Martin Bryan, IS-Thought
Far too many of the world's problems are caused by the fact that we insist, like Humpty Dumpty, on applying our own meaning to words, rather than the meaning assigned to them by their originators. To counter this tendency we are told by ontologists, that peculiar breed of metaphysicists devoted to recording the "essence of things", that we should define concepts and not worry about the terms used to identify them. It seems that ontologists work on the principle that once everyone has agreed on the concept being described it no longer matters what term you use to represent it. The term the person defining the concept assigns to it will, it seems to be suggested, always be the best one for everyone else to use to refer to the topic. Humpty Dumpty would, of course, approve of this approach.
Language, unfortunately, just doesn't work like this. Words do not have a single meaning. They represent multiple concepts, and we have to use the context of the messages they are contained in to determine which concept to assign to each word. Novelists and politicians deliberately use the interplay between the different meanings of a word to introduce ambiguity into their messages. Humour is often dependent on the use of a word in an abnormal context. The more meanings a word has, the more subtle are the concepts it can be used to express.
It is often the shortest and most commonly used words that have the most meanings. In English the word "run" is probably the best example, with over 40 distinct meanings as both a verb and a noun. I am running a computer program to record the words that are running around in my head. Perhaps I should be out for a run around the block, or spend my time running a business. I could run for parliament in the next elections, run my enemies to ground, run my friends around the bend, or simply run into a brick wall trying to explain why one meaning is not going to suffice for the word run. At a more allegorical level the running water in the stream running past my house could be running red with blood from the running sore of pollution that blights modern society. I could run on and on, but you would soon run out of patience with my attempts to run through the alternatives. Suffice it to say, you get my meanings.
How is it that a word can have so many meanings, and yet not confuse a native speaker of a language? What is it that ties together all of these uses of the word so that we can apply a common interpretation of their use? There is a common thread that ties together the many different interpretations of the word, but it is not one that is recognized in dictionaries. Running is normally used in contexts in which there is a start time or point, and an end time or point. The underlying concept is that of "heading towards a defined goal". Most of its meanings are based on our ability to identify this analogy. It is the comparison between the motion of a runner, or the motion of a stream, that allows us to understand that anything that proceeds in a series of steps, or provides a constant flow, can be said to be running.
Analogy is the key to language. We cannot define the different concepts for everything we do, encounter or wish to communicate to others. It is only by generalizing what we encounter, by noting the shared properties of objects and events, that we can reduce the set of concepts in our minds to manageable proportions. When we describe a new "thing" we seek to associate it with a situation that those we expect to read our description should have previously encountered. For example, actors are said to "run through a play" when they rehearse it without fully acting the parts. While there is no physical movement around a stage, there is a logical progression through the steps outlined in the script which allows the play to be speeded up rather than played at its normal rate. It is this ability to identify the similarities between situations that allows humans to communicate new concepts without having to invent new words to explain them.
Rather than trying to formally define the meanings of the situations in which words can be applied we should be trying to identify the analogies that have been associated with a word. We could, perhaps, define "run" as "moving quickly from situation A to situation B, as in a race or in flowing streams". Sometimes, however, words are used in direct opposition to their analogy. Examples of this include phrases like "running into opposition", "running into a brick wall" and "running out of steam", all of which imply being unable to move quickly from one situation to another. Even allowing for the negation of analogies, our definition would still be insufficient for applications such as "running a temperature" and "run of the mill", which are based on the subsets of the properties of run (i.e. the temperature having moved from its normal position, or a mill's constant motion without change of the situation"). Our definition would, however, cover a wider range of terms than are covered by many of the current definitions for run found in English dictionaries.
Many misunderstandings occur between people brought up in different cultures because their languages do not share the same analogies. While an American can "run for office" a Frenchman has to be "carried as a candidate" (se porter candidate). Both analogies are apt, reflecting the different traditions of the countries concerned. In fact, there are so many different ways in which French translates the word run that this entry is invariably one of the longest in any French/English dictionary.
We must take care not to unconsciously suppress cultural differences by restricting the set of analogies that we allow to be applied to a particular concept. We need to encourage different viewpoints. Poets have traditionally sought to introduce new analogies, as have experimental novelists such as Joyce. Scientists rely on analogies to get advanced ideas across to the public, while politicians all too often use them to confuse the public. Words should retain as many different alternative meanings as possible, rather than being put into straight jackets, as is often proposed by those who want to simplify a language so that they can simplify the processes of translating from one language to another. "One word, one meaning" is a slogan that must be vigorously opposed by those who want to preserve culture.
The developers of the Internet are now trying to put some order into it by defining a set of semantics that will form a "Semantic Web". The technique involved requires that a particular string, or a reference to a web resource, be shared by all assertions made with respect to it. I can assert that John Smith is bald. You can assert that John Smith has red hair. Someone else can assign the string John Smith as the name of a brand of beer. The fact that the same string is used in each case is taken, according to the rules proposed for the Semantic Web, to imply a relationship between the assertions, which is patently not the case. Applying such rules to something that is so anarchic as the set of information available over the Internet is nonsense. What we would need to be able to do first is to distinguish all the unique applications of a particular term. This would require that we assign some form of unique identifier to each application of a term which is fundamentally different from that of previous applications of the term. This is in direct contradiction to the way in which we use our languages.
The basic concept behind the Semantic Web is to help computers to interpret information available over the Internet so that they can help humans to come to grips with the vast range of information that is available to us today. What has been forgotten, however, is that written language relies on the fact that one application of a term inherits properties from other applications of the term. Computers need to be able to identify which properties have been inherited in this way if they are to be able to help us find our way through the web of data now available to us. How can we make these inherited properties accessible to computers? Here is where ontologies come in. Ontologies are supposed to define the set of properties associated with a concept in such a way that computers can understand and use to "reason" using "rules of logic".
Unfortunately logic and analogy do not fit together very well. The set of properties associated with running a race is different from that required for running a computer program. Unless we clearly identify the domain in which we are applying a term we cannot expect a computer to select the most relevant set of properties to reason about. But how can we indicate the relevant domain? How can a computer determine that the domain of one Internet resource is about athletics while that of another is about computer programming? Unless we have some way of identifying the relevant domain to the computer that can be enforced as part of the data generation process, how can computers correctly interpret the information we have made available?
How do we identify a domain of discourse? Can we provide computers with a way of identifying our domain of discourse? One way might be to require that all Internet resources to be associated with an ontology that assigned a single meaning to each term within it. But would this be useful, particularly for resources that contain, for example, allegorical poetry or political speeches? Another possible technique might be to identify the set of domains used to distinguish between multiple uses of terms in a dictionary and require users to identify one of these domains as being relevant to the resource. But, again, this would not work for resources that covered more than one domain, or used terms in deliberately ambiguous ways.
If you look at how you have interpreted the word run within this paper you will see that one paper can cover many different domains, and assign many different meanings to a word. How were you able to differentiate between these meanings? In most cases you did it by associating a domain with the noun that was the subject of the phrase or sentence. Each such noun sets a context in which the actions described by a verb can occur, and also determines the context in which nouns that form the object of a phrase or sentence can be interpreted.
In general we cannot assign meanings to the words that make up a sentence until we know the subject that they are concerned with. Defining ontologies with predefined meanings for words simply will not help computers to interpret general purpose texts. Ontologies may help some closed domains with limited vocabularies, but to try to develop general purpose ontologies is going to be a waste of time. The Semantic Web will only work if it adopts the same techniques as humans to identify the meaning of words. This means that the computer must be able to associate meaning with the subject of individual phrases or sentences, rather than trying to identify a set of subjects that describe a whole document.
How can we adapt the way in which people assign meaning to words to the need to better identify what Internet resources are describing? Firstly we need to throw away the concept that a string identifies anything unique. At the same time we need to discard the idea that a resource identifier is unique, as one resource can be interpreted in different ways by different people. What is necessary is that we identify the domain assigned to those words in the document that are key to its interpretation. [For the time being I'll skip over how to do this complicated part of the task!] We will need a set of uniquely identified, domain specific, "subject identifiers" that clearly map to a particular definition of a concept. For example, urn:terms/computers/run and urn:terms/athletics/run could be used indicate that the word run has a different meaning within the domain of computing than that assigned within the world of athletics. Alternatively we can generalize statements by associating particular resources to specific domains, using references such as urn:domain/computers or urn:domain/athletics. By associating unique subject identifiers with specific resources, either internally, as metadata within the file, or externally, using assignment statements of the type defined for the Semantic Web's Resource Description Format (RDF), we can ensure that resources associated with a specific domain or use of a term can be quickly and accurately identified.
The question then arises as to how will it be possible for a computer to identify the correct set of terms to associate with a particular document. This will require a program that can a) identify the subject of each phrase in a sentence b) identify the actions taken with respect to each subject c) identify the most commonly used subjects and actions and d) identify the set of subject/action pairs that best distinguishes this document from other documents. Note particularly the last of these requirements. While a) and b) are both attainable, for languages such as English, using modern linguistic analysis programs, and c) can be achieved by simple statistical analysis, d) cannot be done using the obvious process of taking the most commonly used pairs. It requires other forms of judgement to be applied, including whether the pair appears at the start or end of a chapter, section, paragraph of sentence, or only occurs in the middle. It also requires knowledge of which pairs occur most commonly within a corpus (the more commonly they occur the less use they are in distinguishing between resources) and which pairs best distinguish the application of a subject within a particular domain. Obtaining the necessary correlations will take time, but are a necessary preliminary to the overall goal of trying to get computers to correctly analyze the significance of our inputs to our readers.
November 2002