Under the direction of Martha Palmer and Jim Martin, my dissertation work focuses on improving machine learning approaches to metaphor detection, . This is done be combining linguistic analysis and deep learning: syntactic structures are predictive of metaphor in many cases, and we use this to identify patterns in corpora that can then be leveraged by state of the art neural networks. We've shown significant improvement on a variety of tasks, but many open problems remain. For me, the most interesting is how to develop good semantic representations of metaphor that are useful for downstream applications.
I work extensively with lexical resources, particularly VerbNet. For the last six years, I have worked on maintaining and updated VerbNet through annotation projects and lexical semantic analysis. I've developed tools to improve functionality, and all this has lead to for efficient applications of VerbNet tools to NLP tasks.
I've worked with Leysia Palen and Ken Anderson in Computer Science and Information Science on joint projects with the National Center for Atmospheric Research. We employed modern NLP to social media data generated during natural disasters, combining comprehensive annotation projects and machine learning. We classify relevance of tweets, user behavior, threat information, and more. This innovative research allows for better understanding of how information propogates during crises
Recently I've been working on developing embeddings for structures in lexical resources. My preliminary work is on developing VerbNet class, sense, and role-based embeddings. These are generated by replacing contituents in corpora with their respective VerbNet elements, then employing any standard embedding learner. I hope to be able to extend these to other resources (like FrameNet) soon!