Research

We humans can comprehend words and sentences both quickly and accurately. This enables us to do our daily tasks such as talking with friends or reading books in a rather effortless way, which is quite surprising considering the complexity of the tasks and the amount of information processed. One mechanism considered to be contributing to this incredible ability is prediction: humans anticipate upcoming parts of utterances and start processing them even before they actually hear/see them. This can then facilitate the processing of those upcoming parts when they actually appear. I am interested in how people realize such predictive processing and trying to understand the mechanism of prediction using experimental and computational methods.

I mainly work on cases called argument role reversal anomalies as a window to prediction in sentence processing. For example, the verb served is appropriate in sentence (1), but it is not in sentence (2), where the arguments waitress and customer have the different roles. Therefore, it should be only predicted in (1) but not in (2).

(1) The restaurant owner forgot which customer the waitress had served.
(2) The restaurant owner forgot which waitress the customer had served.

However, many studies have found that neural and eye-movement prediction measures show no contrast for cases like (1) and (2). This suggests that people cannot use argument roles for prediction, even though who is doing what to whom is a very important piece of information in sentence comprehension. This is surprising because there is ample evidence showing that people can use different sources of information immediately and mostly make accurate predictions.

Interestingly, we found that people's predictions seem to be sensitive to argument roles when they are explicitly asked to articulate their predictions following the contexts such as (3) or (4) in the speeded cloze task. They often answer served in (3) but very rarely in (4). This was true even when we used the same sentences and forced them to produce responses in the same time windows as a preceding neural study. This suggests that people can actually use argument roles to generate predictions in the speeded cloze task.

(3) The restaurant owner forgot which customer the waitress had ______
(4) The restaurant owner forgot which waitress the customer had ______

Why do different measures of prediction show conflicting results? I am now trying answer two questions by addressing this puzzle: (i) Can people use argument roles for prediction? (ii) What do different measures of prediction actually reflect?

I have been also interested in developing the methodology of the speeded cloze task as a prediction measure. Because this can be easily conducted online, we can collect prediction data very quickly and even from people in remote places. In fact, I even collected Japanese speeded cloze data in a couple of days from America during the pandemic. I have developed the TransCloze pipeline to facilitate transcription and onset coding of speech data, which can also be used for other production experiments. I also have implemented a computational model for the speeded cloze task.