Nobuhiro Shimogori

Automatically generated captions: will they help non-native speakers communicate in english?

Many people find it difficult to communicate in a foreign language. In order to help these people, one approach being studied is the use of captions generated by automatic speech recognition (ASR). Captions are known to facilitate comprehension of foreign languages, but ASR-generated captions may be subject to problems attributable to recognition errors and recognition time. We conducted two experiments using subjects who are native Japanese speakers to determine how these differences caused by ASR affect understanding when listening to English. We found that captions with 80% accuracy will increase the understanding of the subjects with intermediate English skills, which would apply to about half of native Japanese users. Additionally, changing the display timing of the caption from after speech to before speech would contribute to improving the understanding more than increasing accuracy from 80% to 100%. These findings suggest that captions generated with todays ASR can help non-native speakers communicate in English when used carefully

An inductive learning algorithm based on regression analysis

There are two types of learning. One is symbol learning such as inductive learning in artificial intelligence and the other is pattern learning such as multivariate analysis and neural networking. This paper presents an inductive learning algorithm which obtains propositions whose errors are at a minimum. The algorithm is based on regression analysis and works better than C4.5. The algorithm consists of preprocessing data, obtaining linear functions by multiple regression analysis, and approximating the functions with Boolean theory. Approximating linear functions with Boolean theory is a pseudo maximum likelihood method and regression is the least square method. They are the principles of pattern learning, and so an algorithm for symbolic learning can be obtained by these principles.