What is a Computational Model?
A computational model is like any model: it is a representation of some phenomenon in the real world. However, a key distinguishing feature of computational models is the medium of their creation. When we think of model airplanes or ships, we tend to think of building them out of physical materials. But we could also create a simulation of a ship or airplane inside a computer program. We could program this ship or airplane to behave in a computerized environment analogously to how a real ship or plane would behave in the world. This is a what is meant by a computational model.
We also probably think first of models of ships and airplanes as objects of hobby. But manufacturers use models in order to test, for example, the aerodynamics of aircraft designs before building the full-size prototype. Computational models tend to be used for these research and development purposes. That is, they help us better understand the properties of some real-world phenomenon, via its simulation in a computer.
The Role of Simplification in Computational Models
Often the computational models that we build are simplifications of the real-world phenomena. For example, meteorologists use weather models that are far from a complete replication of the Earth’s atmosphere! Such simplifications are necessary because either the techniques we use to detect the real-world phenomenon cannot give us a complete picture of it, or because our building techniques do not allow us to achieve real-world complexity, or both.
In fact, we often build computational models precisely because we cannot perfectly replicate the real-world. That is, we make assumptions or hypotheses about how features of the real-world phenomenon probably behave. Then, we build these assumptions into the model. Next, we run simulations with the model, and we observe how the model’s behavior deviates from real-world behavior. This deviation tells us about what we got right in our hypotheses, and what we need to improve on. A modeler’s dream scenario is when the model exhibits other real-world behaviors that were not intended in the original design—this is really a good indication that you’re on the right track! As a model is refined further and further, and its behavior converges more and more to that of the real-world phenomenon, we can be increasingly confident that the design properties of the model are like the unobservable properties of the real-world phenomenon.
Computational Modeling in Psycholinguistics
Psycholinguistics is a field that is concerned with how language is acquired, comprehended, produced, and processed by the human brain. In psycholinguistics, we build computational models as hypotheses of how we think aspects of these cognitive processes work.
To take an example, one question that has attracted quite a bit of attention is how language users consider multiple possible alternative completions of a sentence as they are processing that sentence word-by-word. These models posit that the expectations one has for possible completions of an ongoing sentence are based on one’s prior observations of how similar sentences were completed. Often, these models predict that language users will encounter processing difficulty when a sentence is completed in an unexpected way, versus processing ease when a sentences is completed as expected—and indeed, when one looks at experimental data, humans appear to encounter processing slow-downs and speed-ups in the same places predicted by these models!
Computational Modeling and Idiom Processing
As part of the ISLA project, we are building a psycholinguistic computational model related to these incremental processing models just described. Specifically, we are modeling how idioms are processed in a word-by-word fashion by bilinguals during both comprehension and production of language. To our knowledge, this will be the first implemented computational idiom processing model.
We are also collecting experimental data of idiom processing in German-Dutch bilinguals (see projects 1 and 2). Once the model is built, and the experiments are complete, we will run simulations with the model to try to reproduce the experimental results. To the extent that these simulations are successful, it is evidence that our hypotheses about idiom processing—as instantiated in the design of the model—are accurate.
We are currently in the stage of prototyping the model in Python. Stay tuned for more developments!