Publications

The field of human language technology (HLT) encompasses algorithms and applications dedicated to processing human speech and written communication. We focus on two types of HLT systems: (1) machine translation systems, which convert text and speech files from one human language to another, and (2) speech-to-text (STT) systems, which produce text transcripts when given audio files of human speech as input. Although both processes are subject to machine errors and can produce varying levels of garbling in their output, HLT systems are improving at a remarkable pace, according to system-internal measures of performance. To learn how these system-internal measurements correlate with improved capabilities for accomplishing real-world language-understanding tasks, we have embarked on a collaborative, interdisciplinary project involving Lincoln Laboratory, the MIT Department of Brain and Cognitive Sciences, and the Defense Language Institute Foreign Language Center to develop new techniques to scientifically measure the effectiveness of these technologies when they are used by human subjects.

This paper reports results from two recent psycholinguistic experiments that measure the readability of four types of speech transcripts for the DARPA EARS Program. The two key questions in these experiments are (1) how much speech transcript cleanup aids readability and (2) how much the type of cleanup matters. We employ two variants of the four-part figure of merit to measure readability defined at the RT02 workshop and described in our Eurospeech 2003 paper [4] namely: accuracy of answers to comprehension questions, reaction-time for passage reading, reaction-time for question answering and a subjective rating of passage difficulty. The first protocol employs a question-answering task under time pressure. The second employs a self-paced line-by-line paradigm. Both protocols yield similar results: all three types of clean-up in the experiment improve readability 5-10%, but the self-paced reading protocol needs far fewer subjects for statistical significance.