A Connectionist Model for Part of Speech Tagging

AutoTutor is a fully automated tutoring system that attempts to comprehend learner contributions and formulate appropriate dialogue moves. This paper reports the mechanisms and performance of one of AutoTutor’s language modules, the word tagging module. AutoTutor’s word tagging module determines the part of speech tag for every word in the learner’s contributions. It uses a two part procedure: it first consults a lexicon to identify the set of possible tags for each word, then it uses a neural network to select a single tag for each word. Performance assessments were made on a corpus of oral tutorial dialogue, as opposed to well-formed printed text. The lexicon provided the correct tag, as one member of a set, for 97% of the words and 91.6% of the neural network’s first-choice tags matched assignments by humans.