We present a Question Answering system aimed to answer natural language questions over open RDF spending data provided by LinkedSpeding. We propose an original machine-learning approach to learn generalized SPARQL templates from an existing training set of (NL question, SPARQL query) pairs. In our approach the generalized SPARQL templates are fed to an instance-based classifier that associates a given user-provided question to an existing pair, that is used to answer the user question. We employ an external tagger, delegating the Named-Entity Recognition (NER) task to a service developed for the domain we want to question. The problem is particularly challenging due to the small training set size available, counting only 100 questions/SPARQL queries. We illustrate the results of our new approach using data provided by the Question Answering over Linked Data challenge (QALD-6) task 3, showing that it can provide a correct answer to 14 of the 50 questions of the test set. These results are then compared to existing systems, including QA3, our previous work where templates were provided by an expert instead of being generated automatically from a training set.
Machine learning of SPARQL templates for question answering over LinkedSpending
Atzori M.;
2019-01-01
Abstract
We present a Question Answering system aimed to answer natural language questions over open RDF spending data provided by LinkedSpeding. We propose an original machine-learning approach to learn generalized SPARQL templates from an existing training set of (NL question, SPARQL query) pairs. In our approach the generalized SPARQL templates are fed to an instance-based classifier that associates a given user-provided question to an existing pair, that is used to answer the user question. We employ an external tagger, delegating the Named-Entity Recognition (NER) task to a service developed for the domain we want to question. The problem is particularly challenging due to the small training set size available, counting only 100 questions/SPARQL queries. We illustrate the results of our new approach using data provided by the Question Answering over Linked Data challenge (QALD-6) task 3, showing that it can provide a correct answer to 14 of the 50 questions of the test set. These results are then compared to existing systems, including QA3, our previous work where templates were provided by an expert instead of being generated automatically from a training set.
I metadati presenti in IRIS UNICA sono rilasciati con licenza Creative Commons CC0 1.0 Universal, mentre i file delle pubblicazioni sono protetti da diritto d'autore, salvo diversa indicazione.
