Uma abordagem de componentes combinados para geração de funções de ordenação usando programação genética

Abstract

Due to the advent of the Web and other textual repositories, such as digital libraries, the information retrieval task has become a very complex and challenging problem. In this context, search engines became valuable tools for the information retrieval task in document collections. These tools are based on information retrieval models whose main goal is to produce, given a query, a set of documents ranked by relevance as an answer. For doing so, the so-called ranking functions are employed. Several ranking functions have been investigated throughout the years. However, most of them attempt to be very general in nature, i.e., they were designed to be effective in any type of collection. In this work, we propose a new method to discover collection-adapted ranking functions based on Genetic Programming (GP). The evolution process of our Combined Component Approach (CCA), differently from other approaches based on GP, uses several components extracted from effective and well-known ranking functions. Our assumption is that these components are representative and meaningful and can be combined for generating a more effective and specific new ranking function for a given document collection. Experimental results show that our approach was able to outperform in more than 40% standard TF-IDF, BM25 and other GP-based approach (named FAN-GP) in two different collections. The CCA evolution process also was able to reduce the overtraining, commonly found in machine learning methods, especially genetic programming.