Natural Language Processing
Dialogue Analysis centres on understanding the intent behind people’s utterances. What do they mean when they say something? If syntax is the structure of what is said, and semantics is the content, then dialogue analysis is concerned with the pragmatics – what was meant by what was said, taking into account the context of what was said before, and our knowledge of the scenario about which the participant is talking.
I am interested in automatically assigning an appropriate Dialogue Act (DA) tag to user input. There are many models for doing this, but most are machine learning approaches based on a set of specific, rather than general, features of dialogue. I am trying to identify set of general cues which enable dialogue analysis across domains. Further, I am seeking to discover re-usable patterns of DAs, in combination with cues, which form pattern of game-like structures of dialogue, such as error correction, clarification, or conventional politeness which enable us to structure dialogues.
If DA recognition is a building block, Conversation Management is one of the goals. Spoken language dialogue systems are increasingly a commercial reality, but often in limited domains, with highly structured language and constrained range of options. We are looking to expand the functionality of these systems by using more complex methods of creating dialogue systems. CM goes beyond the dialogue management seen in many deployed systems, by having a higher level view of the interaction, as a long-term, persistent goal. Such Conversation Managers can leverage patterns of dialogue to be able to control appropriate interaction, and handle errors and inconsistencies within the framework of the wider picture, or conversation.
Interactive Question Answering
In moving from factoid Question Answering (QA) to answering complex questions, it has become apparent that insufficient attention has been paid to the user’s role in the process, other than as a source of one-shot factual questions or a sequence of related questions. Rather, users both want to and can do a lot more: With respect to answers, users can usually disambiguate between a range of possible factoid answers and/or navigate information clusters in an answer space; With respect to the QA process, users want to ask more types of questions and respond to the system’s answer in more ways than another factual question. In short, real users demand real-time interactive question and answer capabilities, with coherent targeted answers presented in context for easy inspection. Repeat users will require user models that treat information already provided as background to novel information that is now available.
Such developments move the paradigm of QA away from single question, single answer modalities, toward interactive QA, where the system may retain memory of the QA process, and where users develop their understanding of a situation through an interactive QA dialogue. Dialogue systems already allow users to interact with simple, structured data such as train or flight timetables, using a dialogue component based on variations of finite-state models. Such models make intensive use of the structure of the domain to constrain the range of possible interactions. To move forward, one needs the combined capabilities of dialogue systems and open-domain QA systems.
[ back to top ]
Robots are increasingly used in collaboration with humans, in human-scale environments. For example, the Roomba vacuum cleaner is a popular, commercial product. However, this current generation of robots has limited, if any, social capacity. As roots become more complex, and we demand more from them, robots will require increasingly clever social mechanisms in order to work alongside humans effectively. From simple social measures, such as human navigation (how to navigate in areas populated by humans, applying human conventions of personal space, for example), to full human-robot interaction, where robots require models of our beliefs, goals and intentions, the realm of social robotics is an excellent environment for inter-disciplinary study.
[ back to top ]
Computer Science Education is vital, in a world where computers play an ever increasing role in all aspects of society. There are two main strands to my interest. First, computer science does not have to be a dry subject, disconnected from the real world. We can use interactive models of education, and platforms such as educational robotics, to create engaging problems and learning experiences for CS students. Second, no matter what their career path, a general education in some core principals of computer science will enable students to better understand why computers are good at some things, and very bad at others, empowering better use of technology in throughout the workplace.
[ back to top ]