To date, large-scale applied ontology mapping has relied greatly on label matching and other relatively simple syntactic features. In search of more holistic and accurate alignment, we offer a suite of partially overlapping ontology mapping heuristics which allows us to hypothesise matches and test them against the knowledge in our source ontology (OpenCyc). We thereby automatically align our source ontology with 55K concepts from Wikipedia with 93% accuracy.
We have developed a method of specifically ontological quality control in
ontology mapping which combines a suite of partially overlapping mapping
heuristics with common-sense knowledge in OpenCyc. Our approach differs from
previous largely label-matching approaches
Concept to Wikipedia article mapping is governed by a priority queue which iteratively evaluates potential mappings ordered via continuously updated weightings. The process begins with concept-to-article mappings (Table 1), then verifies these using article-to-concept heuristics. The weight of each potential mapping is equal to the product of weights produced by the two sets of heuristics. Concept → Article TITLE MATCHING SYNONYM MATCHING CONTEXT-RELATED SYNONYM MATCHING Article → Concept TITLE MATCHING LABEL MATCHING
Example
Batman-TheComicStrip → {Batman (comic strip):1.0} ComputerWorm → {Worm:1.0, Computer worm:0.39, ... (+5 more)}
ComputerWorm → {Computer worm:1.0, Worm:0.59,... (+4 more)} representing appropriate classes. The mapping weight is multiplied by the proportion of assertions not rejected using OpenCyc’s disjointness knowledge. Example 1: “Bill Laswell is an American [[bassist]], [[record producer|producer]] and [[record label]] owner.” Only three of the four assertions in this sentence are kept: BillLaswell is a UnitedStatesPerson, BassGuitarist, and Producer. BillLaswell cannot be a RecordCompany because OpenCyc knows a person cannot be a company.
Example 2: The concept Basketball-Ball initially maps as follows (Basketball:1.0, Basketball (ball):0.95, College basketball:0.02). The second candidate is the correct one, as the first refers to the team sport. The algorithm attempts to map its first choice Basketball back to Basketball-Ball, which succeeds but also creates a new potential reverse mapping Basketball → Basketball. Consistency checking now tests “Basketball-Ball is a TeamSport”, which fails, removing this potential mapping. The next highest reverse-mapping is Basketball → Basketball, which is found to be consistent, so a mapping is recorded for that. The process now backtracks to hypothesising the second-best option from the original list: Basketball (ball):0.95, which also successfully reverse-maps and is consistent, creating a new (correct) mapping. It is worth emphasising how similar the two ‘basketball concepts’ are by standard semantic relatedness measures, and thus the subtlety our methods are capable of.
The algorithm identified 54,987 mappings of OpenCyc concepts to Wikipedia articles. Applying manual analysis to a random 300 mappings, 266 were judged ‘True’ (88.5%), 21 ‘False’ (7%) and 13 (4.3%) were assigned ‘B’ for ‘Broader term’ (the mapping was largely correct but one side generalised the other). Thus 93% of our mappings were either ‘True’ or highly related. Although YAGO reports 95% accuracy, what is being rated is not mapping joins between Wordnet and Wikipedia, but the truth of assertions in infoboxes. Although our efforts so far lack the scale of projects such as YAGO, we suggest they have a role to play in longterm development towards maximum accuracy in this field. We offer our results at: http://bit.ly/10MlLjl.