Search Results

In this paper, natural language processing (NLP) and machine learning methods and tools are applied to the task of topic (thematic or semantic) classification of Portuguese proverbs. This is a difficult task since proverbs are usually very short sentences. Such classification should allow an easier selection of the most relevant proverbs for a given situation, considering their context in discourse or within a text. For that, we used, on the one hand, a collection of +32,000 proverbial expressions organized "thematically" into a large set of previously attributed topics (+2,200) and, on the other hand, the Orange data mining toolkit, along with the NLP and machine learning tools it provides. Since the classification provided in the collection of proverbs is, for the most part, based only on a keyword in the body of the proverbs, 2 experiments were set up, to determine the feasibility of the task with a modicum of effort and the most promising configurations applicable. Different sample sizes, 100 and 50 proverbs randomly selected per topic, corresponding to Scenario 1 and 2, respectively, were contrasted; several preprocessing strategies were explored, and different data representation methods tested against several learning algorithms. Results show that Neural Networks is the best performing model, achieving the best classification accuracy of 70% and 61%, in the two different experimental scenarios, Scenario 1 and 2, respectively. Some of the inaccurate classification cases seem to indicate that the machine learning approach can sometimes do a better job than a human classifier, especially considering the manual attribution of the topics by the collection’s author, the sheer number of topics involved, and the very unbalanced distribution of proverbs per topic. Based on the results achieved, the paper presents some proposals for future work to cope with such difficulties.