Research

Like many other linguists I am impressed by the rich structural properties of human language, by the fact that speakers of a language agree on so many abstract and detailed facts about their language that they are not consciously aware of, and by the scope and limits of variation among the world’s 7000 languages. Also like many other linguists I regard these abilities as part of humans’ cognitive and neural capacities. What is more distinctive about my approach — though by no means unique — is how seriously I take the integration of linguistic, cognitive, and neuroscientific approaches.

My overall research goals are straightforward: I want to understand how rich linguistic representations are encoded and manipulated in neural tissue. And I want to understand how children are so successful at learning those representations. My research is primarily curiosity-driven, but I also believe that if we can understand how human language works at its best, in the brains of speakers who learn from an early age, then we will be better able to solve language-related problems in education, technology, and health.

But achieving these goals is a little complicated. I expect it to take a while.

Our approach to these big challenges includes several key components.

1. Framing the problem.

A key step is to figure out what an answer might even look like. And to figure out how to break the problem down into sub-parts, and to identify how to link up those sub-parts. This all sounds easier than it really is.

2. Multiple fields.

Our work necessarily engages with multiple fields: linguistics, psychology, neuroscience, and computer science. And in each case it is important to take the details seriously. We take seriously what linguistics tells us about the richness of the phenomena to be encoded or learned, and we have learned a lot from focusing on more complex or more abstract cases.

3. And tools.

Our tools range from low-tech pencil-and-paper research to mid-tech studies using reaction times and eye-movements, to high-tech experiments using electro- and magneto-encephalography. Computational models provide important support for all of the research, as they help us to generate and test explicit hypotheses.

4. And languages.

We can do a lot using only English. But we can do a lot more if we work with diverse languages, as they often allow us to test hypotheses that we cannot readily test in English. We have run experiments in around 15 languages. Language diversity is even more central for the learning problem, as it defines the scope of what children must learn.

5. Create a suitable environment

All of this is more feasible if it is done within a conducive environment. We have addressed this through all-shared state-of-the-art lab facilities, through diverse collaborations, through interdisciplinary student training programs, and now through the Language Science Center.

Setting aside the details of individual studies, there are a few consistent themes in what we have learned in our research.

1. Adult parsing is grammatically sophisticated

Across many different studies we find that the representations that adult comprehenders as they read/listen are tightly constrained by their grammar, including by abstract and obscure constraints.

2. … mostly

For a long time I was trying to show that adult comprehension is grammatically perfect in every regard. That didn’t work out. And that turned out to be great news. The fact that speakers are systematically good and bad with different linguistic constraints (“selective fallibility”) is far more informative than uniform success or uniform failure would have been.

3. The same is true for kids

Children show a similar mixed profile of grammatical successes and failures. And although it is cool to marvel at the cases where children succeed early, we probably learn more from the cases where they show a delay.

4. And some of the specific adult-child parallels are uncanny

A large body of our adult and child findings are covered by the following generalization, assembled from findings that Jeff Lidz and I assembled either individually or jointly: adults’ first interpretation is children’s only interpretation.

5. That language understanding is like language production

Comprehension is really a form of production, where the comprehender tries to generate an internal representation that matches what he’s reading/hearing. This is clearest in cases of ‘prediction’, where the comprehender gets ahead of the speaker. We were early to join the prediction bandwagon, and we’re certainly not alone in this. But as in other areas we’re starting to find that the predictions are not always so fast/accurate, and that this may be more revealing about how the system works.

I am placing my bets on the following problems being the most informative in the coming years. This shapes my priorities.

• Selective fallibility: we have learned a lot from speakers’ uneven profile of successes and failures in implementing grammatical constraints in real-time comprehension. We have recently learned how to “turn on/off”  linguistic illusions. Understanding this is sure to be revealing.
• Real-time semantic interpretation is a new area that will become more important for us in the coming years. It’s starting to look like speakers are less fast and accurate than we thought (when they don’t already know what you’re talking about). This makes it more interesting.
• We have started to dig more deeply into the learning of hard-to-observe phenomena. Solutions to these problems should readily “scale down” to the easier problems. The opposite is not true.

I would gladly pursue research on cross-population differences or computational modeling, if the right collaborators came along.