A comparison of Lucene search queries evolved as text classifiers

In this article, we use a genetic algorithm to evolve seven different types of Lucene search query with the objective of generating accurate and readable text classifiers. We compare the effectiveness of each of the different types of query using three commonly used text datasets. We vary the number of words available for classification and compare results for 4, 8, and 16 words per category. The generated queries can also be viewed as labels for the categories and there is a benefit to a human analyst in being able to read and tune the classifier. The evolved queries also provide an explanation of the classification process. We consider the consistency of the classifiers and compare their performance on categories of different complexities. Finally, various approaches to the analysis of the results are briefly explored.