The Practicality of Studying Language Origins

The study of language origins might offer the way to determine the best explanation for why language is the way it is, said Gary Marcus in a presentation that led off the Evolang conference in Barcelona this week (March 13-15, 2008). Marcus cited Noam Chomsky’s remarks (in rather more baroque language) that while a good theory covers the data, a great theory explains why the data have the character they display, and Marcus urged that the path to a great linguistic theory goes through understanding language’s evolution.

Marcus is a linguist who studied under Steven Pinker so the data he refers to is syntactic. He noted that there has never been a lack of decent theories to describe the syntactic data, but there are still no clear explanations of why we have the syntax we do. (Note: I’m using the term ‘decent’ because there are no theories yet that cover all the data.)

Hovering behind Marcus’s presentation was the distinction between linguistic performance and competence. They provide two separate sources of data:

Linguistic performance: this data comes from observing what people actually say.

Linguistic competence: this data comes from observing what people know they should say.

Chomskyan approaches tend to focus on the competence side of the issue, asking what logical powers are required to generate syntactically correct sentences. As Marcus commented, there are a variety of decent theories of linguistic competence. To choose among them, Marcus proposes looking at the biology of linguistic performance. We don’t always say what we should say, and we don’t always understand what we should understand. Why not?

Researchers who focus on linguistic competence alone tend to say that language is very well designed, nay, near perfectly designed. When pressed to explain such optimization, the researchers tend to fall back on natural selection over many generations. Marcus quoted Tooby and Cosmides (1995)

… natural selection tends to cause the accumulation of superlatively well engineered functional design.

But Marcus pointed out two familiar examples of inferior design—the rotated human spine which leads to many backaches and the “backwards” retina—and he provided a nice quote from Stephen Jay Gould:

Remnants of the past that don’t make sense in present terms—the useless, the odd, the peculiar, the incongruous—are the signs of history.

The bulk of Marcus’s presentation focused on “peculiarities” of linguistic performance that are so entrenched that they suggest a less than optimal competence. Marcus did not address the idea that perhaps these peculiarities indicate that syntacticians have gotten hold of the wrong end of the stick and are mistaken about what should be optimized. Instead, he accepted the assumptions of linguistic competence and showed how human brains do not support that competence.

The central point is the well established fact that human memory does not use the same system used by computers. Memory in computers is “location addressable,” meaning that a program tells the processor to get the data located in memory box X and use it. Human memory, in truth vertebrate memory in general, does not work that way. Marcus listed a number of ways that linguistic performance shows the absence of that kind of memory:

• Speech reflects frequency of use; e.g., speakers access common words immediately but pause before using a more unusual word. This “search for le mot juste” is so familiar that we see the pause as a sign of a careful speaker. But location-addressable memory shows no such tendency.
• People are terrible at repeating word for word what somebody said, but they are good at getting its gist. Machines with location-addressable memories show just the opposite results.
• Many psychological experiments have established the ways in which recent experience interferes with current thinking. No such phenomenon is found in computers with addressable memory.
• Old data interferes with new. For example, we may refer to a woman friend by her old, married name even though we know she has divorced and reverted to her maiden name. With addressable memory, you just replace the name in the memory box and that problem is vanquished.
• People get confused by their own complex sentences. For example, a person may say The keys to the cabinet is… instead of are because while keys are the cabinet is. Computers seem to have an easier time with syntactical agreement than people do.

These limitations and interferences directly challenge the notion that linguistic competence takes the form of using syntactic “trees.” (A syntactic tree is a set of formal rules that are used to generate or interpret a sentence.) One limitation of interpretation can be seen with the sentence The coach smiled at the player tossed a Frisbee. No editor, of course, would ever allow such a sentence to be used in print because it is plainly confusing. A computer using a tree will quickly discover that the phrase “the player tossed a Frisbee” makes no sense in this context and so parse correctly that the coach smiled [and] tossed a Frisbee. Human readers will also recognize the incoherence of the business about the player tossing a Frisbee, but humans typically have a very hard time figuring out what the sentence does mean.

This point gets to the heart of Marcus’s argument. Because our brains do not work like computers with addressable memory, they cannot always handle languages that seem logical and efficient. It is well known that machines often have great trouble interpreting “natural language;” Marcus pointed out that the reverse is also true. Programs with many embedded elements can be hard to follow. Back in the 1950s mathematicians tried to produce an unambiguous language (LogLang) but nobody could learn it.

These problems suggest that the variation between performance data and competence data may not merely reflect the simple frictions of the moment. Something deeper may get in the way of competent speech and interpretation. Some the syntactic theories may assume a competence that is simply not available to humans. A theory that explains our true competence will have to take into account the process that gave us the brains we actually did evolve.

Comments

[JanetK]

I think a good rule of thumb is that (unless there is overwhelming evidence to the contrary) all processing in the brain is parallel and not sequential. The idea of parsing sentences in a sequential manner does not seem natural or very probable to me.