Active Error Correction for learning kinship terms
Kinship Analysis requires learning definitions of all kinship terms used in a target culture from data that is gathered incrementally through interviews with informants. We exploit a collection of previously learned kinship definitions from different cultures (logical “domain theories”) to minimize the cost of learning the definitions in a new domain theory. We use Transfer Learning and Inductive Logic Programming (ILP) to learn kinship definitions. We propose a novel method for identifying potential errors in the data by comparing our data with definitions in previously learned models of similar cultures. We actively ask informants to confirm or correct these potential errors. This Active Error Correction is enhanced by computing the similarity between the target domain theory and those of other cultures, and using these similarities to guide the search for definitions. Our similarity-guided Active Error Correction learns more than 90% of the kin terms, even when training examples are up to 19% incorrect. Suppressing error correction reduces learning accuracy by 29%. Because Active Error Correction correctly identifies almost all of the errors in the data for human correction, it reduces the need to collect further data by 20%. Furthermore, Similarity-Guided Search reduces repeat data requests of the User by 40%. The combination of Active Error Correction and Similarity-Guided Search can be used to reduce data requirements for ILP learning problems. As a side benefit, this work has produced a machine readable collection of 50 kinship domain theories that may provide broader benefits in anthropology, by surfacing interesting anthropological issues.
Artificial intelligence|Computer science
Morris, Gary, "Active Error Correction for learning kinship terms" (2008). Dissertations available from ProQuest. AAI3328624.