This thesis presents an approach to the automatic acquisition of structured descriptions from unstructured data of a given domain. The acquired knowledge is represented in the lexical knowledge representation language DATR, an inheritance-based formalism that permits the representation of default information. The basic components of the learning system are a set of inference rules that establish inheritance relationships between data, and a default-inference algorithm that reduces a monotonic description to a default description and thereby generalizes the data. For a given set of data there usually exist many different DATR descriptions. Therefore, a heuristic inference strategy is suggested that identifies promising candidates. This is achieved by the use of a set of criteria for evaluating the quality of a given DATR theory. Different domains may require different criteria or give different priority to a set of criteria.
The learning approach is applied to a number of tasks from two different linguistic domains. In the first group German inflectional classes are acquired from examples of inflected nouns. In the tasks of the second domain verbs are classified according to their syntactic properties.
The inferred DATR theories capture relationships between data in a systematic way and generalize over them. The acquired classes are linguistically plausible and can be associated with traditional linguistic classes like 'strong nouns' or 'transitive verbs'. The results show some characteristic properties in the usage of available descriptive means as well as in the form of the descriptions. A comparision of the automatically inferred descriptions with DATR theories written by linguists suggests some enhancements of the approach, such as the exploitation of statistical information.
E-mail this message to a friend
