Striking a Balance: Human and Computer Contributions to Learning through Semantic Analysis

Manual acquisition of high-quality, broad-coverage knowledge needed by knowledge-based NLP systems is commonly considered too expensive a procedure, and has been known to cause "the knowledge acquisition bottleneck". The use of the web as a corpus to support automating knowledge acquisition has been gaining in popularity in the recent years. This approach tends to introduce noise at the early stages of learning which can have a compounding impact on the quality of the final results. If the goal is to alleviate the expense of manual knowledge acquisition in the short term, a combination of automatic knowledge learning and human validation / correction must be considered. People can either post-edit automatically produced candidate knowledge elements or intervene at various stages in the acquisition process to facilitate high-quality automatic output. In this paper, we report on a sequence of experiments analyzing the utility of the latter methodology in the framework of a mutual-bootstrapping environment in which new knowledge resources are acquired both for and through the operation of an automatic meaning extraction system.