Conclusions

Acquiring a semantic lexicon from a corpus of sentences labeled with representations of their meaning is an important problem that has not been widely studied. We present both a formalism of the learning problem and a greedy algorithm to find an approximate solution to it. WOLFIE demonstrates that a fairly simple, greedy, symbolic learning algorithm performs well on this task and obtains performance superior to a previous lexicon acquisition system on a corpus of geography queries. Our results also demonstrate that our methods extend to a variety of natural languages besides English, and that they scale fairly well to larger, more difficult corpora. Active learning is a new area of machine learning that has been almost exclusively applied to classification tasks. We have demonstrated its successful application to more complex natural language mappings from phrases to semantic meanings, supporting the acquisition of lexicons and parsers. The wealth of unannotated natural language data, along with the difficulty of annotating such data, make selective sampling a potentially invaluable technique for natural language learning. Our results on realistic corpora indicate that example annotations savings as high as 22% can be achieved by employing active sample selection using only simple certainty measures for predictions on unannotated data. Improved sample selection methods and applications to other important language problems hold the promise of continued progress in using machine learning to construct effective natural language processing systems. Most experiments in corpus-based natural language have presented results on some subtask of natural language, and there are few results on whether the learned subsystems can be successfully integrated to build a complete NLP system. The experiments presented in this paper demonstrated how two learning systems, WOLFIE and CHILL, were successfully integrated to learn a complete NLP system for parsing database queries into executable logical form given only a single corpus of annotated queries, and further demonstrated the potential of active learning to reduce the annotation effort for learning for NLP.