Aprendizado de ranking de entidades aplicado aos dados do governo brasileiro

Abstract

With the growth in the amount of information, according to the governmental transparency available in recent years due to legislative requirements, access to information becomes increasingly difficult. Traditional search engines like Google, Yahoo and Bing return as desired information ordered by searching the document before the informed query. The area whose objective is to return relevant documents to the user is known as Information Retrieval which can be aided by machine learning algorithms to improve the ordering of documents, called in this context as Learning to Rank (L2R). There are several algorithms in the literature to solve L2R problems which each one seeks to solve the ranking problem in the best possible way. In the context of government documents, there is a possibility of identifying which are the main entities present in the most relevant documents relevant to a given query. This work aimed obtaining an ordering of the documents available on the Brazilian Government Data Portal using Learning to Rank and extracting information from entities from unstructured, semi-structured and tabular databases, which are common among the sources available on the Portal. To achieve this goal, used state-of-the-art techniques to recognize named entities and convex optimization models to model the L2R. The results obtained proved to be superior to the search engines available on the market (Google, Yahoo and Bing) since these index only the summary of data sets from the Data Portal.