreter’s mind, we must attempt to understand these signs. In this attempt, we must clarify not only our language but the interrelations of its terms to our thoughts, our emotions and other personal motivations.

In this post we will take flight, despite the forecast, through the storms of controversy and the clouds of confusion. I will take you to a place where the meaning is sound and discovery is a realization. It is a long and difficult exploration so get comfortable before you begin.

Because people think with words and communicate with language; because we learn by reading and writing, and because laws are written with words, it makes sense to understand what it is about words and language that is connected to thinking. We know it is meaning that connects it altogether.

What every sense-maker wants is to abstract the significant.  In the face of incompleteness we may settle for the explanatory power of the elements and dimensions of that meaning. This so that a course may be predicted with greater confidence and that we may reach a more calculated response.  Because as any sense-maker knows: impressions evoke more perceptive thought and such thoughts provoke action.

Because human behavior is the cause of severe problems in society, it would be a great assistance if computers were able to help clarify to people, to us, the elements and dimensions of thought processes and patterns underlying languages, logics, laws, computing, and all inventions of the mind. If only so we can be more certain we have the same things in mind.

It is like that in the case of the moniker semantic search engine.  What is a semantic search engine? It is obvious that we agree about the search engine part. Agreement about the word semantic is less certain; we do not have the same thing in mind at all.

Computational linguists, generally speaking, do not study meaning in the general sense outlined above. Instead, linguists, under the influence of leaders such as Bloomfield and Chomsky, left the study of meaning and mental representations to psychologists and to others.  The study of meaning, mental representations or semantics, was deemed unscientific.  Chomsky called it “mental gymnastics” and “pseudosemantics”.  Whoever has a linguistics degree now, was trained in that tradition.

In this article I will post about semantic maps and the semantic map employed by Readware technology in particular; mainly because that is the one I know. I will also emphasize where Readware’s approach and semantic mapping departs from the traditional approaches that combine natural language processing (grammar-based NLP) with artificial intelligence and database techniques.

(Full Disclosure – I developed Readware technology along with Dr. Tom Adi.)

A semantic map is related to the schema used in relational database technologies with a data dictionary. It is used to model the data in the system.  In semantic web products like Powerset, and some NLP-based products, a semantic map is used to model the data in the system according to the ontology being utilized. Those that don’t use an ontology specifically, consider graphs of triples (e.g. a,r,b, or a relates to b) to be semantic mappings that map language or some other item into a logical assertion –also called a concept.  Let me depart to define the term ontology.

In AI technology, and on the semantic web, an “ontology” is a set of classes, attributes and relationships (essentially a set of assertions) that are used to model a domain of knowledge in a language that is close to that of formal logic (according to Tom Gruber). In addition to use in mechanical inference, the main purpose of such ontologies is not to study natural phenomena but to integrate “heterogeneous databases, enabling interoperability among disparate systems,” Gruber says.  This is different than the long time use of the term in philosophy where it refers to the study of being (in the world) and where its purpose is the study of natural phenomena.

An ontology of information assets (e.g. the Dublin Core) can help me sort and identify media but it cannot help me understand why some people of some nationalities want to kill everyone of another nationality.  What is the meaning of that? Isn’t that phenomena worth studying so we can avoid that sort of behavior in people where we live?  Talk about a use case for disambiguation — but you won’t find any NLP technology up to that task in this case.  They say it is not possible with current technology but what they mean is: it is not possible with their technology.

So the reason I bring this up is because some researchers, computer scientists, developers and even product managers who talk about a “semantic map” are not mapping words onto psychological or cultural meaning at all.  It is not even about natural phenomena or “meaningful” experience. They are mapping words onto the artificial classes, and or attributes and relationships (assertions) that are used to model a domain of knowledge coded in their computational specification.

According to their announcement, Cognition is mapping words by word forms and word senses (attributes of words) and by synonyms and antonyms (relationships between words) among other mappings. They call that meaning and semantics.  Besides the significant difference in our understanding of meaning this is also where there is a significant departure between the traditional logic and techniques of artificial intelligence and those of Readware technology.

You see.  The objects of NLP-driven search engines –and AI programs and RDF files– are assertions.  Millions of assertions if you heard what the folks at Cognition said.  Each vendor is trying to prove they have or make more assertions than the others.  It is as if they reason that if you can catalog all the possible choices, you can sort out which ones apply at specific instances.

It is not so easy. Because information is generally incomplete and inconclusive, just like it is in the case of language and semantics, one must study the subjects to get to the truth. And, in cases like this, the truth is often hard to recognize and may be hard to pin down.

Every natural phenomena we sense has to be studied to some extent.  It is how we learn. The best way to study any natural phenomena, such as language and meaning, is with the scientific method.  For literate people brought up in the western traditions of the world, thinking is less rigorous and formal but it is patterned on this method, and: it is how we learn.

So let me say how I think and learn.  I think with conjecture, guesses… theories… about the nature of the real things and principles I recognize. This leads me to choices and to assertions. I conclude with what is true or not and what assertions might be made; I do not start with them.  I start, usually, with a conjecture and begin to refine a simple idea or mental representation of the reality of the situation.  Descarte called these initial thoughts innate ideas and Kant referred to them as a priori judgments.

Let me ask you seriously:  Do we want computers to check our logic or do we want them to help us create more useful and concrete theories for solving the problems we face?  Do we want computers to be spectators of the human condition or do we want computers to be useful participants assisting us in thinking about how we can improve the human condition?

I think constriction to assertions and traditional logic limits the help computer programs could provide in clarifying theories, evaluating choices and making predictions.  In the words of the eminent mathematician Dr. Vaughn Pratt:

Traditional logic, like classical mechanics, is a spectator sport: there is an apparatus and a separate observer. Information flows from the apparatus into and around the observer, whose measurements are assumed not to disturb the apparatus. The observer is therefore an information processing system, the essence of which is a graph with nodes A,B,… along whose edges f:A->B (measurement f with source A and result B) information flows. The apparatus itself does not see these edges (but constitutes the sources of some of them) and is not disturbed by the observer. The graph of an idealized observer is a Heyting or even Boolean algebra in the case of nonconstructive logic and a cartesian closed category in the case of constructive logic. Considerations of computational complexity and relevance may call for weaker observers, but not so weak that they disturb the apparatus.

The essence of traditional logic then is an intelligent graph reaching its edges into an unsuspecting structure and contemplating its behavior.

This is useful for static structures and well-known procedures but language and the world itself is made up of dynamically changing structures and interrelated processes. Nothing is really static.  So a key difference between Readware technology and the AI technology used by NLP approaches is the difference between being a spectator and being a participant.  The nature of Readware logic is to order and interrelate the elements of a structure and thereby determine the essence of its controlling processes.

The Readware appartus or semantic map is an information processing device (a regular sign system) and the observer (the mind) is a controller. It receives partially processed (ordered) information from the apparatus and it responds with decisions (and even assertions). The abstract objects of Readware are not assertions and Readware algorithms generate theories not assertions.  There is a big difference between these two notions and the consequences of their common use and deployment.

Notwithstanding the search relevance, and as for the rest of the results NLP products achieve, let me say they show an incredible amount of sophistication.  The parsing and recognition of word forms and word senses is world class in all the systems on the market.  The products that implement the complex parsing and indexing of documents into word forms and senses and entity classes and relations are world class products.  I would love to have a stack comprised of Readware and any one of these language processors.

Because what this crop of vendors do not do (even though their claims imply they do), is pattern experience well enough to induce meaning (in computational memory) from the word forms themselves, as I described above and as literate people do.  Let’s look at an example.

Psychologists who study emotional trauma’s consider language as some of the best evidence available.  If someone says they hurt, they should also know that they are hurt and therefore the self-report of being hurt is valid evidence along with other behavioral and physiological evidence, or the lack thereof. It is important for the care provider to recognize, or create a mental representation, of what it means for the patient to be hurt particularly in the instance where there is a lack of physiological or other behavioral evidence.

In order to interpret self reports, (and testimony, text reports, etc.) one needs to study language. This is best done using natural language semantics. To study the semantics of language one needs universal elements or objects. Linguists, psychologists and others who do study semantics look for such semantic universals—concepts that are cross-lingual and cross-cultural.  Being in psychological pain is not an English or even a Chinese language object, it is a meaningful impression of biological or psychological activity.

Such universal concepts are the major defining characteristic of a “semantic map” for a computer program whose vendor makes the incredulous claim that:

We have taught the computer virtually all the meanings of words and phrases in the English language,

As Cognition chief executive Scott Jarus told AFP.

He could have claimed that Cognition had cross-referenced all word forms and senses known to the English language with the definitions of all the words and phrases in their lexicon.  I believe they may have done that much.  However, if the “meanings” of the words were really taught to the computer, then the computer ought to be able to look-up a word and use its definition and any other entailments to perform a search on the subject and report the result.

The Cognition search engine clearly cannot do this. It can only search on the word or words a user provides.  Pick any word and try it for yourself.  Being able to look up any word and cross reference all its details is a skill that goes a long way.  And reading, interrelating and using those definitions to explain reality or experience is something else entirely; this is where a semantic map becomes necessary.

In psychology, a semantic map is a pattern imposed on reality or experience to assist in explaining it, mediating perception, or guiding response. That is the conception of a semantic map I want the reader to have in mind as we continue.  By understanding a semantic map in this way, one has an intimate way of evaluating the efficacy of a proposed semantic map without buying into the computer products first. Because evaluation of semantic maps is the critical and necessary step before adopting them, let me be redundant.

  1. A semantic map is a pattern imposed on reality or experence.
  2. The purpose of a semantic map is to assist in explaining experience.
  3. The purpose of a semantic map is to mediate perception or guide response.

Those annoucing semantic maps should meet this criteria and explain what part or how much of reality they have successfully mapped. My own participation in the research and devlopment of a semantic map (we called it a semantic matrix in the original work) began in 1982.

Getting through the online noise and storm around the concepts of semantics and relevance-– to the actual elements and dimensions of “meaning” and “human understanding” -–is a long term, often frustrating and sometimes harrowing experience. Now that I am thinking about it, it puts me in mind of spiraling down to the ground through thunderclouds in severe weather.  The problem is that things can get out of hand quickly.

For those that maintain control and make it to the ground, there are ways to understand these concepts and all concepts of the mind. This is simply because in order to be shared and to persist; a) any favorable concept must become less abstract and nebulous and form into a more concrete idea, scheme or plan, and; b) there are abstract and specific and recognizable elements and dimensions to every well-formed plan, idea and understanding.  By favorable, I mean a concept likely to survive for whatever purpose: good or bad.

Ultimately language is intimately involved and plays a great role in all forms of human understanding.  And so, many researchers accept that there is a mapping, between the abstract elements and dimensions of meaning and the signs of language.  Readware technology maps the signs of language onto abstract elements and dimensions of emotional and physical control using a matrix of sound symbols as the semantic map according to Adi’s Semantic Theory (ATS).

Some may ask: why choose the elements and dimensions of emotional and physical control? In truth, we did not choose them.  They were derived from a semantic study.  But in hindsight, one wonders why others  did not already recognize them. Going back to the psychologist interpreting their patient is hurting from psychological pain, caregivers want to know how to control that ‘pain’ so it can be mediated in the constitution of their patients.  To get control of pain means we must rest that control from someone or something else.  Physicians have tactics and best practices for this case.

I cannot think of many things in the world that are not directly interrelated to or affected by some form of emotional, physical or environmental control.  Because emotional and physiological control is a large part of the human condition, a shared and interpersonal semantic space is readily patterned by its elements and dimensions.

The Readware semantic matrix has a small number of elements and dimensions for mapping a large number of interpretations.  This is why I read with some amazement the announcement by Cognition that they have the world’s largest semantic map. It motivated me to write this post.  One reader commented that those of us that disagree with Cognition’s claim should just hold our objections and let them tout their wares.

The problem is that the claim they make gives the reader the wrong impression.  Here is the impression Anthony C. shared with his own words:

The academics can discuss the Olde English and definitive dictionaries that have a set number of words, but I’d prefer an NLP system that understands all the meanings of those dictionary entries. That one sounds like it can build a business by licensing “the bit about them that’s unique.”

Anthony has the (wrong) impressions that the OED is a dictionary of a fixed number of words in the Olde English language and that the NLP system (Cognition) understands all the meanings of it’s dictionary entries. And he reaches a dubious conclusion because of those impressions.

It does not take much to prove that there is no NLP logic capable of interpreting the meaning of simple word forms like feel, fear, hope and love and using those interpretations for locating instantiations in natural language expressions (text). There is only keyword search.  People speaking other languages also have words that refer to the same meanings indicated by the words feel, fear, hope and love becasue these human emotions are experienced irrespective of the language spoken. Keywords don’t work cross-lingually on texts.

I expect most sense-makers will hold, as I do, that there is no possibility of achieving “a more accurate or relevant understanding” without understanding the universal elements and dimensions of the meaning of such signs.

An Introduction to Semantic Mapping with Readware technology.

Some researchers believe semantic universals can be found in simple terms like feel, fear, hope and love that are shared across cultures and languages. Many believe that more complex concepts are not shared by many languages. Along with my colleague Dr. Tom Adi I believe that certain sounds are symbols material objects used to represent something invisible. These sounds are shared by all languages. The symbols represent abstract semantic universals that are used by the mind for symbolic processing tasks such as thinking, reading and writing.  The modern phonetic alphabet represents these symbols thus:

a b c d e f g h i j k l m n o p q r s t u v w x y z

All the sounds of every word of the English language are mapped into the writing system using these symbols.  The Adi Theory of Semantics (ATS) maps these symbols onto 11 dimensions of emotional and physical process control. The Roman alphabet and phonetic symbols are arbitrary, of course, as are the conventions for combining the sounds of the English language and any natural language for that matter. Therefore the symbols of any language can be mapped to the universal elements and dimensions of these elementary, compound and interrelated processes without changing or disturbing them in any way.

Every symbol maps onto one of seven abstract processes—assignment, manifestation, containment, assignment of manifestation, assignment of containment, manifestation of containment, and assignment and manifestation of containment—each with  one of four abstract polarities—closed-self, open-self, closed-others, and open-others  This is visualized in the table I have included below.

You may notice that some cells are empty and some have multiple symbols.  There are reasons for this though listing them here would take us away from the present discussion. You may also notice the abstract objects closed-self and open-self are opposites as are closed-others and open-others. These pairings of polarities can be taken to represent abstract interpersonal engagement conditions (self and others) with abstract interpersonal boundary conditions (closed and open).

So, in Readware technology, intuitions about the words of a language– such as feel, fear, hope and love — are obtained by mapping the phonemes of these words —considered by most linguists to be the smallest elements of meaning– onto these abstract semantic universals. This is done with a simple algorithm used for transforming a word into the abstract objects indicated by the structure of the (word’s-root) phonemes (an abstract word theory) representing the intuitive meanings of the word.

These abstract theories (sign functions) produce impressions that have many possible interpretations or realizations.  We would not want to put a number on this and rather believe that the number of realizations are open-ended.  Such realizations have explanatory power that can be studied outright, Readware technology quantifies them for use in computational algorithms.  Now let’s get into the practical implementation so we see how it works.

Most words are ambiguous because sounds produce ambiguity (multiple meanings) when combined in a word root. This is at least partly because each phoneme symbolizes compound abstract objects that convey different aspects and characteristics of the natural phenomenon referenced by the word.  Polysemy is a linguistic term that means that a word root may refer to different objects in different contexts.  All phonemes are polysemic.  All words appear to be polysemic too.

According to ATS, every word can be transformed into one of several forms of quantifiable functions defined over the discrete domains and ranges (of control) dimensioned by these abstract semantic universals. Thus, even emotional word roots that are ambiguous can be included in the evidence studied to understand their nature.

Adi’s theory of semantics has rules to convert any word root into an abstract mathematical mapping such as f: X->Y or f(X), etc..  Consonants that refer to processes with higher precedence play the role of the function f of the mapping f: X->Y and the remaining consonants represent the domain X or range Y of the mapping.  The mapping f(X) is a mapping with an unspecified range.  The words feel, fear, hope and love are mapped by this formula, implying that the context for these word can range across anything at all.

For some, the formulas and the abstract mappings of semantic universals, may be too abstract to be of much practical use, yet each abstract mapping can be interpreted into more concrete terms suitable not only as a definition, but also as a knowledge representation with extraordinary explanatory power.  In other words these abstract impressions induce more concrete interpretations.  Such interpretations can be corroborated with personal experience

Consider, for example, that the sound /fe/ symbolized by the letter “f” represents the abstract semantic universals open-self and manifestation. Remember that open-self is a polarity (think charge, inclination, valence) and that manifestation is a process.  Action and activity are effects of processes of manifestation, i.e., one manifests behavior and actions, and activity is manifested.  So this is how a phoneme from a language is mapped to a compound abstract object that evokes multiple interpretations –meanings– from the sound-symbol.

Both words, feel and fear are used to refer to sorts of emotional activity. That is not a definition from a dictionary though it is defining.  In the word fear, the emotional activity applies to the domain of self –is assigned inward– (by the polarity) indicated with the consonant “r” in the formula.  In the word feel, the “l” has outward polarity and assigns the manifestation outward– the domain is open to the outside.  These abstractions give us some impressions of what it instinctively means to fear or feel–yet they are even more than that. They are theories. We can use them to generate more concrete theories about what the words fear and feel mean.

For the “f” in feel, these abstract universals can be realized simply as “opening oneself to outside manifestations”. For the sound “f” in fear, the universals induce the more concrete realization: vulnerable state. Vulnerable is a realization of the open-self. A state is a realization of a particular manifestation.  Of course, any realization is conjunct to the situation, circumstances or context of use.

An event is also an interpretation of manifestation and the open-self is the universal negative, so a concrete realization of fear is a negative event. The open-self polarity is also realized as unfamiliar emotion and both the words feel and fear are used to communicate a sense of unfamiliar emotional activity. The unfamiliar can induce fear and result in feeliings of anxiety and agitation.  And a feeling is often initially unfamiliar enough to get our attention.  Do you feel me?

The explanatory power of the semantic universals of this mapping enable us to make predictions such as:

–the advent of uncertain or unfamiliar circumstances can evoke fear in the minds of people.

This prediction can be applied to events concurrent with the uncertainty of the political future in America.  Given a collection of American news articles covering this year (2008) from January until September, Readware algorithms can identify the instances that evoke fear and contribute to increasing uncertainty, anxiety and agitation of popular opinion from articles that do not.  This can be done with a query of the form: fear, because.  The entire process for this case, would take less than a few hours, including installing software, parsing and indexing documents and achieving the results. It would cost pennies per article given a few million artilces exist in that range.

By finding instances, textual evidence from passages in press reports, Readware technology can be used to inform political strategists, for example– to locate and track relevant issues that produce fear in the populace and deserve focused attention. Such information can be used to great advantage, to damage an opponent, or not at all.

So, If you are really thinking about using semantic technology, you should know about the limitations of products based on traditional AI-style logic and mechanical inference.  Learn to recognize the difference between what they claim to do and what they actually do.  And be aware of the availability of alternative methods that exploit the explanatory power of text.

Read Full Post »