Abstract—Most of Web applications combines differents services, features and contents in order to enable the creation of new features and services. Such systems are called mashups. One of the most popular kind of mashups are the location ones that use geographic data to provide functionalites to users. The RotaCerta is a location system that uses the Google Maps and perform Natural Language Generation to provide textual descriptions of routes between two different locations. The great advantage of RotaCerta is the use of points of interest (POI) to describe routes. POIs help the user to understand and assimilate the route. However, RotaCerta suffers from a several limitation: the need for manualy updating of a POIs dataset. Such work is exhausting, costly and greatly limits their use. Another point to highlight is the poor linguistic variability of texts it provides. In this work, we propose a mechanism to enable automatic feeding of POIs and a corpus-driven approach to enhance the linguistic variability of location mashups such as RotaCerta.We adopt both manual and automatic generation of new textual templates. In order to assess the quality of the routes descriptions, we use TF-IDF and cosine distance to calculate the similarity between descriptions of routes created by human volunteers and descriptions generated by the proposed approach. Route generation examples have been performed for three different brazilian cities. We also show that the text generated from the new template base is more similar to the texts used by people when describing routes if compared to Google Maps.

Index Terms—mashups, location systems, natural language generation, points of interests  

Cite: Rafael Teles, Bruno Barroso, Adolfo Guimaraes, and Hendrik Macedo, "Automatic Generation of Human-like Route Descriptions: A Corpus-driven Approach," Journal of Emerging Technologies in Web Intelligence, Vol. 5, No. 4, pp. 413-423, November 2013. doi:10.4304/jetwi.5.4.413-423