Knowledge and Learning in Natural Language (Charles

Introduction. This review was written from the perspective of someone working in the field of Machine Learning of Natural Language: the application of techniques from statistical pattern recognition and Machine Learning to problems in Natural Language Processing (NLP). As in any subfield of Artificial Intelligence, the goals of this type of research can be either practical or theoretical. From the point of view of practice (language engineering), Machine Learning methods have been shown to allow the construction of more efficient, more robust, and more accurate NLP systems in a faster way than by handcrafting rules. The explanation for this is that these inductive systems are trained on large corpora of real language use, rather than being based on linguistic intuitions of one or a few (computational) linguists. On the other hand, the induced theories lack the linguistic sophistication of hand-made models. Theoretical results include an increased understanding which linguistic knowledge can be learned from primary linguistic data, which sources of knowledge are necessary for learning a particular NLP task, what distribution of effort and importance should be assigned to programmed knowledge and induced knowledge, and which learning algorithm has the right ‘bias’ to learn linguistic knowledge.