Learning regular expressions to template-based FAQ retrieval systems

Template-based approaches have proven to be one of the most efficient and robustest ways of addressing Question Answering problems. Templates embody the expert’s knowledge on the domain and his/her ability to understand and answer questions, but designing these templates may become a complex task since it is usually carried out manually. Although these methods are not automatic, companies may prefer to undertake this solution in order to offer a better service. In this article, we propose a semiautomatic method to reduce the problem of creating templates to that of validate, and possibly modify, a list of proposed templates. In this way, a better trade-off between reliability—the system is still monitored by an expert—and cost is achieved. In addition, updating templates after domain changes becomes easier, human mistakes are reduced, and portability is increased. Our proposal is based on inferring regular expressions that induce the language conveyed by a set of previously collected query reformulations. The main contribution of this work consists of the definition of a suitable optimisation measure that effectively reflects some important aspects of the problem and the theoretical soundness that supports it.