There’s something more than a little creepy about writing. Not the act, but the letters themselves. Habit, of course, dulls most of us to the actual shapes of Roman letters, but it’s strange nonetheless that someone learned how to fix the noises of speech onto a piece of paper or stone in such a way that they outlast the person who thought of them and find their way into books that will travel to places he’s never even imagined. Writing quite literally is re-presentation: the symbols of letters, ideograms, and hieroglyphics create, as if by magic, the voice of the person who wrote them.

Now, it is admittedly a fairly cheap pedantic trick to imply that two meanings for a word are actually the same, but since this essay is, in fact, actually about multiple meanings for words, I’ll offer no apology for the suggestion that representation in politics is basically the same problem as representation with words, and that there is much to be learned in politics from a young boy trying to capture the color of his first love’s hair in a sonnet.

Considering how often we use words to get what we want, it’s not surprising how many times philosophers have asserted that words are fundamentally political organisms, whose meanings are manipulated to control debate or even thought. This is a rather well-beaten horse, and I’d like to leave Foucault alone for the moment. The connection I want to establish is instead at the level of the process of representation. Because our brains are so good at reading, writing, and speaking, we really didn’t understand how complicated representation is, until we invented computers which are, alas, about as thick as politicians.

There is a branch of mathematics–information theory–that describes representation. It is perhaps more useful for its metaphors than for its measurements. The basic idea is that any message has both a sender and a receiver. The sender operates some mysterious mechanism, which could be a pen, a computer keyboard, a card punch, or a touch-screen voting machine, and the receiver, on the other end of the black box, does something in response. The relationship between the symbols used by the sender and the receiver’s action is the codebook. Codebooks can be simple or complex. A simplistic, and somewhat troublesome, example is the big red button that launches nuclear missiles. If the operator enters the right code, we’re all screwed, whereas all the wrong codes put things off for a while. On the other end of the scale is human language, a far more complicated, subtle thing. Words often have meanings that depend on the context (for instance, rose by itself could be a flower or the past tense of a verb), and in other cases different words may have the same general meaning, but with subtle differences that can be used to set tone or make fine distinctions.

The way in which we discern intent from speech or a written document is still not entirely understood, but it depends on years of hearing and using the language, a process by which we become aware of the vast web of associations and definitions that comprise a language. Computers don’t have access to this immense codebook, and so the symbols that we use are utterly opaque to them, merely an arrangement of characters.

One of the most ubiquitous examples of this opacity is the difficulty computers have in searching through text. It’s easy, of course, to find a particular word in a document; it’s almost impossible to find a document dealing with a specific topic when it’s buried somewhere on your hard drive or out amongst the trillions of documents on the Internet. Everyone who’s used a computer for research has played the search engine game, in which the object is to come up with the right word or combination of words that will actually pull up something useful. What this amounts to is a strange semantic exercise in which the searcher has to guess what sort of words would have been used by the author of the document. Search programs don’t know what words mean; their only forte is in being able to rapidly sift through mountains of data for the particular set of letters they’re given.

The alternative to using the searcher’s semantic skills is to use the author’s. Card catalogs work on this principle: references are filed under various subjects, allowing the researcher to pull up a much smaller and more appropriate selection of documents than if he had searched on the basis of text alone. This solution, however, has the inherent flaw that the indexer must anticipate the possible uses of the document. Furthermore, as search engines become useful for commercial activities, resourceful entrepreneurs have a large incentive to produce incorrect indexing data that will misdirect searchers to their site. The Google search engine, which uses links from other documents to index and rank their search results, is subverted in this way all the time.

Clearly, if computers could index documents themselves in a meaningful way, or infer the meaning of a researcher’s query, search engines would be far more effective. But human language, even on the level of single words, is terribly complex. As we’ve noted earlier, a word can mean all sorts of things. Only in the narrow and artificial world of scientific and technical literature are single terms both unambiguous and complete. Gravity, to a physicist, is exactly one thing, and no other term can be used to identify it. But in ordinary language, gravity signifies the force holding things to the earth as well as solemnity and dignity of manner. Both concepts derive from a single source, and the poets have made much hay from these sorts of deep semantic connections, but for a search algorithm, widely variant denotations can be troublesome. Likewise, words in ordinary language are not complete. There are probably 200 synonyms for car, all of which designate some kind of transportation entity. Humans tend to appropriate words for new concepts, and then re-appropriate them for all kinds of reasons. Nowhere is this more egregious than with political terms. There are several different meanings for “fascist”, a word that was coined less than a century ago, and the history of “liberal” involves so many reverses it could easily be a feature on “Behind the Music”. Lacking the cultural history of words, the desire to communicate, and the mechanisms for learning language, computer programs can only see human language as a black box.

Oddly enough, humans are in precisely the same position as computers when it comes to understanding their own brains. Neurons communicate with each other by means of complex trains of “spikes”, or action potentials. Individual spikes are believed to carry some information—for instance, the number of spikes may indicate how closely some visual stimulus resembles a particular object, so that one neuron may produce a lot of spikes when you’re looking at a face, but none at all when you’re looking at something else. This kind of “rate code”, however, rarely tells the whole story. Some neurons appear to have a complex code book, in which certain patterns of activity communicate a specific meaning that cannot be expressed by the number of spikes in the pattern. Figuring out the neural codebook is a classic chicken and egg problem, because in order to deduce the codes, you have to know the meaning of the messages, and in order to know the meaning of the messages you need to know the codes.

Recently, a mathematical technique called Principal Component Analysis has shown some promise for helping computers to understand humans–at least enough to be useful in searches–and humans to understand brains. In principle it’s quite simple, although the actual implementation demands a lot of computing power. I’ll use the example of search engines to explain how it works. The first step is to count up the number of times a symbol appears; in the case of documents, the symbols are words. This allows the document to be represented by a point in “word space”. Just as you can describe a location on the planet with two numbers, latitude and longitude, a document’s location in word space is determined by the number of times each word appears in the document. Word space has a lot more dimensions than ordinary space, but the mathematics are basically the same. The advantage of representing documents this way is that the computer algorithm can spot clusters. For instance, recipes often contains a lot of the same words (cup, teaspoon, baste, stir, etc.), and in word space the points for recipes would be closer to one another than to those for other types of documents. These clusters define a semantic class–a group of words that together communicate a concept. In my example, that class is recipe, and though a computer wouldn’t have the acumen to actually call it that, it would be able to group recipes on the basis of their contents. Or, with the example of gravity, clustering would be able to distinguish the technical use of the term from the poetic use, because when gravity is used technically it tends to be surrounded by other technical terms.

The disadvantage to this kind of representation of documents is that even with a dictionary of modest size, the number of dimensions is in the thousands. Computers are just now becoming powerful enough to deal with this quantity of data, but an additional problem is created by the sparseness of the data—which is to say that most words do not occur in a given document, making accurate classification much harder.

This is where principal component analysis (or PCA) comes in. Its effect is to distinguish dimensions that are especially relevant from those that carry little or no information. These relevant dimensions would actually be composed of multiple words. The recipe dimension would consist of the words that tend to appear in recipes and not in non-recipes. Interestingly, the dimensions PCA identifies depend on the set of documents used for the analysis, which allows it to work around situations where there are a lot of synonyms for a single concept. In general literature a single dimension might adequately describe most of the synonyms for cars, whereas if the analysis were performed on used car advertisements the distinct meanings of the synonyms would emerge, allowing the software to distinguish between sedans and trucks, for example. PCA only distinguishes multiple concepts when there is an operational difference between them. The meaning of the concept is latent in the text, which gives rise to the name for the technique when used in this context, Latent Semantic Indexing.

The application to politics lies in the fact that political positions are also defined by a set of ideas. Fascism, for example, is not simply authoritarian, but also involves aggressive militarism, a religion of the State, and state control of the economy through the guise of industrial, expansionist capitalism. This is perhaps a poor example if only because we are concerned here with improving representative government, which is one of the things that generally has to be discarded in order for fascists to maintain power. Nonetheless, in democracies voters would like their viewpoints to be represented in the government’s laws and policies, and the primary mechanism they have for communicating those preferences is through the ballot box. Those of us in America should be well aware that this information channel is less than reliable. We have, in practice at least, two choices in most elections, and the signals we send our leaders only rarely produce the desired results (the question of whether this system is actually designed for true representation or merely for gaining and maintaining power is left as an exercise for the reader). Voters are forced to discard some values as irrelevant, either because both parties have the same position on that question, or because the party with which they would normally side has the opposite position. The Lutheran Farmer’s Party, sadly, will never exist in America, because their pro-life, pro-conservation, small-government, pro-social welfare positions will never garner enough support to elect anybody to national office (and we must admit they would probably split into Norwegian pro-lutefisk and German pro-beer factions, further complicating the problem).

There have been efforts to more accurately represent political affiliation. Instead of a simple Left-Right axis, schemas like the Political Compass distinguish between a person’s position on economics (socialism vs. laissez-faire) and his position on social issues (authoritarianism vs. libertarianism). In practice, parliamentary democracies tend to support a larger number of parties, which allows voters to choose politicians more representative of their own positions. But the best method of analysis may be to avoid the notion of parties altogether. It’s common in political psychology studies to determine indices that assess a person’s views on a particular topic, and thus an individual’s political position could be represented by the values of those indices—in effect, a point in “political space”. Latent semantic indexing could process the thousands of political indices into “relevant dimensions”, groups of voters who tend to feel the same way about many different issues, and parties could spring up on an ad hoc basis to address a single group’s position.

Until it becomes possible for small parties to elect their own representatives to power, political indexing will only be useful as a tool for analyzing the political landscape. The existing parties already use a crude form of indexing, in the form of polls. Candidates regularly use polls to identify groups of voters who are more and less attached to either party, and they direct their speeches and advertising dollars to these voters, hoping to convince them that the candidate more accurately represents those voters’ values.

More accurate identification of voting blocs, unfortunately, has the potential for abuse. In an age where more and more of the daily transactions and predilections of ordinary citizens are recorded and stored in databases, the amount of data available for this kind of indexing continues to grow. When your political values can be predicted by the type of potato chips you prefer, there is a great deal of room for manipulation. Voters whose political positions make them sure bets for one party or the other will be increasingly marginalized, while the swing voters will be subjected to ever more carefully targeted propaganda. On the other hand, if voters become more aware of their own values and those who share them, they can combine to ensure that their common interest is not neglected. The tools are out there for politicians and grassroots activists alike, and as they become more sophisticated in using them, politics will continue to transform itself from a conflict of ideas into a war of information.

Whether this is a good thing depends, ultimately, on what you believe government is meant to do for people. If governments can best preserve liberty and justice by satisfying the political desires of as many people as possible, then increased information and better indexing will only make that task easier. But if liberty and justice depend on something more than making government do what you want—if there really is a common good—then sophisticated electioneering can only be inimical to the process of discovering what is good for everyone.