Dr. Jackie Cheung
Using deep learning to understand text and speech in order as it relates to pronouns, identifying individuals in sentences and assigning proper pronouns.
|Type of researcher|
|Introduce yourself, your experience and your credentials||
My group conducts research in natural language processing (NLP), an area of artificial intelligence in which we build computational models of human languages such as English or French. The goal of our research is to develop computational methods for understanding text and speech, in order to generate language that is fluent and appropriate to the context.
In our lab, we investigate statistical machine learning techniques for analyzing and making predictions about language. Several current projects include summarizing fiction, extracting events from text, and adapting language across genres.
|Describe your research||
My research area is in natural language processing, this is the subfield of artificial intelligence that is concerned with human language. The overall goal of my research is to produce systems that can perform natural language understanding and natural language generation.
In particular in my group we are focused on doing this in a way that can incorporate external sources of information. You can think of this as trying to imbue our AI systems with common sense knowledge. One way that we’ve been doing this is by looking at the types and sources of knowledge that these systems can use.
For example, you can imagine systems that can read some training data in its training set as in a typical supervised learning system. But that also reads external sources of information from Wikipedia or from other definitions from encyclopedias, as well as large amounts of general web texts.
What we’re hoping to do is to integrate all of these different sources of knowledge to reason about some linguistic ambiguity resolution decision in context. For example, one task that might group has been working on is something called the Winograd Schema Challenge. This is a set of challenging pronoun resolution questions: figuring out what pronouns like he or she refer to.
For example in the sentence “John yelled at Kevin because he was upset” the task is to figure out who is the first he. Was it John or Kevin. Our approach has been to look at integrating modern machine learning and deep learning methods that reads in large amounts of text with more classical approaches in AI such as information retrieval.
We’re hoping that by combining both sources of techniques and data knowledge we can come up with a system that better understands the context going on in there. The fact that there are two entities involved in some situation and there’s some social expectations. Using that knowledge, we’re hoping to design systems that can work better in the wider variety of contexts and situations.
|Explain its significance||
My work is important because AI technology is becoming much more widely deployed. What this means is that people are interacting with these systems on a wide variety of topics in a wide variety of settings. By imbuing these systems with common sense what I hope is that systems can actually handle these new situations that perhaps it hadn’t seen during the training of the deep learning and machine learning system.
In addition, besides the fact that we want these systems to be able to be robust and handle new situations, the other aspect to this is that we want these systems to make decisions and inferences for reasons that are justified.
You can imagine a system that thinks that all engineers are male or all nurses are female just because of some statistical quirks that it encountered during its training. But this is not what we want the system to do. We want the system to actually reason about that particular context that it is in so it can make decisions that are well justified based on the information in that particular context.
McGill University, Sherbrooke Street West, Montreal, QC, Canada
|Type of institution|