Much of the previous year’s work on answer coalescence involved improving the performance and stability of prototype services. This work has shown that answer coalescence provides novel insights into collections of answers, and that this insight can be provided efficiently.
Improvements to the current algorithms will be twofold. First, the current algorithm can only group answers by a single node, and requires that all other nodes and edges in the grouped answers are equal. While powerful, this leads to situations in which one group of answers is merged around one node, and another group around a second node. A more complete and understandable answer will allow for merging across multiple nodes simultaneously.
Second, we will improve the enrichment calculations underlying answer coalescence. In particular, the calculation involves the number of possible KG nodes that may have been bound to a particular query graph node. Currently, this number is estimated as the overall number of possible nodes of that type (e.g. the number of chemicals in a large KG). However, a given question places further constraints on the entities. A question involving a “treats” edge will probably not simply involve any chemical, but the smaller set of drugs. Without taking that into account, the current algorithm (correctly) returns properties enriched for all drugs. Estimating the effects of query constraints will further sharpen the returned properties, providing only the most interesting properties to the user.
Y1 History:
Extending the prototyping work, we will first determine answers that are close in a graph edit-distance sense [2], then programmatically relate the edits using a series of ontological and similarity checks, including comparison to the graph enhancement question library to check for multiple ways to express similar concepts.
Much of the previous year’s work on answer coalescence involved improving the performance and stability of prototype services. This work has shown that answer coalescence provides novel insights into collections of answers, and that this insight can be provided efficiently.
Improvements to the current algorithms will be twofold. First, the current algorithm can only group answers by a single node, and requires that all other nodes and edges in the grouped answers are equal. While powerful, this leads to situations in which one group of answers is merged around one node, and another group around a second node. A more complete and understandable answer will allow for merging across multiple nodes simultaneously.
Second, we will improve the enrichment calculations underlying answer coalescence. In particular, the calculation involves the number of possible KG nodes that may have been bound to a particular query graph node. Currently, this number is estimated as the overall number of possible nodes of that type (e.g. the number of chemicals in a large KG). However, a given question places further constraints on the entities. A question involving a “treats” edge will probably not simply involve any chemical, but the smaller set of drugs. Without taking that into account, the current algorithm (correctly) returns properties enriched for all drugs. Estimating the effects of query constraints will further sharpen the returned properties, providing only the most interesting properties to the user.
Y1 History: Extending the prototyping work, we will first determine answers that are close in a graph edit-distance sense [2], then programmatically relate the edits using a series of ontological and similarity checks, including comparison to the graph enhancement question library to check for multiple ways to express similar concepts.