Sunday, February 7, 2010

Dolphin-Bikes and The Iconicity Effect

Since the journal Cognition typically allows free online access to its current volume, I was able to read a recent paper on a topic that I've always found interesting: the role of embodied experience in language processing. The basic question is, how does our size and shape and orientation as human beings affect our language? Think about a creature that's physically very different from us, like jelly fish or bacteria or dolphins. Now imagine those creatures magically had the same cognitive capacity that we do. 

Would our language system work for them or would it necessarily have to be different? 
As an analogy, think of a simple bicycle. Bicycles are designed for human bodies. Now think of dolphins. Dolphins are smart and can be trained to do many things, but could you train a dolphin to ride a bicycle? No. Because a dolphin body would not fit a bicycle properly. You would have to design another mechanism around the dolphin body. In all likelihood, the dolphin-bike would not be ridable by a human because our bodies just wouldn't fit properly. And we could point to the features that were not right.

Now, think of language as a mechanism that was designed by evolution to fit humans. Can we discover what parts of the language mechanism require human-like experiences (e.g., standing upright, two-eyes in front of our head, etc). Yes, we can. For example, researchers have already shown what's called an iconicity effect for semantic similarity recognition. When participants are shown the word ceiling visually presented above the word basement, they are faster to recognize that the words are semantically related than when the same words are shown in the opposite presentation. Note that a creature like bacteria might perceive the relationship between ceiling/basement very differently than we humans do.

In The linguistic and embodied nature of conceptual processing, Louwerse & Jeuniaux (full citation below) performed four related experiments to test what exactly was driving these kinds of results. In particular, they wanted to know "to what extent, and under what conditions, both embodied and linguistic factors are used in conceptual processing."

From their abstract
The embodiment factor predicted error rates and response time better for pictures, whereas the linguistic factor predicted error rates and response time better for words. These findings were modified by task, with the embodiment factor being strongest in iconicity judgments for pictures and the linguistic factor being strongest in semantic judgments for words. Both factors predicted error rates and response time for both semantic and iconicity judgments. These findings support the view that conceptual processing is both linguistic and embodied, with a bias for the embodiment or the linguistic factor depending on the nature of the task and the stimuli.

This was well done research and I liked the paper, but I feel the need to nit-pick (it's what I'm best at). 
  • Interesting use of LSA to determine semantic similarity (LSA is a sophisticated statistical analysis of corpora). But I wonder if this is the right measure. In fact, they discuss a 15% error rate (where participants decided two particular words are not similar when LSA says they are) and they didn't feel this was an issue. But in the bigger picture, why use LSA at all? These are small data sets. You could just run a bunch of word pairs through a paper-pencil judgment task and use those results. That would match the experiment's task more closely. Don't get me wrong, I Like LSA. I'm impressed with its value as a tool for linguistic research. But it's not right for everything.
  • I didn't see the frequency of word pairs as a specifically linguistic factor. This is a quip regarding terminology, not methodology. I think it was a smart metric to utilize. There certainly are plenty of frequency effects in language, but to call those effects linguistic factors seems misleading. 
  • They only used one iconic dimension (over/under) to create all of their iconic stimuli (e.g., monitor/keyboard, boot/heel, steeple/church). It may be the case that it's just easier to find example pairs for this dimension (which would be an interesting fact unto itself). But it occurred to me that a mix of dimensions would be good. 
NOTE: In my first draft of this post I used dogs and doggie bikes in my analogy, then got wise and checked with the youtubes; sure enough, some Japanese guy taught a dog to ride a bike. Cute video here.

Full Citation
Louwerse, M., & Jeuniaux, P. (2010). The linguistic and embodied nature of conceptual processing Cognition, 114 (1), 96-104 DOI: 10.1016/j.cognition.2009.09.002

1 comment: said...

Many institutions limit access to their online information. Making this information available will be an asset to all.

Putting the Linguistics into Kaggle Competitions

In the spirit of Dr. Emily Bender’s NAACL blog post Putting the Linguistics in Computational Linguistics , I want to apply some of her thou...