Section: New Results
Linking Data Graphs
Participants : Angela Bonifati, Radu Ciucanu, Joachim Niehren, Aurélien Lemay, Grégoire Laurence, Antoine Ndione, Slawomir Staworko.
In  , Bonifati, Ciucanu and Staworko investigate the problem of inferring arbitrary n-ary join predicates across two relations via user interactions. The relations can be found on the Web, thus they lack integrity constraints. In such a scenario, the user is asked to label as positive or negative a few tuples depending on whether she would like them in the join result or not. Deciding whether the remaing tuples are uninformative, i.e. do not allow to infer the query goal, can be done in polynomial time.
The PhD thesis of Ndione focuses on probabilistic algorithms to decide approximate membership of words in a language by using property testing. In  , Ndione, Lemay and Niehren presented an algorithm that tests the membership modulo the edit distance. Their algorithm run in polynomial time, as opposed to other property testing algorithms, leveraging the Hamming distance or the edit distance with moves, that are exponential.
In  , Laurence, Lemay, Niehren, Staworko and Tommasi (project leader of the Magnet team) studied the problem of learning sequential top-down tree-to- word transducers (STWs). They present a Myhill-Nerode characterization of the corresponding class of sequential tree-to-word transformations (STW). Next, they investigate what learning of stws means, identify fundamental obstacles, and propose a learning model with abstain. Finally, they present a polynomial learning algorithm.
In  , Niehren, Champavère (former PhD student in the team), Gilleron and Lemay addressed the problem of learnability of regular queries in unranked trees. The idea is that tree pruning strategies and the schemas (DTD in the specific case) can guide the learning process and lead to a class of queries that are learnable according to those. The obtained learning algorithm adds pruning heuristics to the traditional learning algorithm based on tree automata and exploiting positive and negative examples.