Testing the JANUS model of noun-phrase anaphor interpretation

One of the most obvious features of well-written text and normal conversation is that the same people, places, and things are referred to a number of times. Surprisingly, from a scientific perspective, such repeated references often use different linguistic expressions. So, for example, someone referred to initially as "the person serving in the butcher's shop" might later be referred to as "the man" or simply "he", using what are called "anaphoric" expressions, or expressions that refer back to the previous text. To a language user these changes of linguistic expression seem perfectly natural, but what is the underlying explanation for such variation? In recent years there have been a few tentative attempts to provide an overall account, in scientific terms, of how reference and co-reference work in language understanding and language production. One of these accounts was developed by our group on a previous ESRC-funded project. It is called JANUS, because it says that to understand how an expression such as "the man" or "he" works in text it is necessary to take account of both how they link to what has already been said and to what might be said in the upcoming text. The main way of linking back to what has already been said is to identify previous words that refer to the same person or thing as the current word (or sometimes to link up to that person or thing in a less direct way). How we make links to what might be said has not been studied so much, but might include such things as anticipating a change in perspective on a person or thing. The JANUS model claims that when something like a change in perspective is anticipated, the word used will be different from what it otherwise might have been. Generally, it will be more specific (e.g., if the person working in the butchers is referred to as "the footballer", rather than just "he", the text is probably going to go on to talk about his sporting activities, rather than his work).

Understanding the basic processes of text comprehension allows the design of texts that are easy to understand. Such texts are increasingly important in a world in which much of our comprehension takes place in impersonal contexts, such as the internet (e.g, web browsing, and on-line form-filling). In the longer term, our work will have implications for designing tools for extracting information from text automatically. Another important applied issue on which our research bears is the understanding of how comprehension breaks down, for example, in cases of brain injury or degenerative disease, and hence it may suggest ways of helping people with comprehension problems.

This project will try to provide empirical support for the JANUS model. It will test some specific claims of that model about how links to what has already been said are made: claims about, for example, which people and things mentioned in a text are considered as the person or thing now being talked about. Some of these claims contrast with those made in other, rival, framework's, such as Amit Almor's Informational Load Hypothesis (ILH). It will also consider a specific prediction from JANUS about links to what might be said. If using an expression that is more specific than expected signals a change in perspective on the person or thing referred to, then following information that is consistent with that change of perspective (e.g. something about a football match, when a person is unexpectedly called a footballer) should be easy to understand. On the other hand, if the next bit of information is not about football, comprehension should be slightly more difficult than it otherwise would have been.

The research addresses one of the most basic questions about text and discourse comprehension (and the production of comprehensible discourse). How are links made between very different expressions referring to the same person or thing (e.g. "the London Eye" and "it") and what factors make these links easy or difficult to discern for the listener or reader. It aims to test a theory of the processes underlying the making of these links, based on a small number of general psychological principles. The most immediate impacts of the research will, inevitably, be on academic researchers carrying out similar research. However, the research has broader implications within psycholinguistics and cognitive psychology more generally. A proper theory of coreference processing has implications, for example, for developmental theories of how discourse comprehension develops, and provides a framework for describing the different skills that children have at different ages, as their comprehension abilities develop. Such a theory also has implications for what goes wrong in comprehension for people with various types of brain damage or degenerative diseases. For example, Amit Almor, author of one of the rival theories to JANUS (the ILH or informational load hypothesis) has applied his ideas to the breakdown of comprehension seem in patients with advance Alzheimer-type dementia.

Work on coreference processing can also have implication for linguistic analyses. Peter Gordon, author of another rival theory (derived from Centering Theory, as developed in computational linguistics) has used his work with Randall Hendrick to question an account of coreference based on Chomsky's theory of Government and Binding.

Casting a wider net, comprehensible text is increasingly important in everyday life, as more and more interpersonal communication is replaced by computer-based communication. Given that people have to understand information presented via computers with no help or with further computer-based help, it is important that the "design" of the material presented via computer-based system takes account of the way that people typically process text. For example, our JANUS model predicts that expressions that appear to be more specific than they need to be serve specific functions, such as emphasis or change of perspective. If no change of perspective is intended, an overspecific expression may be misleading. The results of our research will be of use to people who design text for the internet.