Learning Object-Language Alignments for Open-Vocabulary Object Detection