To be able to talk and to understand each other, we have to continuously store and retrieve information. Some words in fact would be impossible to understand if we were not able to recall relevant information from memory - think of words such as pronouns like he, she, or words such as there, reflexives like himself etc. But the role of memory in language is much more persuasive. For example, even in a simple sentence like "Students definitely understand the assignment", we have to recall the subject "students" when reading "understand" otherwise we would not know who understands the assignment.
In linguistics and cognitive science, recent research identified core properties of memory that we rely on in communication. The resulting model of memory has been supported by rich research. The model, however, has been applied only very narrowly. For instance, it can explain how we recall the word "students" when we read the verb "understand" in the example above. But many other functions of memory in communication, for instance, our knowledge of previous discourse, grammatical knowledge or knowledge of words, falls outside the approach.
The main objective of this project is to expand the current memory framework a way beyond its original application.
The project’s core idea is that the identified memory model is in principle right and can be used elsewhere in the study of language if we combine it with domain-specific theories developed in linguistics and cognitive science. First, we will link the model of memory to computational models of knowledge of words. Second, we will link it to models of grammatical knowledge and incremental parsing (representing how language users arrive at the meaning of sentences). Finally, we will link it to discourse theories to have an analysis of storage and recall of textual information. As a result, we will have model that is cross-domain and can describe how we generally retrieve linguistic information when we use language.
The objective of the project is relevant for current research in AI. More concretely, it is well-known that large language models (like GPT) have memory limits that are non-human like. They can be worse than humans (think of cases in which past conversations fall outside of model's memory window and the information is consequently forgotten or hallucinated) or better than humans (perfect recollection of previous discourse or texts seen during pre-training) but in any way, non-human like in some respects. One final goal of the project is to study whether the proposed computational model of memory can be combined with large language models, so that large language models can more closely approximate actual human memory systems used in language.
