I read a language related article today where the researchers presented their subjects sentences with the words making up the sentences being presented serially on a screen. Their goal was to determine if brain responses correlated with the presentation of words in a sentence carrying more information according to several probabilistic models of language. The material that they presented to subjects seemed basically like a slowed down version of this:

The authors did find the effect they expected – words bearing more information lead to a certain kind of increased brain activity, but they also made some more far reaching suggestions based on the probabilistic models they used.

Two of the probabilistic models the researchers employed didn’t make use of hierarchical structure. The predictions made by these models correlated better with the brain responses they found. Based on this finding, they state in the abstract: “These findings suggest that different information measures quantify cognitively different processes and that readers do not make use of a sentence’s hierarchical structure for generating expectations about the upcoming word.”

I understand that the researchers needed to present words serially in order to isolate the variable they were considering, namely higher information bearing words. But, how natural of a task is this really to make assumptions about something beyond the word like the role of hierarchical structure in language? In reading, our eyes don’t typically take in just one word at a time. At least, I know I don’t read like that.

For me, the brain responses in and of themselves are interesting. For instance, some corpus approaches to linguistics look at co-occurrence to see which words occur more often together than they should based on chance alone. This is referred to by some as mutual information. But, what is really interesting in these brain responses is that the words in the sentence that don’t normally cluster with the others receive the most significant brain responses. In terms of information theory, these words carry more information in the sentence because they are less predictable. So, perhaps it’s not just correlations, but also negative correlations that we need to pay attention to in probabilistic approaches to language when we are working at the level of the sentence. As interesting as that is, I’m not sure it’s necessary to reach farther than that to talk about hierarchical structure when it seems like the task the subjects performed was pretty unnatural.