Thoth : um algoritmo para geração de frases curtas em linguagem natural a partir de dados abertos e conectados

Abstract

The current consolidation and availability of linked open data have fomented several initiatives, among them it is possible to observe the use of the content stored in them for natural language generation. The generation of natural language phrases can benefit from using these bases in at least two aspects, which are the large amount of information available and the existence of additional notes on the meaning of this information. As for the resources used for the lexicalization of sentences, the works in this area can be grouped into three categories: the first one characterized by the use of sets of templates to define the sentence structure; the second by the use of machine learning algorithms to the generation of sentences in an unsupervised way; and the third the use of both approaches in a hybrid model. The approaches generate interesting results but have difficulties in relation to the naturalness of the sentences generated. It is observed that the works related to the topic do not use on a large scale the information of the RDF properties present in the ontologies, factors that can be considered as support in the generation of more natural phrases. Among these are semantic relationships between concepts that can help construct sentences in natural language. In this context, the current research aims to explore these properties for the generation of natural language for the English language from a set of templates developed by linguists and the use of lexical resources. Two evaluations were performed to evaluate criteria and variables for the proposed language generation algorithm and a third one for final validation of the research. The first evaluation sought to identify ways of generating natural language phrases from the RDF properties. Starting from the analysis of the results of the first evaluation, a new experiment was conducted to measure the naturalness of the sentences generated from the RDF properties. Finally, a third evaluation was designed and executed, where linguistic professionals and native English speakers evaluated the short sentences generated by the algorithm. The results of the final evaluation were considered promising for applications that aim to generate natural language from the information of RDF properties with the support of lexical resources.