Research Interests

My primary research area is natural language processing, though I am also interested in machine learning, information retrieval, and artificial intelligence in general. Much of my research focuses on automatically acquiring the knowledge needed for conceptual natural language processing using unannotated text corpora. Some current and previous research projects are described below. ("We" refers to work done by me and members of the NLP group at Utah.)

• Learning Patterns for Information Extraction
  We have developed several algorithms to automatically learn extraction patterns. AutoSlog was the first system, which learned extraction patterns using an annotated training corpus. AutoSlog-TS was its successor, which learned extraction patterns using an unannotated but preclassified training corpus (preclassified means the texts are identified as relevant or irrelevant to the domain). Most recently, we have developed a meta-bootstrapping algorithm that can learn both extraction patterns and a semantic lexicon simulataneously using only unannotated texts.

• Learning Semantic Lexicons
  Semantic dictionaries are essential to support language understanding. We have developed two different algorithms for automatically acquiring domain-specific semantic lexicons. One method is the meta-bootstrapping algorithm described above, and the second method is a statistical collocation algorithm that also requires only an unannotated text corpus.

• Extraction-based Text Categorization
  Our forays into information retrieval have used information extraction techniques to support high-precision text categorization. We have developed several text classification algorithms that use extraction patterns to recognize specific contexts and role relationships, which can be essential for some classification tasks. The most notable algorithms use relevancy signatures and augmented relevancy signatures to represent extraction-based text classification terms.

• Anaphora Resolution
  We have recently begun to investigate the problem of resolving anaphora. One key challenge is to identify which noun phrases need to be resolved - many do not. We created a corpus-based method for automatically identifying definite noun phrases that are not anaphoric because they do not have antecedents in the text.

• Question Answering
  Another recent area of interest is computational question answering, especially for reading comprehension. We developed a rule-based system called Quarc that can take reading comprehension exams. We also participated in the Summer 2000 Johns Hopkins Workshop on Reading Comprehension, where we developed a question answering system called Spot.

• Natural Language Interfaces for Programming
  An ongoing research project involves creating a natural language interface for Java programming. This interface accepts natural language sentences as programming instructions (e.g., "create a for loop that iterates from 1 to 10") and automatically creates Java source code.

• Sundance
  Most of our research uses our home-grown natural language processing system called Sundance. Sundance is a heuristic-based shallow parser that also activates and instantiates case frames for information extraction.