Generating Customizable Natural Language Descriptions

Abstract

In the Natural Language Generation field, the task of Reference Expression Generation (REG) consists of selecting the semantic content that should compose the linguistic description of a discourse object, as in `the man in black' or `the boy in the centre'. REG algorithms tend to use predefined parameters to model qualitative and quantitative aspects of the description to be generated, and are often difficult to adapt to new domains or applications. In addition, obtaining linguistic examples in all possible reference situations is also often impracticable, which limits the use of machine learning approaches in this task. Based on these observations, this work proposes an extension of a standard REG algorithm, in which certain aspects of a reference strategy are customized without necessarily resorting to training data. The proposal was tested by making use of three REG domains, and its results compared favourably against a standard REG approach.