A growing number of studies use a "ghost-driver" vehicle driven by a person in a car seat costume to simulate an autonomous vehicle. Using a hidden-driver vehicle in a field study in the Netherlands, Study 1 (N = 130) confirmed that the ghostdriver methodology is valid in Europe and confirmed that European pedestrians change their behavior when encountering a hidden-driver vehicle. As an important extension to past research, we find pedestrian group size is associated with their behavior: groups look longer than singletons when encountering an autonomous vehicle, but look for less time than singletons when encountering a normal vehicle. Study 2 (N = 101) adapted and extended the hidden-driver method to test whether it is believable as online video stimuli and whether car characteristics and participant feelings are related to the beliefs and behavior of pedestrians who see hidden-driver vehicles. As expected, belief rates were lower for hidden-driver vehicles seen in videos compared to in a field study. Importantly, we found noticing no driver was the only significant predictor of belief in car autonomy, which reinforces prior justification for the use of the ghostdriver method. Our contributions are a replication of the hidden-driver method in Europe and comparisons with past US and Mexico data; an extension and evaluation of the ghostdriver method in video form; evidence of the necessity of the hidden driver in creating the illusion of vehicle autonomy; and an extended analysis of how pedestrian group size and feelings relate to pedestrian behavior when encountering a hidden-driver vehicle.

Pradipto Das, Yandi Xia, Aaron Levine, Giuseppe Di Fabbrizio, Ankur Datta. Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers. 2017.

E-commerce catalogs include a continuously growing number of products that are constantly updated. Each item in a catalog is characterized by several attributes and identified by a taxonomy label. Categorizing products with their taxonomy labels is fundamental to effectively search and organize listings in a catalog. However, manual and/or rule based approaches to categorization are not scalable. In this paper, we compare several classifiers to product taxonomy categorization of top-level categories. We first investigate a number of feature sets and observe that a combination of word unigrams from product names and navigational breadcrumbs work best for categorization. Secondly, we apply correspondence topic models to detect noisy data and introduce a lightweight manual process to improve dataset quality. Finally, we evaluate linear models, gradient boosted trees (GBTs) and convolutional neural networks (CNNs) with pre-trained word embeddings demonstrating that, compared to other baselines, GBTs and CNNs yield the highest gains in error reduction.