Utilização de mecanismos de roteamento para seleção de sistemas de Question Answering
Resumo
The evolution of the interaction between humans and computers has accompanied the technological evolution of computers themselves.
This process culminated in the rise of a subfield of computing called Questioning Answering (QA), which provides a form of natural interaction between machines and humans --- the Question-Answer interaction model.
This model manifests itself in at least two forms of systems: restricted domain systems, which are specific, limited, more complex, and open domain systems, which address general subjects, not constrained to a particular topic, exhibiting a diversity that prevents the presentation in greater detail on any particular topic.
Ideally, a QA system should combine in a practical way the main characteristics of the existing models in order to unite the variety of topics covered by the open domain systems and the thoroughness of the restricted domain systems.
One possible solution is to combine several instances of restricted domain systems into a single open domain system. For this, a routing mechanism able to select one of the available instances must exist. Then, the selected instance should answer the question about the represented domain.
In this work, a mechanism for selecting instances of QA systems is presented in shape of a hierarchical question domain classifier. Domains are naturally organized in a hierarchical taxonomy.
When classifying the proposed questions into one of them, the classifier tries to select the most suitable QA system to answer the question.
Although, for the purpose of training the classifier, quality data is mandatory. To tackle this dependency, an automatic question generation strategy based on documents was applied, resulting in a large synthetic question dataset.
Results were promising when the classifier was evaluated against a real question dataset, suggesting that automatic question generation is feasible to train the classifier.
In conclusion, the developed routing mechanism can be used to build a solid and universal hybrid QA system, ensembling the best qualities of each kind of system stand-alone.