Language is a remarkable human capacity. No group of people has been found that lacks language and no other species has been found to communicate in a similar way. To elucidate what it is that characterizes language, linguists study human language scientifically, seeking to describe the implicit rules that govern how we use it.
Given that linguistics is the study of human language, it is perhaps surprising that there is a subfield of linguistics called computational linguistics. Computers, of course, are not people. How then is the use of computers relevant to the study of human language? And what does computational linguistics have to offer to the study of human language that is linguistics?
To answer this question, I must note that there are several distinct practices that we refer to as "computational linguistics." I myself might group them into three tracks:
- Applied computational linguistics
- Mathematical computational linguistics
- Cognitive computational linguistics
Of course, other divisions might be drawn. I'm not making any claims to total objectivity!
Applied computational linguistics
Taking each in turn, applied computational linguistics deals with the application of computers as tools for doing things with language. Most current projects in this track develop computational language models that are trained on a large corpus of text to predict which word is most likely to follow all preceding words (these are perhaps better known now as "large language models," or LLMs).
Depending on the nature of the corpus, these models can do various things. For instance, if a computational language is trained to predict which word is most likely to follow preceding words on a corpus of text where a long text is paired with a summary, it can be used to predict the most likely summary of a previously unseen long text.
Or a corpus might pair product reviews with tags indicating whether the review was positive or negative, and then could be deployed to rapidly categorize a new review by predicting which tag is most likely to follow a previously unseen review.
If a model is trained on a body of text that looks like a helpful dialog between a user and an assistant, it will learn that the most likely thing to follow a user's input is something that has the form of a helpful response to that input. This is how chatbots, like Chat GPT, Gemini, Claude and Copilot are trained.
Mathematical computational linguistics
The goal of mathematical computational linguistics is to characterize in a formal sense the expressive capacity of human language and to develop explicit, understandable algorithms that describe what human language is and how it works. For instance, mathematical computational linguists have proposed procedures that describe how people construct representations of sentences (called "parsing") and thereby understand them. The meaning of "computational" for this track, then, should be understood to refer to the process of how something is computed and not necessarily to the use of computers.
Cognitive computational linguistics
Finally, cognitive computational linguistics builds computational models of human language and interrogates how well they reflect human linguistic behaviors. Nowadays, a common practice in this track is to train or use pre-existing LLMs and obtain various measurements of linguistic probabilities from them.
Recall, as mentioned above, that an LLM is a statistical model of language, which when given some input, predicts how likely each word in its vocabulary is to come after that input. These predicted probabilities can be compared to human linguistic behavior in various ways. For instance, Hale (2001) and Levy (2008) have proposed that how difficult it will be for a person to process some word should depend on how probable that word is in context.
With an LLM, we can obtain estimates of how probable a word is in a particular context, and see how well this probability explains the difficulty people actually show when processing that word. Huang et al. (2024) have found that current models are able to explain where people will find sentences difficult to understand, but tend to greatly underestimate how difficult the sentences are at those points.
In these various ways, computational linguistics seeks to apply our understanding of human language to build useful tools, to understand the procedures that characterize the inner workings of human language and to build cognitive models that could enhance our understanding of human language. Of course, the tools, algorithms and cognitive models that clever computational linguists have already developed are imperfect.
If you have an interest in learning more and making them better, in contributing your skills and perspective to these tasks, come join the computational linguistics major. We could use all the help we can get!
About the professor
Michael Wilson, Ph.D., is a professor of computational linguistics. His work explores how language constructs meaning using formal, experimental and computational techniques.
