Answering Reading Comprehension Tests:N-gram and Multi-View Regression

Computer “reading” of natural languages is a growing research topic with numerous applications, ranging from areas of artificial intelligence to linguistics. A successful “reading” of natural languages can be interpreted in two ways: 1) extracting facts out of a passage and 2) capturing “meaning” of a text. We focus on the latter interpretation, defining “meaning” as is done in humans, namely, can the system answer the same reading comprehension questions that are given to early learners of language. We present two approaches to answering cloze questions, which are fill-in-the blank, multiple-choice reading comprehension questions. The first approach utilizes n-gram, a basic statistical methodology that captures the local context of a passage to predict an answer. The second approach uses Multi-View Regression (“MVR”), based on canonical correlation analysis (“CCA”), which captures both the local and broader context. In this statistical approach, the “meaning” of each word is represented as a real-valued state vector associated with each word occurrence. MVR is a dynamical belief net which, like a Kalman filter or Hidden Markov Model (“HMM”), estimates the state based on the preceding and following sequence of words and predicts the word to be “emitted” based on that state. While there has been much research on these statistical methodologies in the field of machine learning and on generation and evaluation of cloze questions in the field of natural language processing, the connection between the two hasn’t been explored much. Prior research in cloze questions suggests that proper context capturing leads to correct answers. We create a system that reads in a set of cloze questions and predicts two sets of answers with the two statistical methodologies that have different context capturing schemes, and then analyze the results. Because MVR captures both the local and broader context, it answers cloze questions better than n-gram, which only captures the local context.