Current research

• NLP in biomedical subdomains

  My research at Cambridge studies problems presented by the domain-specific linguistic properties of biomedical texts. The massive amount of data that is added each year to medical journals is impossible for end-users (doctors, nurses, etc) to use directly. Moreover, scientific language can differ substantially from everyday language, making general-purpose resources unsuitable. Determining the degree to which this affects various specialties will demonstrate where resources are most lacking.

• Verb subcategorization

  Lexical information about verbs is particularly useful for a field like biomedicine, where new terms are frequently coined and information extraction is a common task.

• Linguistic effects of transmission and translation

  Our most culturally important literature is often of great antiquity, originally composed in the dead language of non-contemporary people, and has passed through any number of modifying forces to reach its modern form. The range of historical certainty about these forces may be used to model these forces, which can be applied in an unbiased fashion to uncover new hypotheses. The most straightforward situation is authorship identification (c.f. the canonical example on the Federalist Papers).

• Graphical models

  In all these areas, I'm interested in building generative models that explain the data. Such models have the potential to tie together many linguistic dimensions in a principled fashion.

Past research with Becky Passonneau

• Anatomy of Art

  Part of the CLiMB project of Judith Klavans at UMD, we developed linguistic feature extraction and machine learning approaches to help associate art history text with its role with respect to the work of art, e.g. whether it describes the physical properties of the work, the historical background in which it was produced, or contrasts it with other works. We particularly focused on the difficulty of this task, how humans (dis)agreed on manual annotation, and the implications for supervised machine learning from such a gold standard. Given the sparseness of the data, we developed a semantic clustering method that leveraged the WordNet ontology to combine bag-of-words features.

• MASC ANC-WordNet

  This study, with Nancy Ide at Vassar and Christiane Fellbaum at Princeton, was to gather sentences manually-annotated with WordNet synsets for a given polysemous word. The sentences were taken from the American National Corpus, and the goal was to create a disambiguated subset while simultaneously evaluating and updating the WordNet inventory of synsets for the words.

• Con Edison power grid failure prediction

  I worked on a small part of this project that involved standardizing "trouble tickets" written in the field by ConED technicians, and geocoding the locations of incident reports.