Reasoner Accountability

Christopher Guy Yocum

I wrote an email to the Semantic Web email list in July 2022 on this topic. While I received some interesting responses, it did not stir the kind of interesting discussion that I had hoped for.

One of the basic problems of the Semantic Web is accountability. In a world where anyone can say anything about facts, it is hard to discern what the output of the system should be. This was discussed by Cory Doctorow and described as metacrap. In terms of IrishGen, this is less of a concern because of two things: first, the type of information that is encoded is bounded by its subject matter; second, the statements can be traced back to their origin information in the various Irish manuscripts by the more intrepid individual. However, there is one weak link in this: reasoner accountability.

There are a couple of aspects of this. First, as IrishGen is a human curated database, human errors occur. Catching these and correcting them is a part of ensuring that everyone has confidence that the information translated from primary sources is faithful to them. Second, catching coding errors in the reasoner so that they can also be corrected. When reasoners are used as heavily as they are in IrishGen, the question of accountability becomes critical.

Of the two systems that I have the most experience with (Stardog and GraphDB), Stardog does have the “Explain Reasoning Results” capability. However, this does not work when the question is: what path did the reasoner take to obtain this particular result? The Stardog reasoning explain system will only explain what inferences were done for particular classes. The Stardog reasoning system will not explain triples that it has created while evaluating the query (for more on forward and backwards chaining reasoners see Triplestores, Ontologies, and Reasoning. GraphDB provides two plug-ins for this: Proof and Provenance. I have not yet had a chance to use these as they seem to have been added in the GraphDB 10 version of the system. When I have some more time, I will have a chance to use these and see if they do what I hope that they do.

However, the overriding point of this post is that for users who are more sceptical of systems like IrishGen, auditability of the reasoner’s choices is critical to convincing them that these systems can be held to account in the decisions that are made. Additionally, auditability allows curators to investigate any anomalous results and demonstrate that the reasoner has created a sound interpretation of the data and how it has derived its results. Hopefully, more Semantic Web systems will start having these kinds of plugins so that users can interrogate the integrity of their systems.