홈 › Papers

Best K-best: Efficient Item Selection for Rapid Data Annotation

We wish to leverage low-resource parsers in a human-in-the-loop process for rapidly increasing available training data. Historically, constructing rich natural language interfaces has been enabled through large annotated datasets with complex annotations. Research has shifted to developing low resource semantic parsers that make more efficient use of limited training data. While it is significant that initial parsing solutions are now viable based on limited examples, these parsers are not state of the art: the goal remains to enable highly accurate and rich interactions, the sort supported by earlier models when provided sufficient data. Here we investigate how to improve upon the simple idea of selecting new training examples from a list of model predictions. What makes for the "best $K$-best" solution for minimizing the effort of a human annotator, and maximizing the amount of new training data we can collect on a fixed budget?

Code (0)

등록된 구현이 없습니다.

Similar Papers 제목 키워드 기반