Research Statement Research Accomplishments

Current research in probabilistic models of natural language is heavily biased towards systems that do not take into account the fact that language is typically generated and recognized incrementally. Even systems designed to operate on spontaneous speech tend to take as input an entire utterance, already transcribed by a pre-process, and then perform downstream processing on the entire string at once. While these methods are effective on edited text and correctly transcribed fluent speech, a pipelined approach still separates uncertainty about different levels of the signal which might be useful when combined, and cannot operate in real time on streaming input. My research, in contrast, focuses on developing incremental processing models for streaming input, such as speech or ambiguously punctuated text (e.g., unedited text in blogs or other digital media). This line of research offers contributions to both engineering and scientific research on models of human language. From an engineering perspective, these incremental models enable continuous interaction in human-computer interfaces, offer more realistic and useful representations of user input, and can potentially enable faster processing of language. From a scientific perspective, these models are attractive to psycholinguistics researchers, because their incremental nature means that they take advantage of the same information that would be available to a human listener, and can thus contribute to psycholinguistic models of language recognition.