Background
Language is predominantly viewed as a means for information exchange in the field of AI and natural language processing. However, this overlooks the important role language plays as a vehicle for creative thought. This aspect of cognition is unique to mankind, crucial for innovation, and currently under-explored in AI. As an example, inventions such as the mobile phone, or the glass canoe, have come to life in people’s minds by means of combining known concepts in unseen ways. The field of NLP has mature methods to capture the meaning of words and the way they relate to each other. At the same time, a number of researchers have started questioning the learning capacity of these large language models and concerns regarding short-cut learning, bias and inability to generalise are growing. Creative tasks, such as the ones specified in this project, require a high level of generalisation that allows the system to cut across patterns found: across domains, time periods, as well as different languages. They are therefore a perfect testbed for measuring the generalisation power and actual level of intelligence of current systems.
Objectives
This project aims to investigate what aspects computational models need to perform creative cognitive tasks, from generating relatively simple novel concepts to more complex and structured ideas, across multiple domains and languages. More in particular, it aims to answer what types of structured and unstructured knowledge are needed and what models best integrate these types of knowledge.
Methods
Current neural language models are trained on large amounts of text data and used successfully in many NLP applications as models of language use. However, as shown in previous work, such statistical models might capture word associations well, but they are not sufficient for tasks that involve reasoning or generalisation across domains, which is the case for the creative cognitive task we study. Previous work from cognitive science used symbolic methods to model conceptual blending. However, these methods are often not scalable. We plan to use hybrid neural-symbolic methods for modelling creative thinking, which will allow us to add structure in our models and integrate insights from cognitive science. Furthermore, we will exploit differences in embeddings and knowledge across domains and languages/cultures to inform the models for novel concept creation.
Expected results and impact
We expect to build a variety of computational models for the range of creative tasks we defined from novel concept generation to the generation of more structured complex ideas, across domains and languages. This will lead to insights w.r.t. the limitations and capabilities of neural and neuro-symbolic approaches on such creative tasks. In addition, we will learn in how far we can leverage cognitive theory for building creative systems. Our research into creative processes will impact the field by pushing towards AI tools that are more flexible and resourceful than current technologies, which will lead to the level of innovation and diversity needed for human progress, while better exploring the wealth of data available.